Optimal Scheduling of Refinery Crude-Oil Operations

Carnegie Mellon University
Research Showcase
Dissertations Theses and Dissertations
12-1-2010
Optimal Scheduling of Refinery Crude-Oil

Operations
Sylvain Mouret
Follow this and additional works at: http://repository.cmu.edu/dissertations
Recommended Citation
Mouret, Sylvain, "Optimal Scheduling of Refinery Crude-Oil Operations" (2010). Dissertations. Paper 23.
This is brought to you for free and open access by the Theses and Dissertations at Research Showcase. It has been accepted for inclusion in
Dissertations by an authorized administrator of Research Showcase. For more information, please contact research-showcase@andrew.cmu.edu.
Optimal Scheduling of Refinery
Crude-Oil Operations
A DISSERTATION
Submitted to the Graduate School
in Partial Fulfillment of the Requirements
for the degree of
Doctor of Philosophy
in
Chemical Engineering
by
Sylvain Mouret

Pittsburgh, Pennsylvania
December, 2010
Acknowledgments
First of all, I would like to express my most sincere gratitude to my advisor Professor Ignacio
E. Grossmann for his inestimable guidance and support over the course of my Ph.D. He has
managed to create a productive yet friendly environment and proved to be an abundant
source of knowledge for myself. I cannot thank him enough for his confidence in me and his
deep implication in my studies and in my life.
Besides my advisor, I would like to thank my thesis committee members – Professors
Lorenz Biegler, Nikolaos Sahinidis, John Hooker, and Willem-Jan van Hoeve for their time
and valuable comments.
I would like to thank Pierre Pestiaux, my supervisor at Total, whose strong commitment
to the project and never-ending enthusiasm has made this thesis possible.
I would also like to thank Philippe Bonnelle for bringing his experience and his insightful
suggestions into the project as well as other collaborators at SOG and CReG, for their useful
feedback on my work and friendly support. Furthermore, I am grateful to Total Refining &
Marketing for financial support of this project.
I wish to express my thankfulness for all my past and present workmates in the PSE group
for setting a productive mood in the office and a diverting atmosphere out of work. Among
them I would like to specifically mention Rosanna Franco, Gonzalo Guillén Gosálbez, Ri-
cardo Lima, Rodrigo López-Negrete de la Fuente, Mariano Martin, Roger Rocha, Sebastian
Terrazas, and Victor Zavala with whom I share many unforgettable memories.
I would also like to thank my fellow football and tennis teammates, Tarot card players,
French speaking lunchers, barbecue grillers, etc... who made my Pittsburgh experience a
very enjoyable one.
I want to express my gratitude to my family who has always been there when I needed
them, and to my 18-month-old niece Anna for being so cute and joyful.
Acknowledgments ii
Last but not least, I cannot thank enough my beloved fiancée Charlotte for her patience
and for standing by me during the past three and a half years. Her unconditional love is
never to be forgotten.
Acknowledgments iii
Abstract
This thesis deals with the development of mathematical models and algorithms for optimiz-
ing refinery crude-oil operations schedules. The problem can be posed as a mixed-integer
nonlinear program (MINLP), thus combining two major challenges of operations research:
combinatorial search and global optimization.
First, we propose a unified modeling approach for scheduling problems that aims at
bridging the gaps between four different time representations using the general concept of
priority-slots. For each time representation, an MILP formulation is derived and strength-
ened using the maximal cliques and bicliques of the non-overlapping graph. Additionally,
we present three solution methods to obtain global optimal or near-optimal solutions. The
scheduling approach is applied to single-stage and multi-stage batch scheduling problems
as well as a crude-oil operations scheduling problem maximizing the gross margin of the
distilled crude-oils.
In order to solve the crude-oil scheduling MINLP, we introduce a two-step MILP-NLP
procedure. The solution approach benefits from a very tight upper bound provided by the
first stage MILP while the second stage NLP is used to obtain a feasible solution.
Next, we detail the application of the single-operation sequencing time representation
to the crude-oil operations scheduling problem. As this time representation displays many
symmetric solutions, we introduce a symmetry-breaking sequencing rule expressed as a
deterministic finite automaton in order to efficiently restrict the set of feasible solutions.
Furthermore, we propose to integrate constraint programming (CP) techniques to the
branch & cut search to dynamically improve the linear relaxation of a crude-oil operations
scheduling problem minimizing the total logistics costs expressed as a bilinear objective.
CP is used to derived tight McCormick convex envelopes for each node subproblem thus
reducing the optimality gap for the MINLP.
Abstract iv
Finally, the refinery planning and crude-oil scheduling problems are simultaneously solved
using a Lagrangian decomposition procedure based on dualizing the constraint linking crude
distillation feedstocks in each subproblem. A new hybrid dual problem is proposed to update
the Lagrange multipliers, while a simple heuristic strategy is presented in order to obtain
feasible solutions to the full-space MINLP. The approach is successfully applied to a small
case study and a larger refinery problem.
Abstract v
Contents
Acknowledgments ii
Abstract iv
Contents vi
List of Tables x
List of Figures xii
1 Introduction 1
1.1 Single-Stage and Multi-Stage Batch Scheduling . . . . . . . . . . . . . . . . 2
1.2 Optimization of Oil Refineries . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.2.1 Refinery Planning . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.2.2 Crude-Oil Operations Scheduling . . . . . . . . . . . . . . . . . . . . 8
1.3 Mixed-Integer Optimization Tools . . . . . . . . . . . . . . . . . . . . . . . 9
1.3.1 Mixed-Integer Linear Programming . . . . . . . . . . . . . . . . . . . 10
1.3.2 Mixed-Integer Nonlinear Programming . . . . . . . . . . . . . . . . . 12
1.3.3 Constraint Programming . . . . . . . . . . . . . . . . . . . . . . . . 13
1.3.4 Lagrangian Relaxation . . . . . . . . . . . . . . . . . . . . . . . . . . 14
1.3.5 Symmetry-Breaking Approaches . . . . . . . . . . . . . . . . . . . . 16
1.4 Overview of Thesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
1.4.1 Chapter 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
1.4.2 Chapter 3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
1.4.3 Chapter 4 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
1.4.4 Chapter 5 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
1.4.5 Chapter 6 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
1.4.6 Chapter 7 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
2 Time Representations and Mathematical Models for Process Scheduling

Problems 19
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
2.2 Case Study . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
2.3 Time Representations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
2.4 Mathematical Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
2.4.1 Sets and Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
2.4.2 Variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
2.4.3 MOS Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
Contents vi
2.4.4 MOS-SST Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
2.4.5 MOS-FST Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
2.4.6 SOS Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
2.5 Strengthened Reformulations . . . . . . . . . . . . . . . . . . . . . . . . . . 33
2.5.1 Non-overlapping Graph Properties . . . . . . . . . . . . . . . . . . . 33
2.5.2 MOS Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
2.5.3 MOS-SST Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
2.5.4 MOS-FST Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
2.5.5 SOS Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
2.6 Solution Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
2.6.1 Additive Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
2.6.2 Multiplicative Approach . . . . . . . . . . . . . . . . . . . . . . . . . 40
2.6.3 Direct Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
2.7 Single-Stage Batch Scheduling Problem . . . . . . . . . . . . . . . . . . . . 42
2.7.1 MOS Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
2.7.2 MOS-SST Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
2.7.3 MOS-FST Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
2.7.4 SOS Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
2.7.5 Models Comparison . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
2.8 Multi-Stage Batch Scheduling Problem . . . . . . . . . . . . . . . . . . . . . 57
2.8.1 MOS Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
2.8.2 MOS-SST Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
2.8.3 MOS-FST Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
2.8.4 Models Comparison . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
2.9 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
3 Short-Term Scheduling of Crude-Oil Operations 67

3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
3.2 Problem Statement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
3.2.1 General Description . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
3.2.2 Case Study . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
3.3 Mathematical Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
3.3.1 Sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
3.3.2 Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
3.3.3 Variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
3.3.4 Objective Function . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
3.3.5 General Constraints . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
3.3.6 Strengthened Constraints . . . . . . . . . . . . . . . . . . . . . . . . 78
3.3.7 Symmetry-Breaking Constraint for MOS Models . . . . . . . . . . . 79
3.3.8 Full Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
3.4 Solution Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
3.5 Computational Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
3.5.1 Scheduling Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
Contents vii
3.5.2 Performance of the MOS Model . . . . . . . . . . . . . . . . . . . . . 85
3.5.3 Performance of the MOS-SST Model . . . . . . . . . . . . . . . . . . 87
3.5.4 Performance of the MOS-FST Model . . . . . . . . . . . . . . . . . . 88
3.5.5 Performance of the MILP-NLP Decomposition Strategy . . . . . . . 89
3.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
4 Single-Operation Sequencing Model for Crude-Oil Operations Scheduling 92

4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
4.2 Strengthened Constraints . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
4.3 Symmetry-Breaking Constraints . . . . . . . . . . . . . . . . . . . . . . . . 95
4.3.1 Symmetric Sequences of Operations . . . . . . . . . . . . . . . . . . 95
4.3.2 A Sequencing Rule Based on a Regular Language . . . . . . . . . . . 95
4.3.3 Rule Derivation for COSP1 . . . . . . . . . . . . . . . . . . . . . . . 97
4.3.4 Regular Constraint . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
4.4.1 Performance of the SOS Model . . . . . . . . . . . . . . . . . . . . . 101
4.4.2 Effect of the Number of Priority-Slots . . . . . . . . . . . . . . . . . 102
4.4.3 Remark on the Optimality of the Solution . . . . . . . . . . . . . . . 103
4.4.4 Effect of Symmetry-Breaking Constraints . . . . . . . . . . . . . . . 105
4.5 Comparison of Crude-Oil Scheduling Models . . . . . . . . . . . . . . . . . . 106
4.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
5 Tightening the Linear Relaxation of a Crude-Oil Operations Scheduling

MINLP Using Constraint Programming 109
5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
5.2 MINLP Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110
5.3 Reformulation and Linear Relaxation . . . . . . . . . . . . . . . . . . . . . . 113
5.4 McCormick Cuts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
5.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118
6 Integration of Refinery Planning and Crude-Oil Scheduling using La-

grangian Decomposition 120
6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120
6.2 Problem Statement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
6.2.1 Refinery Planning Problem . . . . . . . . . . . . . . . . . . . . . . . 121
6.2.2 Crude-Oil Scheduling Problem . . . . . . . . . . . . . . . . . . . . . 125
6.2.3 Full-Space Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
6.3 Lagrangian Decomposition Scheme . . . . . . . . . . . . . . . . . . . . . . . 127
6.4 Solution of the Dual Problem . . . . . . . . . . . . . . . . . . . . . . . . . . 130
6.5 Heuristic Solutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134
6.6 Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136
6.6.1 CDU Feedstocks and Lagrange Multipliers . . . . . . . . . . . . . . . 136
Contents viii
6.6.2 Multi-Period Refinery Planning . . . . . . . . . . . . . . . . . . . . . 137
6.6.3 CDU Feedstocks Aggregation . . . . . . . . . . . . . . . . . . . . . . 138
6.6.4 Handling Nonlinearities in Crude-Oil Scheduling Model . . . . . . . 139
6.6.5 Handling Nonlinearities in the Refinery Planning Model . . . . . . . 140
6.6.6 Detailed Implementation . . . . . . . . . . . . . . . . . . . . . . . . . 140
6.7 Numerical Illustration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142
6.8 Larger Refinery Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148
6.9 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 154
7 Conclusion 156
7.1 Time Representations and Mathematical Models . . . . . . . . . . . . . . . 156
7.2 Short-Term Scheduling of Crude-Oil Operations . . . . . . . . . . . . . . . . 159
7.3 Single-Operation Sequencing Model for Crude-Oil Operations Scheduling . 160
7.4 Tightening the Linear Relaxation of an MINLP Using CP . . . . . . . . . . 161
7.5 Integration of Refinery Planning and Crude-Oil Scheduling . . . . . . . . . 163
7.6 Contributions of the Thesis . . . . . . . . . . . . . . . . . . . . . . . . . . . 164
7.7 Recommendations for Future Work . . . . . . . . . . . . . . . . . . . . . . . 165
8 Bibliography 168
Appendices 177
A On Tightness of Strengthened Constraints 179
B Crude-Oil Operations Scheduling Examples 181
C Mathematical Models for Crude-Oil Operations Scheduling Problems 185

C.1 MOS Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 185
C.2 MOS-SST Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 186
C.3 MOS-FST Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 187
C.4 SOS Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 188
D Mathematical Model for the Refinery Planning Problem 189
Contents ix
List of Tables
1.1 Optimization techniques used in different MINLP solvers. . . . . . . . . . . 12
2.1 Resource requirements for case study. . . . . . . . . . . . . . . . . . . . . . 22

2.2 Time representations nomenclature. . . . . . . . . . . . . . . . . . . . . . . 27
2.3 Data for single-stage batch scheduling problems. . . . . . . . . . . . . . . . 43
2.4 Unit cardinality bounds depending on parameter n for SSBSP29. . . . . . . 46
2.5 MOS computational results for single-stage batch scheduling problems. . . . 49
2.6 MOS-SST computational results for single-stage batch scheduling problems. 52
2.7 MOS-FST computational results for single-stage batch scheduling problems. 54
2.8 Data for multi-stage batch scheduling problems. . . . . . . . . . . . . . . . . 58
2.9 MOS computational results for multi-stage batch scheduling problems. . . . 61
2.10 MOS-SST computational results for multi-stage batch scheduling problems. 63
2.11 MOS-FST computational results for multi-stage batch scheduling problems. 65
3.1 Data for COSP1. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71

3.2 MOS computational results for crude-oil scheduling problems. . . . . . . . . 87
3.3 MOS-SST computational results for crude-oil scheduling problems. . . . . . 88
3.4 MOS-FST computational results for crude-oil scheduling problems. . . . . . 89
3.5 Performance of different MINLP algorithms for crude-oil scheduling problems. 90
4.1 Maximal cliques and bicliques for COSP2 and COSP3. . . . . . . . . . . . . 93

4.2 Cliques and bicliques selections a, b, and c for COSP2 and COSP3. . . . . . 94
4.3 List of sequences belonging to regular language L7 . . . . . . . . . . . . . . . 98
4.4 SOS computational results for crude-oil operations scheduling problems. . . 101
4.5 Size and performance of Basic and Extended models on COSP1 (13 slots). . 106
4.6 Size of MOS, MOS-SST, MOS-FST, and SOS models for crude-oil scheduling
problems. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
5.1 Cost data for crude-oil operations scheduling problems. . . . . . . . . . . . 113

5.2 Results obtained with BasicRelaxation and ExtendedRelaxation algorithms. 118
5.3 Results obtained with diferrent MINLP algorithms on COSP1 and COSP2. 119
6.1 Crude-oil scheduling data for case study. . . . . . . . . . . . . . . . . . . . . 126

6.2 Lagrangian iterations statistics (6 priority-slots, NLP=SNOPT). . . . . . . 142
6.3 Lagrangian iterations statistics (7 priority-slots, NLP=SNOPT). . . . . . . 143
6.4 Comparative performance of different MINLP algorithms. . . . . . . . . . . 146
6.5 Crude cut prices and specification for larger refinery problem. . . . . . . . . 150
6.6 Crude-oil scheduling data for larger refinery problem. . . . . . . . . . . . . . 151
List of Tables x
6.7 Lagrangian iterations statistics for larger refinery problem (6 priority-slots,
NLP=CONOPT). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152
6.8 Optimal Lagrange multipliers for each crude and each CDU mode. . . . . . 153
6.9 Comparative performance of several MINLP algorithms for larger refinery
problem (NLP solver: CONOPT). . . . . . . . . . . . . . . . . . . . . . . . 153
6.10 Blend compositions in the optimal solution of larger refinery problem. . . . 154
B.1 Data for COSP1. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 181

B.2 Data for COSP2. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 182
B.3 Data for COSP3. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 183
B.4 Data for COSP4. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 184
List of Tables xi
List of Figures
1.1 A typical multi-stage batch process from Pinto and Grossmann (1995). . . . 3
1.2 A typical oil refining process from Méndez et al. (2006b). . . . . . . . . . . 6
1.3 Schematic flow diagram of a typical oil refinery from Wikipedia (2010). . . 7
1.4 Crude-oil scheduling problem 1 from Lee et al. (1996). . . . . . . . . . . . . 9
2.1 Four steps optimization approach. . . . . . . . . . . . . . . . . . . . . . . . 20

2.2 Non-overlapping matrix and graph for case study. . . . . . . . . . . . . . . . 23
2.3 A unique schedule obtained through different time representations. . . . . . 24
2.4 Biclique ({v1 , v6 }; {v4 , v5 }). . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
2.5 Assignment constraint using consecutive time-points. . . . . . . . . . . . . . 37
2.6 Non-overlapping graph with isolated cliques for SSBSP8. . . . . . . . . . . . 44
2.7 Effect of the minimum priority-slot usage constraint. . . . . . . . . . . . . . 50
2.8 Equivalent MOS and SOS assignments for SSBSP12. . . . . . . . . . . . . . 55
2.9 Comparison of time representations for single-stage batch scheduling problems. 56
2.10 Partial non-overlapping graph with isolated cliques for MSBSP5. . . . . . . 59
2.11 Comparison of time representations for multi-stage batch scheduling problems. 65
3.1 Example of tank schedule. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70

3.2 Sub-optimal schedule for COSP1 (profit: $6,925,000). . . . . . . . . . . . . 72
3.3 Optimal schedule for COSP1 (profit: $7,975,000). . . . . . . . . . . . . . . . 73
3.4 Refinery crude-oil scheduling system for problem COSP2 and COSP3. . . . 79
3.5 Non-overlapping graph for crude-oil examples 2 and 3. . . . . . . . . . . . . 80
3.6 Two step decomposition strategy. . . . . . . . . . . . . . . . . . . . . . . . . 81
3.7 Optimal schedule for COSP2 (profit: $10,117,000). . . . . . . . . . . . . . . 83
3.8 Schedule obtained for COSP3 within 2.3% optimality gap (profit: $8,540,000). 84
3.9 Optimal schedule for COSP2 with late vessel arrivals (profit: $9,775,000). . 85
3.10 Optimal schedule for COSP2 with late vessel arrivals and fixed initial deci-
sions (profit: $9,609,000). . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
4.1 Symmetric sequences of operations for COSP1. . . . . . . . . . . . . . . . . 96

4.2 Automaton DFA7 recognizing regular language L7 . . . . . . . . . . . . . . . 98
4.3 Automaton recognizing the regular language L. . . . . . . . . . . . . . . . . 99
4.4 Performance of the SOS model on crude-oil scheduling problems (MILP
solver: Xpress). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
4.5 Performance of the Basic and Extended models on COSP1 (6 to 13 slots). . 106
4.6 Comparison of time representations for crude-oil scheduling problems. . . . 107
5.1 Branch & cut algorithm with McCormick cuts. . . . . . . . . . . . . . . . . 116
List of Figures xii

6.1 Basic refinery planning system. . . . . . . . . . . . . . . . . . . . . . . . . . 122
6.2 Refinery planning case study. . . . . . . . . . . . . . . . . . . . . . . . . . . 124
6.3 Refinery crude-oil scheduling system for COSP1. . . . . . . . . . . . . . . . 125
6.4 Economic interpretation of the Lagrangian decomposition. . . . . . . . . . . 130
6.5 General iterative primal-dual algorithm. . . . . . . . . . . . . . . . . . . . . 132
6.6 Plots of the feasible space of (P̂DK+1 ). . . . . . . . . . . . . . . . . . . . . . . 133
6.7 Iterative primal-dual algorithm with heuristic step. . . . . . . . . . . . . . . 135
6.8 Crude-oil scheduling and multi-period refinery planning integration. . . . . 137
6.9 Disaggregated CDU feedstocks synchronization. . . . . . . . . . . . . . . . . 138
6.10 Complete algorithm implementation. . . . . . . . . . . . . . . . . . . . . . . 141
6.11 Lagrangian iteration objective values (6 priority-slots, NLP=SNOPT). . . . 144
6.12 Lagrangian iteration objective values (7 priority-slots, NLP=SNOPT). . . . 144
6.13 Lagrange multiplier updates (6 priority-slots, NLP=SNOPT). . . . . . . . . 145
6.14 Lagrange multiplier updates (7 priority-slots, NLP=SNOPT). . . . . . . . . 145
6.15 Blend compositions in solutions obtained with 6 priority-slots. . . . . . . . . 147
6.16 Planning model for larger refinery problem. . . . . . . . . . . . . . . . . . . 149
6.17 Refinery crude-oil scheduling system for COSP3. . . . . . . . . . . . . . . . 150
6.18 Lagrangian iteration objective values for larger refinery problem (6 priority-
slots, NLP=CONOPT). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152
B.1 Refinery crude-oil scheduling system for COSP1. . . . . . . . . . . . . . . . 181

B.2 Refinery crude-oil scheduling system for COSP2 and COSP3. . . . . . . . . 182
B.3 Refinery crude-oil scheduling system for COSP4. . . . . . . . . . . . . . . . 183
D.1 Layered artificial neural network. . . . . . . . . . . . . . . . . . . . . . . . . 190
List of Figures xiii

Chapter 1
Introduction
Optimization in the oil refining industry began with the use of linear programming (LP) to
perform process and economic analysis of industrial plants (see Garvin et al., 1957; Manne,
1958). Many refinery problems are now addressed with algorithms based on mathematical
models: refinery planning, crude-oil operations scheduling, final product blending, crude-oil
transportation, final product shipping, and profitability improvement plans (for instance,
see Pinto et al., 2000). In general, problem-specific techniques are used to solve each model
independently from the others. The goal of this thesis is to develop a general methodology
towards enterprise-wide optimization of oil refineries (Grossmann, 2005). Due to the struc-
tural diversity of the problems to be solved, the challenge is to effectively integrate different
optimization techniques in order to generate near-optimal enterprise-wide solutions. The
industrial applications addressed in this thesis are related to single-stage and multi-stage
batch processes, medium-term planning of refining operations and short-term scheduling of
crude-oil operations. The objectives of the thesis are as follows:
1. Develop a unified modeling approach for solving process scheduling problems

2. Apply the proposed time representations to schedule and optimize batch processes
and crude-oil operations
3. Develop and implement general solution methods to effectively solve such problems
and obtain near-optimal solutions with rigorous optimality estimates
4. Develop a method for integrating mixed-integer linear programming and constraint
programming for improving the linear relaxation of mixed-integer nonlinear programs
5. Apply advanced Lagrangian decomposition techniques to simultaneously solve refinery
planning and crude-oil operations scheduling problems
Chapter 1. Introduction 1
1.1 Single-Stage and Multi-Stage Batch Scheduling
In this chapter, an overview of single-stage and multi-stage batch scheduling problems

is presented followed by a description of the refinery planning and crude-oil scheduling
problems. The different optimization techniques used are then reviewed and we conclude
with an overview of the chapters in the thesis.
The chemical industry has been marked by an increase of product diversification, which
in turn has led to an increase in the complexity of operations of plant facilities. Chemical
companies are now facing the challenge of meeting global demands of multiple products
while increasing plant capacities to achieve economies of scale (Wassick, 2009). Operating
optimally such plants can be non-trivial as decision-makers have to account for demand
deadlines, process constraints, and limited resources. Therefore, the scheduling of chemical
processes has received much attention over the past 20 years. Two major categories of
processes have been outlined and addressed: sequential and network-based processes (see
process classification in Méndez et al., 2006a). The main difference lies in the fact that
network processes may display recycle loops which sequential processes do not.
Recently, several research groups have reviewed the different trends in process scheduling
for general purpose plants. Floudas and Lin (2004) provided an extensive comparison
of discrete-time and continuous-time formulations. In Méndez et al. (2006a), a complete
classification of scheduling approaches is presented in addition to the process classification.
In this thesis, we study single-stage and multi-stage batch scheduling problems. Figure 1.1
depicts a typical multi-stage batch process. A finite number of batches with fixed sizes have
to be processed going through a given set of successive stages. In each stage, a finite set of
parallel units is available. The processing times for each batch and each stage may be unit-
dependent. Stages with limited resources or high processing times are often called bottleneck
stages. Different policies can be used for interstage storage: unlimited intermediate storage
(UIS), finite intermediate storage (FIS), no intermediate storage (NIS), and zero-wait (ZW).
1
10
2 11 20 23
7
12
3 21 24
13
4 14
15
16
5 8 17 22 25
18
6 9
19
STAGE Reaction Fluidization Standardization Drying Packing
Figure 1.1: A typical multi-stage batch process from Pinto and Grossmann (1995).
Two main scheduling approaches have been developed to solve batch scheduling problems:
precedence-based and slot-based models.
The idea of using disjunctive programming principles (see Balas, 1985) to solve chemical
process scheduling problems was initially introduced by Cerdá et al. (1997) who proposed a
mathematical model for scheduling single-stage multiproduct batch plants. Among others,
Gupta and Karimi (2003); Méndez and Cerdá (2007); Marchetti and Cerdá (2009) have
successively improved and extended precedence-based formulations and applied them to
multistage batch scheduling problems. They consider complex scheduling features such as
sequence-dependent changeovers, unit release times, or discrete resource constraints. The
basic idea is to use binary variables to represent the precedence relations between each pair
of operations. The corresponding models usually involve many big-M constraints, which
may result in poor LP relaxations. However, these formulations lead to models of modest
size which often makes them tractable. In principle, these formulations can be used to
represent any type of scheduling constraints, but global features such as limited inventory
or discrete resource constraints, which may involve more than two operations, are non-trivial
and often require additional big-M constraints.

Another scheduling approach consists of using the concept of slots in order to assign
a position for each operation in a sequence. Depending on the formulation used, these
positions are directly or indirectly linked to the timing decisions. These types of formulations
are particularly efficient to represent global scheduling features as its inherent sequencing
representation may involve as many operations as needed.
Ku and Karimi (1988) presented an MILP formulation for multi-stage batch scheduling
problem with finite interstage storage and exactly one unit per stage. Two heuristics are
presented in order to solve the problem faster and compared to the optimal MILP approach.
Pinto and Grossmann (1995) developed a continuous-time slot-based mathematical for-
mulation and considered a batch preordering heuristic to improve the computational re-
quirements. This approach was applied to several examples with unit-dependent processing
times and changeovers.
Hui et al. (2000) addressed the issue of sequence-dependent changeovers using a simi-
lar slot-based MILP model. Later, Gupta and Karimi (2003) improved the mathematical
formulation using fewer variables and constraints to reduce the computational times.
Castro and Grossmann (2005) used the resource-task network representation (RTN) to
model multi-stage batch scheduling problem. They developed and implemented several
scheduling approaches and presented extensive computational results.
Recently, Prasad and Maravelias (2008) solved the integrated batching and scheduling
problem. It consists of simultaneously making the following decisions: batch selection and
sizing, unit assignment and sequencing of batches on each unit.
Finally, it should be noted that models for structures others than multi-stage batch plants
have also been extensively studied using representations as the state-task network (Kondili
et al., 1993) and resource-task network (Pantelides, 1994).
1.2 Optimization of Oil Refineries
Oil refineries are a key element in the valorization of crude-oils into energy. They consist of
highly flexible plants that can refine crude-oils produced from many locations in the world,
with very different properties, into useful petroleum products: LPG, gasoline and diesel
fuels, kerosene, heating oil, bitumen, etc.
As depicted in Figure 1.2, the refining process can be decomposed into 3 main phases:
crude-oil unloading and preparation for distillation, fractionation and reaction operations,
final product blending and shipping. Based on this spatial decomposition, the following
off-line optimization problems have been addressed in the literature:
1. Refinery planning (see section 1.2.1)

2. Crude-oil operations scheduling (see section 1.2.2)
3. Final product blending operations scheduling with recipe optimization (see Méndez
et al., 2006b)
4. Scheduling of internal refinery product-specific subsystems:
- for LPG, see Pinto and Moro (2000)

- for fuel oils and asphalts, see Joly and Pinto (2003)
- for lube oils and paraffins, see Casas-Liza and Pinto (2005)
Given the diversity and the complexity of the subsystems to be studied, this thesis is focused
on the integration of the first two refinery optimization problems.
1.2.1 Refinery Planning
Refinery operators typically determine their annual and monthly plan by solving the refinery
planning problem. It is based on a steady-state flowsheet optimization of a plant with
multiple heterogenous units (see Fig. 1.3). The objective is to maximize the plant’s net
present value determined by sales revenues minus purchase and operating costs. In addition
to process constraints, the problem also considers crude availabilities as well as product
demands and specifications.
Figure 1.2: A typical oil refining process from Méndez et al. (2006b).
The refinery planning problem was one of the first industrial applications of linear pro-
gramming (Bodington and Baker, 1990). However, the solution methods have evolved
towards successive linear programming (SLP) in order to better account for the nonlinear
nature of the refining process. In particular, the nonlinearities in the refinery model arise
from pooling equations and advanced process models.
The pooling problem has received much attention in the literature since the 70’s. It usu-
ally consists of optimizing feedstocks purchases, product blending operations and product
sales while taking into account product availabilities, product demands, property specifi-
cations, and pool capacities. Haverly (1978); Hart (1978); Haverly (1980) developed and
experimented the distributive recursion approach, a technique proved to be equivalent to
classical SLP (see Lasdon and Joffe, 1990) in order to solve it.
Several authors have successively addressed the issue of global optimization of pure pool-
ing problems. The techniques used range from generalized Benders decomposition (Floudas
and Aggarwal, 1990), Lagrangian relaxation (Floudas and Visweswaran, 1993; Adhya et al.,
1999), or spatial Branch and Bound (Foulds et al., 1992; Quesada and Grossmann, 1995a;
Audet et al., 2000) to reformulation-linearization techniques (Audet et al., 2000; Meyer and
Floudas, 2006), or piecewise convex relaxations (Meyer and Floudas, 2006; Pham et al.,
Figure 1.3: Schematic flow diagram of a typical oil refinery from Wikipedia (2010).
2009). Extensive reviews of pooling formulations and solution methods can be found in
Audet et al. (2004); Misener and Floudas (2009).
Some examples of nonlinear refinery planning problems including pooling constraints
and nonlinear process models can be found in Pinto and Moro (2000); Li et al. (2005);
Alhajri et al. (2008). Although commercial solvers such as GRTMPS (Haverly Systems),
PIMS (Aspen Tech), and RPMS (Honeywell Hi-Spec Solutions) implement successive linear
programming algorithms to solve this problem (see Zhang et al., 1985), any standard NLP
solvers can also be used, although they may not guarantee global optimality of the solution.
A major issue with refinery planning is that most models are single-period models in
which the refinery is assumed to operate in the same state over the whole planning period
(typically 1 month). Therefore, the planning solution is used as a tactical goal for refinery
operators rather than as an operational tool. In particular, crude distillation unit (CDU)
feedstock decisions returned by the refinery planning problem are usually not applicable in
the field due to crude logistics constraints. These are described in the crude-oil operations
scheduling problem.
1.2.2 Crude-Oil Operations Scheduling
The optimal scheduling of crude-oil operations have been studied since the 90’s and has
been shown to lead to multimillion dollar benefits by Kelly and Mann (2003) as it is the
first stage of the oil refining process. It involves crude-oil unloading from crude marine
vessels (at berths or jetties), or from a pipeline to storage tanks, transfers from storage
tanks to charging tanks and atmospheric distillations of crude-oil mixtures from charging
tanks (see Fig. 1.4). The crude is then processed in order to produce basic products which
are then blended into finished products (see section 1.2.1). In this thesis, we assume that
the schedule of the crude supply is given. The production demands are determined by
the long-term refinery planning, either sequentially (chapters 3, 4 and 5) or simultaneously
(chapter 6). The following objectives are considered:
1. Maximization of crude gross margins (chapters 3 and 4)

2. Minimization of total logistics costs (chapter 5)
3. Maximization of total refinery profit as expressed in the planning problem (chapter 6)
Shah (1996) proposed to use mathematical programming techniques to find crude-oil

schedules exploiting opportunities to increase economic benefits. Lee et al. (1996) con-
sidered a crude-oil scheduling problem involving crude unloading at berths, and devel-
oped a discrete-time MINLP model, and solved an MILP relaxation of the model. Later,
1.3 Mixed-Integer Optimization Tools
Crude Vessels Storage Tanks Charging Tanks CDU

1 3
4 7
5 8
2 6
Figure 1.4: Crude-oil scheduling problem 1 from Lee et al. (1996).
Wenkay et al. (2002) improved the model and proposed an iterative approach to solve the
MINLP model, taking into account the nonlinear blending constraints.
Pinto et al. (2000), Moro and Pinto (2004), and Reddy et al. (2004) used a global event
formulation to model refinery systems involving crude-oil unloading from pipeline or jetties.
The scheduling horizon is divided into fixed length sub-intervals, which are then divided in
several variable length time-slots.
In parallel, Jia et al. (2003) developed an operation specific event model and applied it
to the problems introduced by Lee et al. (1996) using a linear approximation of storage
costs. A comparison of computational performance between both continuous-time and
discrete-time models was given showing significant decreases in CPU time for the former
model. Also, solutions that are not guaranteed to be globally optimal were obtained using
standard MINLP algorithms.
Recently, Furman et al. (2007) presented a more accurate version of the event-point
formulation, and Karuppiah et al. (2008) later addressed the global optimization of this
model using an outer-approximation algorithm where the MILP master problem is solved
by adding cuts from a Lagrangian decomposition. While rigorous, this method can be
computationally expensive.
Scheduling problems are among the most challenging optimization problems, both in terms
of modeling and solution algorithm. Mostly mixed-integer linear programming (MILP,
see Kallrath, 2002), constraint programming (CP, see Baptiste et al., 2001) and genetic
algorithm (GA, see Mitchell, 1998) techniques have been used to tackle these problems.
CP has proved to be very efficient for solving scheduling problems but it is rarely used to
solve problems arising in the chemical engineering field. One of the reason is that CP is
very efficient at sequencing tasks or jobs that are defined a priori (e.g. job-shop problems
in discrete manufacturing). However, the scheduling of chemical processes usually involves
both defining and sequencing the tasks that should be performed. Defining tasks means
choosing a batch size or a unit operating mode for example. As a consequence, LP based
techniques have been preferred with formulations essentially based on time grids as it easily
allows modeling tank or unit capacity at the end of each time interval (see Floudas and Lin,
2004; Méndez et al., 2006b).
In this thesis, we aim at solving optimization problems involving both continuous and
discrete decisions. Many computational techniques have emerged from the area of mixed-
integer optimization in order to solve problems with different characteristics (linear, non-
linear, convex, non-convex, purely integer, ...), including :
1. mixed-integer linear programming (MILP)

2. mixed-integer nonlinear programming (MINLP)
3. constraint programming (CP)
4. Lagrangian relaxation
In this section, we present brief reviews of these four optimization techniques as well as
symmetry-breaking approaches.
1.3.1 Mixed-Integer Linear Programming
Mixed-integer linear programming is used to model many decision problems from industry
(for example, process scheduling, production planning, resource allocation and supply chain
management). However, solving such combinatorial models is NP-hard (see Nemhauser and
Wolsey, 1999). Therefore, several approaches have been developed and combined in order
to solve these problems in reasonable times. Two main techniques have emerged: branch-
and-bound (Land and Doig, 1960) and cutting planes (Gomory, 1958). Both are based on
the LP relaxation of the MILP. This relaxation is obtained by considering integer variables
as continuous variables with identical bounds: it can also be called continuous relaxation
or polytope relaxation.
The branch-and-bound technique consists of searching through a tree defined by succes-
sive assignments of integer values to integer variables. At each node of this tree, an LP is
solved in order to obtain a local node optimality bound (e.g. upper bound for maximiza-
tion problems); the global optimality bound is the best bound among all open nodes (i.e.
unprocessed nodes). If the solution of this LP is integral, it provides a global feasibility
bound (e.g. lower bound for maximization problems). The search continues until all nodes
have been processed (0% optimality gap) or until the optimality gap is below a specified
tolerance.
The cutting plane algorithm consists of iteratively solving the LP relaxation and gen-
erating additional constraints (called cutting planes or cuts) that cut off the current LP
solution. The procedure is stopped when the LP solution is integral. Although this can be
achieved in a finite number of steps, in practice a large number of iterations are required.
The branch-and-cut procedure is a combination of the two aforementioned techniques.
The cutting plane algorithm is used to tighten the LP relaxation at each node, thus im-
proving the optimality bound (local and potentially global too). Branching occurs whenever
the optimality bound cannot be significantly improved.
Recent reviews of MILP techniques can be found in Bixby et al. (1999); Johnson et al.
(2000). The best known MILP solvers (CPLEX, Xpress, and Gurobi) all implement the
branch-and-cut procedure and are widely used to solve industrial large-scale mixed-integer
problems.
Table 1.1: Optimization techniques used in different MINLP solvers.

SBB and
MINLP solvers DICOPT AlphaECP Bonmin KNITRO
MINLP BB
branch-and-bound x x x
outer-approximation x x
LP/NLP based branch x x
and bound
extended cutting plane x
1.3.2 Mixed-Integer Nonlinear Programming
A number of industrial applications of mixed-integer optimization include nonlinearities

in the objective function or in the constraints. Except for specific cases (mixed-integer
quadratic programming, mixed-integer quadratically constrained programming or mixed-
integer second-order cone programming), standard MILP techniques cannot be used directly
to solve such problems. Therefore, many optimization techniques have been developed to
solve general MINLPs (for review, see Grossmann, 2002), including:
1. NLP-based branch-and-bound (Leyffer, 2001)

2. outer-approximation (Duran and Grossmann, 1986)
3. LP/NLP based branch and bound (Quesada and Grossmann, 1992)
4. extended cutting plane (Westerlund and Pettersson, 1995)
Table 1.1 summarizes the different techniques used in standard MINLP solvers available in
GAMS.
A challenging MINLP topic is global optimization of non-convex problems, that is MINLPs
with non-convex NLP subproblems. Similarly to MILP optimization, all approaches use a
convex relaxation of the MINLP and rely on a spatial branch-and-bound search. These
global optimization techniques include:
1. branch-and-reduce (Tawarmalani and Sahinidis, 2004)

2. α-BB (Adjiman et al., 2000)
3. spatial branch-and-bound for bilinear and linear fractional terms (Quesada and Gross-
mann, 1995a)
4. outer-approximation (Kesavan et al., 2004)
The main global MINLP solvers are BARON (Sahinidis, 1996), which implements an ad-
vanced branch-and-reduce algorithm, and LINDOGlobal (Lin and Schrage, 2009), which
also implements a spatial branch-and-bound procedure. The global optimization of large-
scale industrial non-convex MINLP problems is still unachievable in most cases. However,
in some cases, recent developments in this area can be used to generate good heuristic
solutions to such problems and provide tight global optimality estimates for these solutions.
1.3.3 Constraint Programming
Constraint programming is an alternative optimization approach to classical Operations

Research (OR) techniques that is widely used to solve combinatorial problems such as
scheduling (Baptiste et al., 2001), timetabling (Goltz and Matzke, 1998) or vehicle routing
problems (Shaw, 1998). It is based on variable domain filtering algorithms, constraint
propagation techniques and tree search heuristics. A central tool in constraint programming
is the domain store (also called constraint store), which is used to record the domain of
each variable in the model. At each node, the constraint propagation procedure iteratively
calls each constraint domain filtering algorithm and updates the local domain store. If all
variable domains are eventually reduced to singletons, a new solution if found, otherwise
branching is used and a new node is selected. General reviews of constraint programming
techniques can be found in van Hentenryck (1989); Rossi et al. (2006).
Many solvers implements these algorithms including Ilog CP, Choco, Gecode, and Comet
(Michel and van Hentenryck, 2003). Recent developments in the CP community aim at
improving filtering algorithms for global constraints (Régin, 2003; van Hoeve et al., 2009),
integrating OR techniques with CP (Milano and Wallace, 2006; Hooker, 2007; Yunes et al.,
2010), and replacing the traditional domain store by multi-valued decision diagrams (MDD)
to increase the amount of propagation between constraints (Andersen et al., 2007; Hoda
et al., 2010).
The use of CP to solve optimization problems with continuous decisions, although theoret-
ically achievable, has not yet received much attention and often relies on the discretization
of the domain of the continuous variables. As a consequence, it is not widely used to solve
industrial problems in the chemical industry. However, classical OR techniques can benefit
from CP filtering algorithms and constraint propagation, which can be less computationally
expensive than solving large strengthened LP models.
1.3.4 Lagrangian Relaxation
Lagrangian relaxation is a relaxation technique which aims at solving mathematical models

including complicating or hard constraints. It consists of two major elements: a primal
relaxation and a dual algorithm. The primal relaxation is obtained by transferring the
complicating constraints into the objective function, scaled by a penalty factor, specifically
the Lagrange multiplier. Given a Lagrange multipliers λ ∈ Rm 2
+ (m2 being the number of
complicating constraints), this transformation can be described as follows:



 max cT x 

  max cT x + λT (b2 − A2 x)
(1) s.t. A1 x ≤ b1 ⇒ (2)
  s.t. A x ≤ b

 1 1
A2 x ≤ b2

Problem (2) is a relaxation of (1) as its optimal objective value is always greater than the
optimal objective value of (1).
Given this primal relaxation, the dual algorithm aims at solving the following problem:

 max cT x + λT (b2 − A2 x)
(3) minm
λ∈R+ 2  s.t. A x ≤ b
1 1
In order to solve this dual problem, the following techniques can be used:
1. subgradient method (Held and Karp, 1971; Fisher, 1981)

2. cutting plane procedure (Cheney and Goldstein, 1959; Kelley, 1960)
3. boxstep method (Marsten et al., 1975)

4. bundle method (Lemaréchal, 1974)
5. volume algorithm (Barahona and Anbil, 2000)
6. analytic center cutting plane method (Goffin et al., 1992).
As explained in Frangioni (2005), problem (3) is equivalent to the following convexified

version of the original problem:



 max cT x

(4) s.t. conv(A1 x ≤ b1 )



A2 x ≤ b2

Also, the optimal Lagrange multipliers determined by solving problem (3) correspond to the
marginal values of the constraint A2 x ≤ b2 in an optimal solution of problem (4). Clearly,
if all variables are continuous (x ∈ Rn ), problem (3) is therefore equivalent to problem
(1). However, if some variables are integer, problem (3) yields a relaxation of problem (1).
The difference between their respective optimal values is called a dual gap. In general, the
Lagrangian relaxation is tighter than the LP relaxation, but it is more difficult to obtain.
Although the Lagrangian relaxation technique have been developed for LPs or MILPs, it
can also be used to solve difficult MINLPs (for example, see Neiro and Pinto, 2006).
A key issue in Lagrangian relaxation techniques is the choice of the complicating con-
straints. There is a classic tradeoff between making the relaxed problem (2) easy to solve
and reducing the dual gap. In some cases, the complicating constraints are selected in
order to make the relaxed problem (2) decomposable and therefore much easier to solve.
This technique is called Lagrangian decomposition. The reader may refer to Fisher (1985)
and Guignard (2003) for extensive reviews on Lagrangian relaxation and decomposition
techniques.
1.4 Overview of Thesis
1.3.5 Symmetry-Breaking Approaches
The effectiveness of the optimization techniques previously mentioned often suffers from the
presence of multiple symmetric solutions. In simple words, we define symmetric solutions
as feasible solutions that provide identical practical decisions from the modeler’s perspec-
tive. In particular, symmetric solutions have identical objective values. The presence of
symmetries usually leads to an exhaustive enumeration of many feasible solutions which are
not detected as symmetric by the solver. An extensive review of symmetry detection and
symmetry-breaking approaches can be found in Margot (2008). In the context of this thesis,
we focus on problem-specific symmetry detection and static symmetry-breaking constraints.
1.4.1 Chapter 2
In chapter 2, a unified modeling approach for solving process scheduling problems is pro-
posed. Four different time representations that are based on priority-slots are presented and
compared by deriving the relations between them. For each time representation, a mathe-
matical model is presented and strengthened using the maximum cliques and bicliques of the
non-overlapping graph. We introduce three solution methods that can be used to achieve
global optimality or obtain near-optimal solutions depending on the stopping criterion used.
The proposed approaches are applied to single-stage and multi-stage batch scheduling prob-
lems. Computational results show that the multi-operation sequencing (MOS) time rep-
resentation is superior to the others as it allows efficient symmetry-breaking and requires
fewer priority-slots, thus leading to smaller model sizes.
1.4.2 Chapter 3
In chapter 3, the crude-oil operations scheduling problem is stated and the four time rep-
resentations introduced in chapter 2 are used to derive strengthened MINLP models for
solving this problem. In particular, it is shown how the non-overlapping graph can be used
to generate non-trivial strengthened constraints for a specific example. Computational re-
sults are obtained for all time representations except for single-operation sequencing (SOS).
A two-step MILP-NLP procedure is used to solve the non-convex MINLP models leading
to an optimality gap lower than 4% in all cases.
1.4.3 Chapter 4
In chapter 4, we apply the single-operation sequencing (SOS) time representation introduced

in chapter 2 to the crude-oil scheduling problems presented in chapter 3. The corresponding
MINLP model is based on the representation of a crude-oil schedule by a single sequence of
transfer operations. Therefore, it is possible to reduce the symmetries involved in the prob-
lem using a deterministic finite automaton to represent a symmetry-breaking sequencing
rule. Computational results show the effectiveness of the symmetry-breaking approach. A
final comparison of all time representations applied to the crude-oil scheduling operations
problem is presented showing the superiority of the MOS model.
1.4.4 Chapter 5
Chapter 5 aims at tightening the linear relaxation of a refinery crude-oil operation scheduling
MINLP based on the single-operations sequencing (SOS) time representation. The model
is mostly linear but contains bilinear products of continuous variables in the objective
function (minimization of total logistics costs). It is possible to define a linear relaxation
of the model leading to a weak bound on the objective value of the optimal solution. A
typical method to address this issue is to discretize the continuous space and to use linear
relaxation constraints based on lower and upper bounds of the variables (e.g. McCormick
convex envelopes, see McCormick, 1976) on each subdivision of the continuous space. This
work explores another approach involving constraint programming (CP). The idea is to use
an additional CP model which is used to tighten the bounds of the continuous variables
involved in bilinear terms and then generate cuts based on McCormick convex envelopes.
These cuts are then added to the mixed-integer linear program (MILP) during the search
leading to a tighter linear relaxation of the MINLP. Results show large reductions of the
optimality gap in the two-step MILP-NLP solution method introduced in chapter 3 due to
the tighter linear relaxation obtained.
1.4.5 Chapter 6
In chapter 6, we introduce a new methodology to solve a large-scale mixed-integer nonlinear

program (MINLP) integrating the two major optimization problems appearing in the oil
refining industry: refinery planning and crude-oil operations scheduling. The proposed
approach consists of using Lagrangian decomposition to effectively integrate both problems.
The main advantage of this technique is that each problem can be solved independently.
A new hybrid dual problem is introduced to iteratively update the Lagrange multipliers.
It uses the classical concepts of cutting planes, subgradient, and boxstep. The proposed
approach is compared to a basic sequential approach and to standard MINLP solvers. The
results obtained on a case study and a larger refinery problem show that the Lagrangian
decomposition algorithm is more robust than the other approaches and produces better
solutions in reasonable times.
1.4.6 Chapter 7
Chapter 7 summarizes the major contributions of the thesis. The conclusions are discussed
with suggestions for future work.
This thesis has led to the following papers: Mouret et al. (2008, 2009a,b, 2010a,b)
Chapter 2
Time Representations and Mathematical
Models for Process Scheduling Problems
2.1 Introduction
Rigorous optimization of real-world problems are often based on advanced optimization

tools such as mixed-integer linear programming (MILP) or constraint programming (CP,
see Rossi et al., 2006). These tools rely on a mathematical or symbolic representation of
the problem that is applied by an end-user. In some cases, the relationship between the
problem description and its mathematical model is not clear. Therefore an intermediate step
is included in the optimization approach (see Figure 2.1). In this step, the representation
used is detailed and approximations are made. For instance, in the context of scheduling
problems, using a discrete-time formulation is in general a constraining approximation of
the actual problem, and thus, it may lead to a suboptimal solution as discussed in Floudas
and Lin (2004).
Additionally, it is important to note that several mathematical models may be used
to obtain the global optimal solution of the problem, which is the best possible solution
according to a given optimization criterion. For example, many time representations rely
on a specific parameter representing the number of time points (Kondili et al., 1993), time
intervals (Lee et al., 1996), or event points (Ierapetritou and Floudas, 1998). Therefore, the
scheduling problem is represented by an infinite set of mathematical models, one for each
possible value of this parameter (all positive integers). The global optimal schedule is the
best solution among the optimal solution of all these models. In general, it is not possible
Chapter 2. Time Representations and Models for Scheduling Problems 19

2.1 Introduction
Problem Description
Problem Representation
Mathematical Model(s)
Solution Method
Figure 2.1: Four steps optimization approach.
to know a priori the parameter value that will lead to the global optimal solution, although
it is sometimes possible to derive upper and lower bounds for it. The common trade-off is
that global optimality may be guaranteed with a large value of this parameter, which often
results in prohibitive solution times.
Many different time representations have been introduced to solve scheduling problems
(for review see Floudas and Lin, 2004). Experience has shown that, depending on the
characteristics of the problem, some time representations are more suitable than others. In
this thesis, we focus on scheduling problems that rely on:
a) a set of possible operations, or actions, that can be performed once, several times, or
not at all;
b) scheduling decisions that involve both selecting, parametrizing and sequencing the
operations that should be executed;
c) scheduling constraints such as release dates, due dates, bounds on processing times,
non-overlapping constraints, sequence-dependent changeovers, cardinality constraints,
and precedence constraints;
d) additional side constraints that are used to model more complex features such as
limited inventory management or process constraints.
It should be noted that, for instance, the selection of operations may correspond to the
selection of equipment or discrete resources for tasks in a state-task-network (Kondili et al.,

2.2 Case Study
1993) or in a resource-task-network (Pantelides, 1994). In general, operations are defined

by fully disaggregating all possible discrete selections of actions in the scheduling system.
In contrast, parameterization of operations corresponds to continuous decisions such as
batch sizes, transfer volumes, or process operating conditions. The above assumptions do
not cover all kinds of scheduling problems but are an important part of the classification
presented by Méndez et al. (2006a). A unique aspect of this work is that the unifying
framework of the models presented in this chapter allows it to be applied to single-stage
and multi-stage batch scheduling problems as well as to crude-oil operations scheduling
(see chapters 3 and 4).
The main objective of this chapter is to develop a unified modeling approach for scheduling
problems in order to facilitate the evaluation of several time representations, both in terms of
computational time and solution quality. First, a simple scheduling problem is introduced
as an example. Next, we study four different types of time representations, which have
been used in the literature and clarify the relationships between them. Then, basic MILP
models for pure scheduling constraints are presented for each of these time representation.
Using concepts from graph theory (cliques and bicliques), we show how these models can be
generically strengthened based on the structure of the scheduling problem. Three solution
methods are then developed to solve these mathematical formulations. Finally, single-stage
and multi-stage batch scheduling problems are presented and solved using the different
approaches in order to show the effectiveness of the strengthened formulations, and to
provide elements of comparison between the different time representations.
2.2 Case Study
We introduce a small scheduling system that involves 6 different operations v1 , . . . , v6 and

3 unary resources r1 , r2 , r3 . A unary resource cannot be shared by two or more processing
operations at a given time. Table 2.1 displays resource requirement for each operation.
For example, operation v4 simultaneously requires resources r1 and r2 . In this case and

2.3 Time Representations
Table 2.1: Resource requirements for case study.

Operation v1 v2 v3 v4 v5 v6
Resources r1 r2 r3 r1 ∧ r2 r1 ∧ r3 r2 ∧ r3
in the examples studied in this thesis, unary resource requirements are handled as non-
overlapping constraints between operations. For instance, operations v1 and v4 cannot
overlap as they both use resource r1 . Also, operations v5 and v6 cannot overlap as they
both use resource r3 . Besides, as a given operation v can be executed several times, any
two separate executions of v may not overlap. Thus, any operation v cannot overlap with
itself. Different linear objectives can be considered: maximization of profit, minimization
of makespan, minimization of assignment costs, minimization of tardiness or earliness.
In order to extract useful information from the structure of the problem, we use a global
representation of all the non-overlapping constraints. The non-overlapping matrix, de-
noted by N O, is such that N Ovv0 = 1 if operation v and v 0 must not overlap, 0 oth-
erwise. The non-overlapping graph, denoted by GN O = (W, E), is an undirected graph
where the set of vertices W is the set of operations and the set of edges is defined by
E = {{v, v 0 } s.t. N Ovv0 = 1}. Therefore, the non-overlapping matrix is the adjacency ma-
trix of graph GN O . The concept of non-overlapping graph can be viewed as an extension of
the disjunctive graph (Balas, 1969; Adams et al., 1988), which is used to represent disjunc-
tive constraints between operations that have to be executed exactly once. In this thesis,
we consider operations that can be executed once, several times, or not at all. Figure 2.2
shows the non-overlapping matrix and graph for the case study. For clarity, edges that
connect a vertex to itself, called self-loops, are not represented.
In this section, we study four different time representations and show how they can be
defined using identical concepts. Each representation makes use of a totally ordered set
of priority-slots T = {1, . . . , n}, which are used to assign and order the executions of op-

v2 v4 
1 0 0 1 1 0

0 1 0 1 0 1
 
0 0 1 0 1 1
v6 v1 
1

 1 0 1 1 1
1 0 1 1 1 1
v3 v5 0 1 1 1 1 1
Figure 2.2: Non-overlapping matrix and graph for case study.
erations. The number of priority-slots, denoted by n, has to be postulated a priori. Any

operation may be executed several times by assigning it to multiple priority-slots. We de-
note by A = {(i, v), i ∈ T, v ∈ W } the set of all possible assignments and define a partial
scheduling order on A by:
∀(i1 , v1 ), (i2 , v2 ) ∈ A, (i1 , v1 ) ≺ (i2 , v2 ) ⇔ i1 < i2 ∧ N Ov1 v2 = 1
In each time representation, a solution is defined as a subset A0 of A and is represented

by a sequence of operations. Each selected assignment (i, v) correspond to an execution of
operation v with scheduling priority i during time interval [Siv , Eiv ]. In each time represen-
tation, the partial scheduling order ≺ implies precedence relations between elements of A0
as follows:
∀(i1 , v1 ), (i2 , v2 ) ∈ A0 , (i1 , v1 ) ≺ (i2 , v2 ) ⇒ Ei1 v1 ≤ Si2 v2
It is should be noted that it is not straightforward to select the number of priority-slots.

Indeed, postulating a large number priority-slots increases the chance of obtaining the global
optimal solution, but it also increases the size of the model and the CPU time. The four
time representations are listed below.
a. Multi Operation Sequencing (MOS)

b. Multi Operation Sequencing with Synchronized Start Times (MOS-SST)
c. Multi Operation Sequencing with Fixed Start Times (MOS-FST)
d. Single Operation Sequencing (SOS)

a. MOS (continuous-time) b. MOS-SST (continuous-time)

v1 1 4 v1 1 6
v2 2 v2 3
Operations
Operations
executions
executions
v3 3 v3 4
v4 3 v4 5
v5 2 v5 2
v6 1 v6 1
r1 r1
Resources
Resources
usage
usage
r2 r2
r3 r3
0 5 10 t =0 t t t t t =8
1 2 3 4 5 6
c. MOS-FST (discrete-time) d. SOS (continuous-time)
v 1 9 v 1 7
1 1
v 5 v 3
2 2
Operations
Operations
executions
executions
v 6 v 5
3 3
v 7 v 6
4 4
v 4 v 4
5 5
v 1 v 2
6 6
r r
Resources
Resources
1 1
usage
usage
r r
2 2
r r
3 3
t t t t t t t t t t 0 5 10
1 2 3 4 5 6 7 8 9 10
Figure 2.3: A unique schedule obtained through different time representations.
Figure 2.3 shows how the same schedule for the case study can be obtained within each
time representations. Each execution of an operation is represented by an horizontal bar
in the upper Gantt chart, while resource usage is represented by horizontal lines in the
lower Gantt chart. The priority-slots are represented by number labels on each operation
execution. In each case, the smallest possible number of priority-slots needed to obtained
the solution has been used. From this figure, it is clear that some time representations
require more priority-slots than others.
In the MOS representation, several operations can be assigned to each priority-slots as
long as they may overlap with each other. For instance, in Figure 2.3(a) operations v1 and v6

are allowed to overlap and are both assigned to the first priority-slot. However, operations
v1 and v5 cannot overlap and are consequently assigned to different priority-slots: slots 1
and 4 for operation v1 , slot 2 for operation v5 . If two non-overlapping operations v and v 0
are assigned to priority-slots i and j, respectively, such that i < j, then operation v 0 must
be executed after operation v (i.e. operation v must start after the end of operation v).
For instance, operation v1 assigned to priority-slot 4 is executed after operation v2 assigned
to priority-slot 2. The solution depicted in Figure 2.3(a) is represented by the sequence
of operations ({1, 6}, {2, 5}, {3, 4}, {1}). We denote MOS(n) a scheduling model using
the MOS time representation with n postulated priority-slots. This time representation
was introduced by Ierapetritou and Floudas (1998) as the event point formulation. Their
mathematical model, although significantly different than the model developed in this thesis,
was used to solve several STN problems. As mentioned by Maravelias and Grossmann
(2003), inventory tracking using event points is quite different than inventory tracking using
time points, which might lead to inconsistent enforcement of storage capacity constraints.
This issue was addressed by Janak et al. (2004) by adding additional storage tasks in the
STN problem, which can lead to a significant increase of model size.
The MOS-SST representation is based on the same features as the MOS representation.
Additionally, all operations assigned to the same priority-slot must have the same start
time. For instance, in Figure 2.3(b), operations v1 and v6 are both assigned to priority-slot
1, and therefore both start at the same time t = 0. Thus, each priority-slot i is associated
to variable time-point ti which is represented by a vertical dotted line in Figure 2.3(b). The
time interval between any two successive time-points is variable. The solution depicted in
Figure 2.3(a) is represented by the sequence of operations ({1, 6}, {2}, {5}, {3}, {4}, {1}).
We denote MOS-SST(n) a scheduling model using the MOS-SST time representation
with n postulated priority-slots. This type of representation has been used to solve a wide
variety of problems where time-points are used to track both the start and end events of
each operation (see Zhang and Sargent, 1996; Schilling and Pantelides, 1996; Maravelias
and Grossmann, 2003).

The MOS-FST representation is based on the same features as the MOS-SST repre-
sentation. Additionally, the time-point associated to each priority-slot is fixed a priori.
Thus, the interval between any two successive time-points is fixed. For instance, the so-
lution depicted in Figure 2.3(c) is obtained using time-points that are uniformly spaced
along the time horizon: t1 = 0, t2 = 1, . . . , t10 = 9. Therefore, operation v5 assigned to
priority-slot 4 starts at t = t4 = 3 while operation v4 assigned to priority-slot 7 starts at
t = t7 = 6. The solution depicted in Figure 2.3(c) is represented by the sequence of op-
erations ({1, 6}, ∅, ∅, {5}, {2}, {3}, {4}, ∅, {1}, ∅). We denote MOS-FST(n) a scheduling
model using the MOS-FST time representation with n postulated priority-slots. Discrete-
time formulation for process scheduling problems were initially developed to solve STN and
RTN models where processing times are assumed to be constant (see Kondili et al., 1993;
Pantelides, 1994).
In the SOS representation, at most one operation can be assigned to each priority-
slot. Similarly to the MOS model, if two non-overlapping operations v and v 0 are assigned
to priority-slots i and j (i < j), then v 0 must be executed after v. It should be noted
that this constraint does not apply to pairs of operations that are allowed to overlap. As
operations v2 and v5 are allowed to overlap, their relative position in time is not affected
by their respective scheduling priority. Therefore, the proposed solution is equivalent to
assigning operations v2 and v5 to priority-slots 4 and 3, respectively. The solution depicted
in Figure 2.3(d) is represented by the sequence of operations (1, 6, 2, 5, 3, 4, 1). We denote
SOS(n) a scheduling model using the SOS time representation with n postulated priority-
slots. This time representation was introduced by Mouret et al. (2009a) to solve the refinery
crude-oil operations scheduling problem.
Table 2.2 summarizes the correspondence between our nomenclature and equivalent nam-
ing conventions used in the literature. From these definitions, it can be inferred that for
a given number of priority-slots n the integer feasible space of MOS(n) is larger than
the integer feasible space of models MOS-SST(n), MOS-FST(n), and SOS(n). Indeed,
the latter models are derived from the MOS model by introducing additional constraints,

Table 2.2: Time representations nomenclature.

Nomenclature Equivalent names
MOS unit-specific time grid or multiple time grid
MOS-SST global events or common time grid or nonuniform discretization
MOS-FST fixed events or fixed time grid or uniform discretization
SOS first introduced in Mouret et al. (2009a)
which reduce the set of feasible solutions. Furthermore, the integer feasible space of MOS-
SST(n) is larger than the one of MOS-FST(n) and SOS(n). In particular, any solution
for the SOS model is a solution for the MOS-SST model. Indeed, at most one operation
can be assigned to each priority-slot so the synchronization of start times is automatically
satisfied.
These properties can also be interpreted by considering a scheduling solution z. We
denote z ∈ MOS(n) the membership of schedule z to the integer feasible space of model
MOS(n). In other words, z ∈ MOS(n) means that solution z satisfies all the constraints
of model MOS(n). We introduce the minimum number of priority-slots needed to ”find”
solution z using each time representation.
nMOS (z) = min{n|z ∈ MOS(n)}

n
nMOS-SST (z) = min{n|z ∈ MOS-SST(n)}

n
nMOS-FST (z) = min{n|z ∈ MOS-FST(n)}

n
nSOS (z) = min{n|z ∈ SOS(n)}

n
Then, the following inequalities hold:


 nMOS (z) ≤ nMOS-SST (z) ≤ nMOS-FST (z)
 n (z) ≤ n (z) ≤ n (z)
MOS MOS-SST SOS
The relation between MOS-SST and MOS-FST time representations was previously de-
rived by Maravelias and Grossmann (2006) in the context of state-task network formulations.
Remark. An important limitation of these time representations is that operations are
considered as a whole for sequencing purpose. More precisely, a scheduling priority is

2.4 Mathematical Models
assigned to operations and not to the start and end events of these operations, which may
be necessary to solve some scheduling problems. For instance, operations that require a
cumulative resource with capacity greater than 1 (e.g. manpower limited to 2 workers) are
allowed to overlap, but only a limited number of these operations may overlap at any given
point in time. The models developed in this thesis do not accommodate these features.
Another case is inventory tracking when simultaneous charging and discharging of tanks
is allowed. It is sufficient to enforce capacity limitations only at the start and end events
of each charging/discharging operations, but in order to do so, a precise sequence of such
events needs to be obtained by the model. Possible extensions of the model to handle these
specific features, such as presented in Janak et al. (2004), will not be discussed in this thesis.
In this section, we present mathematical models for each time representation. They all rely
on the same sets, parameters and variables. Objective functions are not presented here (e.g.
minimize makespan, minimize tardiness, or maximize profit) although they can significantly
impact the solution of the corresponding formulation.
2.4.1 Sets and Parameters
The following sets and parameters are used.
• T = {1, . . . , n} is a totally ordered set of priority-slots (indices i, j, i1 , i2 ).

• W is the set of all operations (indices v, v 0 , v1 , v2 ).
• H is the scheduling horizon.
• [Sv , Sv ] ⊂ [0, H] are bounds on the start time of any execution of operation v.
• [Dv , Dv ] ⊂ [0, H] are bounds on the duration of any execution of operation v.
• [Ev , Ev ] ⊂ [0, H] are bounds on the end time of any execution of operation v.
• [Nv , Nv ] are bounds on the total number of executions of operation v.
• [NW 0 , NW 0 ] are bounds on the total number of executions of all operations in W 0 ⊂ W .

• N Ov1 v2 is 1 if operations v1 and v2 must not overlap, 0 if they are allowed to overlap.
• T Rv1 v2 is a sequence-dependent transition time between non-overlapping operations
v1 and v2 .
• T RW 0 is a unique set transition time between any pair of non-overlapping operations
in W 0 ⊂ W .
• Pv1 v2 = 1 denotes a precedence constraint between operations v1 and v2 .
• PW1 W2 = 1 denotes a precedence constraint between sets of operations W1 and W2 .
Remark 1: It should be noted that for operations with fixed processing time, Dv =
Dv = Dv .
Remark 2: A set transition time T RW 0 is defined when ∀v1 , v2 ∈ W 0 , T Rv1 ,v2 = T RW 0 .
It can be used to represent unit changeover times.
Remark 3: A precedence constraint between operations v1 and v2 states that v1 must
be executed before v2 . We assume that precedence constraints are enforced on operations
which are executed exactly once (Nv1 = Nv2 = Nv1 = Nv2 = 1). Similarly, we consider
that a precedence constraint between sets of operations W1 and W2 states that exactly one
operation in each set must be executed (NW1 = NW2 = NW1 = NW2 = 1) and the operation
selected from W1 must be executed before the operation selected from W2 .
2.4.2 Variables
The variables used in all models are composed of binary assignment variables, and contin-
uous time variables.
• Assignment variables Ziv ∈ {0, 1} i ∈ T, v ∈ W

Ziv = 1 if operation v is assigned to priority-slot i, Ziv = 0 otherwise.
• Time variables Siv ≥ 0, Div ≥ 0, Eiv ≥ 0 i ∈ T, v ∈ W
Siv is the start time of operation v if it is assigned to priority-slot i, Siv = 0 otherwise.
Div is the duration of operation v if it is assigned to priority-slot i, Div = 0 otherwise.
Eiv is the end time of operation v if it is assigned to priority-slot i, Eiv = 0 otherwise.

2.4.3 MOS Model
Variable Bound Constraints Bounds on time variables can be expressed using the
following constraints.
Sv · Ziv ≤ Siv ≤ Sv · Ziv i ∈ T, v ∈ W (2.1a)
Dv · Ziv ≤ Div ≤ Dv · Ziv i ∈ T, v ∈ W (2.1b)
Ev · Ziv ≤ Eiv ≤ Ev · Ziv i ∈ T, v ∈ W (2.1c)
Time Constraint Time variables are linked through the following additional constraint.
Eiv = Siv + Div i ∈ T, v ∈ W (2.2)
Cardinality constraint The total number of execution of operations in a set W 0 ⊂ W

is restricted by the following constraint. A cardinality constraint on a single operation v
can be enforced by setting W 0 = {v}.
X
NW 0 ≤ Ziv ≤ NW 0 W0 ⊂ W (2.3)
i∈T
v∈W 0
Assignment Constraint Two non-overlapping operations v1 and v2 such that N Ov1 v2 =

1 cannot be assigned simultaneously to the same priority-slot.
Ziv1 + Ziv2 ≤ 1 i ∈ T, v1 , v2 ∈ W, N Ov1 v2 = 1 (2.4)
Non-overlapping Constraint A non-overlapping constraint between two operations

v1 , v2 ∈ W states that they must not be executed simultaneously. This property is en-
forced using the following big-M constraints where the big-M constant is defined as a valid
upper bound of the left hand side of the inequality.
Ei1 v1 ≤ Si2 v2 + H · (1 − Zi2 v2 ) i1 , i2 ∈ T, i1 < i2 , v1 , v2 ∈ W, N Ov1 v2 = 1 (2.5a)
Ei1 v2 ≤ Si2 v1 + H · (1 − Zi2 v1 ) i1 , i2 ∈ T, i1 < i2 , v1 , v2 ∈ W, N Ov1 v2 = 1 (2.5b)

Due to the assignment constraint (2.4) and variable bound constraints (2.1a) and (2.1c),
equations (2.5a) and (2.5b) can be combined in order to form the following tighter surrogate
constraint (see appendix A).
Ei1 v1 + Ei1 v2
i1 , i2 ∈ T, i1 < i2 , v1 , v2 ∈ W, N Ov1 v2 = 1 (2.6)
≤ Si2 v1 + Si2 v2 + H · (1 − Zi2 v1 − Zi2 v2 )
Non-overlapping Constraint with Transition Times Sequence-dependent transition

times T Rv1 v2 between operations v1 ∈ W and v2 ∈ W can be enforced as follows.
Ei1 v1 + T Rv1 v2 · Zi1 v1

i1 , i2 ∈ T, i1 < i2 , v1 , v2 ∈ W, N Ov1 v2 = 1 (2.7a)
≤ Si2 v2 + (H + T Rv1 v2 ) · (1 − Zi2 v2 )
Ei1 v2 + T Rv2 v1 · Zi1 v2

i1 , i2 ∈ T, i1 < i2 , v1 , v2 ∈ W, N Ov1 v2 = 1 (2.7b)
≤ Si2 v1 + (H + T Rv2 v1 ) · (1 − Zi2 v1 )
If the transition time is not sequence-dependent (i.e. T Rv1 v2 = T Rv2 v1 ) then constraints
(2.7) can be combined in order to form the following tighter surrogate constraint.
Ei1 v1 + Ei1 v2 + T Rv1 v2 · (Zi1 v1 + Zi1 v2 ) i1 , i2 ∈ T, i1 < i2 , v1 , v2 ∈ W,

(2.8)
≤ Si2 v1 + Si2 v2 + (H + T Rv1 v2 ) · (1 − Zi2 v1 − Zi2 v2 ) N Ov1 v2 = 1, T Rv1 v2 = T Rv2 v1
Precedence Constraint Consider two operations v1 , v2 ∈ W such that Pv1 v2 = 1, then

the precedence constraint can be expressed as follows. Note that in each summation, exactly
one term can be non-zero due to the assumed cardinality constraint.
X X
Eiv1 ≤ Siv2 v1 , v2 ∈ W, Pv1 v2 = 1 (2.9)
i∈T i∈T
Consider two sets of operations W1 , W2 ⊂ W . If PW1 W2 = 1, the previous constraint can be

extended as follows. Note that in each summation, exactly one term can be non-zero due
to the corresponding cardinality constraint (2.3).
X X X X
Eiv1 ≤ Siv2 W1 , W2 ⊂ W, PW1 W2 = 1 (2.10)
i∈T v1 ∈W1 i∈T v2 ∈W2

2.4.4 MOS-SST Model
In the MOS-SST model, the start times of all operations assigned to the same priority-slot
i have to be synchronized. Therefore, we introduce positive synchronization time-points
variables ti (i ∈ T ). Variables ti correspond to the start time of all operations assigned to
priority-slot i.
ti ∈ [0, H] i∈T (2.11)
Time-point Sequence Constraint The following constraint ensures that the variables
ti are ordered in time.
ti−1 ≤ ti i ∈ T \ {1} (2.12)
Synchronization Constraints If operation v is assigned to priority-slot i, it must start

at time ti . Therefore, the following synchronization constraints are used. Note that con-
straint (2.1a) already ensures that Siv = 0 if Ziv = 0.
Siv ≤ ti i ∈ T, v ∈ W (2.13a)
Siv ≥ ti − H · (1 − Ziv ) i ∈ T, v ∈ W (2.13b)
2.4.5 MOS-FST Model
In the MOS-FST model, the time-points variables ti of the MOS-SST model are used and
are fixed a priori. They can therefore be considered as parameters. In this thesis, these
time-points are always selected using a uniform time discretization.
i−1
ti = ·H i∈T (2.14)
n
Synchronization Constraint If operation v is assigned to priority-slot i, it must start

at time ti . Therefore, the following synchronization constraints are used.
Siv = ti · Ziv i ∈ T, v ∈ W (2.15)

2.5 Strengthened Reformulations
2.4.6 SOS Model
All constraints from the MOS model are still valid for the SOS model. However, a new
assignment constraint is defined.
Assignment Constraint At most one operation has to be assigned to each priority-slot.

X
Ziv ≤ 1 i∈T (2.16)
v∈W
The models presented in section 2.4 may not be effectively solved by MILP solvers. Indeed,
special attention needs to be paid to their LP relaxation. As MILP solvers usually perform
better when the model has a tight LP relaxation, we introduce strengthened formulations
based on the non-overlapping structure of the problem.
2.5.1 Non-overlapping Graph Properties
We recall the definition of a clique and a biclique in the context of the non-overlapping
graph GN O . A clique of GN O is a subset of the set of operations W 0 ⊂ W such that
any two operations in W 0 must not overlap. A maximal clique is a clique that is not a
subset of any other clique. An isolated clique is a clique that has no edges connecting it
to its complement in the graph. For the case study, the non-overlapping graph displayed
in Figure 2.2 contains 9 non-maximal cliques of two vertices, one for each non self-loop
edge (e.g. {v1 , v4 }). It contains 4 maximal cliques of three vertices {v1 , v4 , v5 }, {v2 , v4 , v6 },
{v3 , v5 , v6 }, {v4 , v5 , v6 }. It contains no clique of larger size and no isolated clique.
We define a biclique of GN O as a pair of sets of operations (W1 ; W2 ) ∈ W 2 such that for
any pair of operations (v1 ; v2 ) ∈ W1 × W2 , {v1 , v2 } is an edge of GN O . Sets W1 and W2
are not necessarily disjoint. A maximal biclique of GN O is a biclique that is not contained
in any other biclique of GN O . For the case study, the non-overlapping graph displayed
in Figure 2.2 contains 9 non-maximal bicliques of two vertices, one for each non self-loop

v1 v4
v6 v5
Figure 2.4: Biclique ({v1 , v6 }; {v4 , v5 }).
edge (e.g. (W1 = {v1 };W2 = {v4 })). It contains 4 maximal bicliques of three vertices that
can be derived from its maximal cliques ({v1 , v4 , v5 };{v1 , v4 , v5 }), ({v2 , v4 , v6 };{v2 , v4 , v6 }),
({v3 , v5 , v6 };{v3 , v5 , v6 }), and ({v4 , v5 , v6 };{v4 , v5 , v6 }). It also contains 3 maximal bicliques
composed of four vertices ({v1 , v6 };{v4 , v5 }), ({v2 , v5 };{v4 , v6 }), ({v3 , v4 };{v5 , v6 }) and 3
maximal bicliques composed of five vertices ({v4 };{v1 , v2 , v4 , v5 , v6 }), ({v5 };{v1 , v3 , v4 , v5 , v6 }),
({v6 };{v2 , v3 , v4 , v5 , v6 }). Figure 2.4 depicts the subgraph of GN O corresponding to biclique
({v1 , v6 };{v4 , v5 }).
Remark 4: Isolated cliques are always maximal as they cannot be extended to larger
cliques.
Remark 5: Using maximal cliques instead of non-maximal cliques generally leads to a
tighter LP relaxation. Constraints based on cliques will therefore be applied to maximal
cliques only. Appendix A shows how applying strengthened constraints to maximal cliques
only generates the tightest and most compact model. The same remark applies to bicliques.
Remark 6: The number of maximal cliques in an undirected graph might be exponential
in the size of the graph. Nevertheless, there are exponential-time algorithms to enumerate
all maximal cliques of a graph and we assume the non-overlapping graph is small enough so
that this task can be performed in reasonable time. The same remark applies to bicliques.
2.5.2 MOS Model
Aggregated Assignment Constraint Let W 0 ⊂ W be a clique of the non-overlapping

graph GN O . Then, at most one operation from W 0 can be assigned to priority-slot i.
Therefore, the following constraint, which is at least as strong as constraint (2.4) as shown

in appendix A, is valid.
X
Ziv ≤ 1 i ∈ T, W 0 ∈ clique(GN O ) (2.17)
v∈W 0
Aggregated Non-overlapping Constraint Given a clique W 0 ⊂ W of the

non-overlapping graph GN O , the following aggregated non-overlapping constraint, which
is at least as strong as constraint (2.6) as shown in appendix A, is valid. Note that only one
term in each summation can be non-zero due to the aggregated assignment constraint (2.17).
X X X
Ei1 v ≤ Si2 v + H · (1 − Zi2 v ) i1 , i2 ∈ T, i1 < i2 , W 0 ∈ clique(GN O ) (2.18)
v∈W 0 v∈W 0 v∈W 0
Aggregated Non-overlapping Constraint with Intermediate Operations Given

a clique W 0 ⊂ W of the non-overlapping graph GN O , and two priority-slots i1 , i2 ∈ T such
that i1 < i2 , constraint (2.18) can be further strengthened (see appendix A) by including
the duration of operations assigned to intermediate priority-slots i (i1 < i < i2 ).
X X X
Ei1 v + Div
v∈W 0 i∈T v∈W 0
i1 <i<i2 i1 , i2 ∈ T, i1 < i2 , W 0 ∈ clique(GN O ) (2.19)
X X
≤ Si2 v + H · (1 − Z i2 v )
v∈W 0 v∈W 0
Aggregated Non-overlapping Constraint with Clique Transition Times Given a

clique W 0 ⊂ W of the non-overlapping graph GN O , the clique transition time T RW 0 is the
minimum time delay between any two executions of operations in W 0 . If W 0 represents a
set of operations executed in a unit, T RW 0 corresponds to a unit transition time. Clique
transition times can be enforced as follows. Note that the value of the big-M constant is

increased to a new valid upper bound for the left hand side of the inequality: H + T RW 0 .
X
(Ei1 v + T RW 0 · Zi1 v )
v∈W 0
X X
+ (Div + T RW 0 · Ziv )
i1 , i2 ∈ T, i1 < i2 , W 0 ∈ clique(GN O ) (2.20)
i∈T v∈W 0
i1 <i<i2
X X
≤ Si2 v + (H + T RW 0 ) · (1 − Z i2 v )
v∈W 0 v∈W 0
2.5.3 MOS-SST Model
Aggregated Synchronization Constraints Given a clique W 0 ⊂ W of the

non-overlapping graph GN O , the following aggregated synchronization constraints, which
are tighter than constraints (2.13) as shown in appendix A, are valid. Note that in each
summation only one term can be non-zero.
X
Siv ≤ ti i ∈ T, W 0 ∈ clique(GN O ) (2.21a)
v∈W 0
X X
Siv ≥ ti − H · (1 − Ziv ) i ∈ T, W 0 ∈ clique(GN O ) (2.21b)
v∈W 0 v∈W 0
Tightening Precedence Constraint Consider a precedence constraint between sets of

operations W1 and W2 (PW1 W2 = 1). Assume that an operation from W1 is assigned to
priority-slot i1 with associated time-point t1 and that an operation from W2 is assigned
to priority-slot i2 with associated time-point t2 . Then, due to the precedence constraint,
we must have t1 < t2 . Therefore, a necessary condition for the precedence constraint to
hold true is i1 < i2 . The following constraint ensures that this condition is satisfied, and
therefore complements constraint (2.10).
X X X X
Zjv ≥ Zjv i ∈ T, W1 , W2 ⊂ W, PW1 W2 = 1 (2.22)
j∈T v∈W1 j∈T v∈W2
j<i j≤i

l m
Dv
i− H/n i
H/n
Dv
Figure 2.5: Assignment constraint using consecutive time-points.
2.5.4 MOS-FST Model
The use of graph theory to strengthen discrete-time scheduling formulations has been stud-
ied by Waterer et al. (2002) who derived facet defining inequalities based on cliques and 5-
holes in a finite graph. This graph can be seen as an extension of the non-overlapping graph
where each node corresponds to an operation and a time interval (with integer bounds).
In this work, we use a basic non-overlapping graph (with no time intervals) which only
generates clique-based constraints because it is not possible to enumerate time intervals
in continuous-time representations. Therefore, the exact same graph is used for all time
representations.
Assignment Constraint using Time-points Consider an operation v ∈ W . If the time

interval between two time-points is smaller than the duration of v, then v cannot start at
both time-points (v cannot be assigned to both priority-slots). As illustrated in Figure 2.5,
l m
Dv
v cannot be assigned to priority-slots i − H/n + 1 and i simultaneously, as well as any
l m
Dv
other priority-slot in between. However, it can be assigned to priority-slots i − H/n and i
simultaneously. Note that the lower bound on the duration of v is used to account for the
case of variable processing times. Therefore, the following constraint, which is a stronger
extension of constraint (2.4), is valid.
Dv
X
Ziv + Zjv ≤ 1 i ∈ T, v ∈ W, δv = (2.23)
H/n
j∈T
i−δv <j<i

Aggregated Assignment Constraint The previous constraint (2.23) can be extended

to cliques of GN O . For fixed processing times, this constraint is equivalent to equation (9)
from Shah et al. (1993).
 
 
X   X 
 Dv

0
Ziv + Zjv ≤1 i ∈ T, W ∈ clique(W ), δv = (2.24)
H/n
v∈W 0 
 
 j∈T 

i−δv <j<i
Aggregated Assignment Constraint with Transitions Times The previous con-

straint (2.24) can also be extended to the case of clique transition times.
 

 

Dv + T RW 0
X  X 
0
Ziv + Zjv ≤ 1 i ∈ T, W ∈ clique(W ), δW 0 (v) =
H/n
v∈W 0 
 
 j∈T 

i−δW 0 (v)<j<i
(2.25)
2.5.5 SOS Model
Aggregated Non-overlapping Constraint Additional aggregated non-overlapping con-

straints can be generated based on bicliques of the non-overlapping graph for the SOS
model. Given a biclique (W1 ; W2 ) of the non-overlapping graph GN O , the following aggre-
gated non-overlapping constraints, which are at least as strong as constraints (2.5) as shown
in appendix A, are valid. Note that in each summation only one term can be non-zero due
to the SOS specific assignment constraint (2.16). Also, note that constraints (2.18) and
(2.26) are not redundant as they are applied to different sets of operations.
X X X i1 , i2 ∈ T, i1 < i2 ,
Ei1 v ≤ Si2 v + H · (1 − Z i2 v ) (2.26a)
v∈W1 v∈W2 v∈W2 (W1 ; W2 ) ∈ biclique(GN O )
X X X i1 , i2 ∈ T, i1 < i2 ,
Ei1 v ≤ Si2 v + H · (1 − Z i2 v ) (2.26b)

2.6 Solution Methods
Determining the minimum number of priority-slots needed to find the optimal schedule is
non-trivial. A commonly used algorithm is to solve several scheduling models, each time
increasing the number of priority-slots. We present two different approaches based on this
idea followed by the more classic direct approach.
2.6.1 Additive Approach
The MOS, MOS-SST and SOS time representations are such that any solution found with
n priority-slots can be found with n + 1 priority-slots: z ∈ MOS(n) ⇒ z ∈ MOS(n + 1).
Therefore, the corresponding models can be solved by successively increasing by 1 the
number of priority-slot. The additive approach is described by Algorithm 1. The parameter
n0 is the initial number of priority-slots. To improve efficiency of the branch & bound
algorithm at each iteration, the cutoff parameter is set using the objective value of the best
incumbent found so far. During some iterations, the MILP model may not have any solution
strictly better than the best incumbent. In such cases, it will not return any solution even
though the model may be feasible.
Three stopping criteria are considered. The first one is (∆ ≤ ε) where ε is an absolute
tolerance on the variation of objective value. In general, this criterion does not guarantee
global optimality of the solution even when setting ε to 0. However, it leads to a small
number of iterations. Also, in most of our experiments, this stopping criterion returned
the global optimal solution, which was only proved optimal by using the following stopping
criterion.
The second stopping criterion is (n > n) where n is an upper limit on the number of
priority-slots. In some cases, it is possible to determine an upper limit on the number of
priority-slots needed to find the optimal solution of the scheduling problem. In such cases,
this stopping criterion guarantees global optimality of the solution.
The third stopping criterion is a time limit on the total computational time, which of

Algorithm 1: Additive approach.

begin
z ∗ ←− ∅ ;
cutoff ←− −∞ ;
n ←− n0 ;
repeat
z ←− Maximize(MOS(n), cutoff) ;
∆ ←− z.objval() − z ∗ .objval() ;
if ∆ > 0 then
z ∗ ←− z ;
cutoff ←− z ∗ .objval() ;
n ←− n + 1 ;
until stopping condition ;
return z ∗ ;
end
course does not guarantee global optimality.

It is important to note that at each iteration, the integer feasible space of MOS(n)
includes scheduling solutions explored during previous iterations. In order to avoid re-
dundant search, we introduce constraint (2.27) that rejects any solution that do not make
use of all priority-slots. By adding this constraint to the MOS(n) model, the property
z ∈ MOS(n) ⇒ z ∈ MOS(n + 1) is no longer valid. The same remark holds true for the
MOS-SST(n) and SOS(n) models.
X
Ziv ≥ 1 i∈T (2.27)
v∈W
2.6.2 Multiplicative Approach
The previous approach could be used to solve a scheduling problem using the MOS-FST
time representation. However, a major flaw is that it is not guaranteed that a solution
found with n priority-slots can be found with n + 1 priority-slots. This can be overcome
by multiplying the number of priority-slots be a factor of 2 instead of using an increment.
Indeed, any solution found with n priority-slots can be found with 2n priority-slots: z ∈
MOS-FST(n) ⇒ z ∈ MOS-FST(2n). The multiplicative approach is therefore very

Algorithm 2: Multiplicative approach.

begin
z ∗ ←− ∅ ;
cutoff ←− −∞ ;
n ←− n0 ;
repeat
z ←− Maximize(MOS(n), cutoff) ;
∆ ←− z.objval() − z ∗ .objval() ;
if ∆ > 0 then
z ∗ ←− z ;
cutoff ←− z ∗ .objval() ;
n ←− 2 · n ;
until stopping condition ;
return z ∗ ;
end
similar to the additive approach and is described by Algorithm 2. The stopping criteria
introduced for the additive approach can still be used for the multiplicative approach.
2.6.3 Direct Approach
Although the additive approach can be used to solve SOS models, it might not be very
efficient. Indeed, each time the number of priority-slot is increased by 1, the solver can
only schedule one additional operation. In MOS or MOS-SST models, several additional
operations can be scheduled at each iteration, which leaves much more flexibility to better
improve the objective value. Instead, a direct approach can be used. It consists of choosing
a fixed value for n and solving the MILP model once. In some cases, the total number of
executions of operations is fixed and known in advanced, so it can be used as the number
of priority-slots, guaranteeing global optimality of the solution obtained. It other cases,
a detailed analysis of the scheduling structure of the problem needs to be performed to
effectively define a value for n. Global optimality may not be guaranteed if n is too small,
or if additional constraints are used to improve the search.

2.7 Single-Stage Batch Scheduling Problem
In this section, we apply the various time representations to model single-stage batch
scheduling problems. We study several instances introduced in Pinto and Grossmann (1995).
The largest instance has 29 orders to be processed before given due dates from time 0 to
time 30. Four units are available to process each single-stage order. Each unit can process
only one order at a time and a minimum unit-specific set-up time is required between any
two orders. Each order can only be processed on a subset of all units with order-and-unit-
specific processing times. Table 2.3 displays all required data. For each order, units that
do not have a corresponding processing time cannot be selected for this order. The ob-
jective is to minimize total earliness which corresponds to maximizing the end times of all
orders. Five instances have been studied with 8, 12, 18, 25, and 29 orders, which we denote
SSBSP8, . . . , SSBSP29. We introduce the following specific sets.
• O ⊂ {o1 , . . . , o29 } is the set of orders (O = {o1 , . . . , o8 } for SSBSP8).

• U = {u1 , . . . , u4 } is the set of units.
• Uo is the set of units on which order o ∈ O can be processed. For instance, Uo1 =
{u1 , u4 }.
• Ou is the set of orders that can be processed on unit u ∈ U . For instance, Ou3 =
{o4 , o5 , o7 , o8 }.
The set of operations W can then be defined as follows. Exactly one operation is defined
for each order and each unit on which the order can be processed. This reduces the number
of indices from 2 to 1, although the combinatorial size of the problem remains identical.
W = {(o, u) ∈ O × U, u ∈ Uo } = {o1 u1 , o1 u4 , o2 u1 , o2 u4 , o3 u1 , o3 u4 , o4 u3 , o4 u4 , . . .}
Figure 2.6 depicts the non-overlapping graph of SSBSP8. It contains 3 isolated cliques each
corresponding to one unit (unit u2 cannot be used in this instance as it is excluded for
all orders, as seen in Table 2.3). The unit-based structure of the non-overlapping isolated
cliques leads to classical strengthened assignment and scheduling constraints that have

Table 2.3: Data for single-stage batch scheduling problems.

Unit processing time (hr)
Order Due date (hr)
1 2 3 4
1 15 1.538 1.194
2 30 1.500 0.789
3 22 1.607 0.818
4 25 1.564 2.143
5 20 0.736 1.017
6 30 5.263 3.200
7 21 4.865 3.025 3.214
8 26 1.500 1.440
9 30 1.869 2.459
10 29 1.282
11 30 3.750 3.000
12 21 6.796 7.000 5.600
13 30 11.250 6.716
14 25 2.632 1.527
15 24 5.000 2.985
16 30 1.250 0.783
17 30 4.474 3.036
18 30 1.429
19 13 3.130 2.687
20 19 2.424 1.074 1.600
21 30 7.317 3.614
22 20 0.864
23 12 3.624
24 30 2.667 4.000
25 17 5.952 3.448 4.902
26 20 3.824 1.757
27 11 6.410 3.937
28 30 5.500 3.235
29 25 4.286
Set-up times 0.180 0.175 0.000 0.237
already been presented in the literature. Therefore, the mathematical models for batch
scheduling problems do not differ very much from previous works. We will denote Wu =
{v = (o, u), o ∈ Ou }, where u ∈ U , the set of operations executed on unit u ∈ U , Wu is
an isolated clique of GN O . Given an order o ∈ O, exactly one execution of this order must
be performed. Therefore, the set Wo = {v = (o, u), u ∈ Uo } has an associated cardinality

o7 u4
o7 u1 o8 u4 o6 u4
o6 u1 o4 u3 o8 u3
o1 u1 o1 u4 o5 u4
o3 u1 o5 u3 o7 u3
o2 u1 o2 u4 o4 u4
o3 u4
Figure 2.6: Non-overlapping graph with isolated cliques for SSBSP8.
constraint of 1, that is NWo = NWo = 1.
2.7.1 MOS Model
The MOS model for the single-stage batch scheduling problem is derived from constraints
(2.1)-(2.3), (2.17), (2.20), and (2.27).

XX
max Eiv
i∈T v∈W
s.t. Div = Dv · Ziv i ∈ T, v ∈ W
Eiv ≤ Ev · Ziv i ∈ T, v ∈ W
Eiv = Siv + Div i ∈ T, v ∈ W
X X
Ziv = 1 o∈O
i∈T v∈Wo
X
Ziv ≤ 1 i ∈ T, u ∈ U
v∈Wu
X
(Ei1 v + T RWu · Zi1 v )
v∈Wu
X X
+ (Div + T RWu · Ziv )
i∈T v∈Wu
i1 , i2 ∈ T, i1 < i2 , u ∈ U
i1 <i<i2
X X
≤ Si2 v + (H + T RWu ) · (1 − Z i2 v )
v∈Wu v∈Wu
X
Ziv ≥ 1 i∈T
v∈W
Siv , Div , Eiv ≥ 0 i ∈ T, v ∈ W
Ziv ∈ {0, 1} i ∈ T, v ∈ W
This model can be strengthened in different ways. For each unit, it is possible to derive
the minimum and maximum number of orders that will be processed on it. The maximum
number of orders that can be processed on unit u ∈ U is defined as the number of operations
in the set.

NWu = Wu u∈U
An improved upper cardinality value for Wu can be obtained by considering the processing
times of operations that might be processed on unit u. Indeed, due to a limited scheduling
horizon, it might not be possible to execute all operations in Wu . Therefore, we use the
following improved definition that solves a one-dimensional knapsack problem where all
operations have a value of 1. A linear algorithm to solve this problem consists of ordering
the operations in Wu with respect to their duration Dv and, starting from the operation

Table 2.4: Unit cardinality bounds depending on parameter n for SSBSP29.

n u1 u2 u3 u4
8 [8,8] [5,5] [8,8] [8,8]
9 [6,9] [2,5] [6,9] [6,9]
10 [4,9] [2,5] [5,10] [5,10]
11 [2,9] [2,5] [4,11] [4,11]
12 [1,9] [2,5] [3,11] [4,12]
13 [0,9] [2,5] [2,11] [4,13]
14 [0,9] [2,5] [2,11] [4,14]
with the lowest processing time, setting γv to 1 until the knapsack limit H + T RWu is
reached. For the remaining operations, γv is set to 0.
( )
X X
NWu = max γv γv ∈ {0, 1}, (Dv + T RWu ) · γv ≤ H + T RWu u∈U
v∈Wu v∈Wu
We denote Wu1 = {v = (o, u) ∈ W, Uo = {u}} the set of operations that can only be
processed on unit u ∈ U . The minimum number of orders that will be processed on unit
u ∈ U can be defined as follows.
NWu = Wu1

u∈U
However, as the number of orders executed on units different than u ∈ U is limited, NWu
can be increased as follows.
 

 

 X 
1
NWu = max Wu , O −
min{n, NWu0 } u∈U

u0 ∈U
 

 0

u 6=u
The values obtained for NWu and NWu in problem SSBSP29 are displayed in Table 2.4 as
an example. Using NWu and NWu , the following cardinality constraint can be derived.
X X
NWu ≤ Ziv ≤ NWu u∈U (2.28)
i∈T v∈Wu
Given a unit u ∈ U , Wu is an isolated clique of GN O , so the optimal sequence of operations

from Wu is not affected by operations from W \ Wu . It is therefore possible to reduce

the set of possible sequences of operations from Wu by assigning these operations to the
first priority-slots only. In other words, an operation Wu can be assigned to priority-slot i
only if an operation from Wu is assigned to priority-slot i − 1. This leads to the following
symmetry-breaking constraint.
X X
Ziv ≤ Z(i−1)v i ∈ T, i 6= 1, u ∈ U (2.29)
v∈Wu v∈Wu
This idea can be further applied to other constraints in the model. A maximum cardinality
constraint on the set of operations Wu can be enforced by setting to 0 assignment variables
corresponding to the last priority-slots for this set.
X
Ziv = 0 i ∈ T, i > NWu , u ∈ U (2.30)
v∈Wu
Also, a minimum cardinality constraint on Wu can be enforced by assigning exactly one

operation to the first priority-slots.
X
Ziv = 1 i ∈ T, i ≤ NWu , u ∈ U (2.31)
v∈Wu
Besides, non-overlapping constraint (2.20) can be tightened for the first priority-slots as
follows.
X
v∈Wu
X X
i1 , i2 ∈ T, i1 < i2 ≤ NWu , u ∈ U (2.32)
i∈T v∈Wu
i1 <i<i2
X
≤ S i2 v
v∈Wu
The single-stage batch scheduling problems are solved with the MOS model plus con-
straints (2.28)-(2.32). The additive approach is used with the second stopping criterion
(n > maxu Nu ) which guarantees global optimality of the solution. Indeed, for each unit
u ∈ U , no more than Nu priority-slots are needed to sequence operations on it. The initial

number of priority-slots is set to n0 = O {u|Wu 6= ∅} which is the minimum number

of orders at least one unit has to process. Computational results are given in Table 2.5.
Experiments were run on an Intel Xeon 1.86GHz processor using GAMS/CPLEX 11. For
each iteration, the LP relaxation, MILP solution, number of nodes, CPU time and cumu-
lative CPU time are displayed. The term “no solution” for MILP solution is used when
CPLEX did not find any solution with higher objective value than the cutoff value which
does not mean that no solution exists. For each problem, the global optimal solution is
underlined. The CPU time elapsed until the first stopping criterion (∆ ≤ 0) is satisfied is
also underlined. The results show that the first stopping criterion leads to significant reduc-
tion of CPU times compared to the second. Also, it is interesting to note the problem with
29 orders, although larger, is solved faster than the 25 orders problem. As the scheduling
horizon is fixed to 30 hours, the density of operations is higher with 29 orders. Therefore,
this problem is more constrained and has potentially fewer solutions. The model size is
increased but the branch & bound tree is smaller, thus leading to smaller CPU times. This
is verified by the fact that fewer nodes are explored.
Additionally, we solved problem SSBSP18 using the full MOS models with or without
the minimum priority-slot usage constraint (2.27). The results obtained with the additive
approach are displayed in Figure 2.7. It shows that constraint (2.27) greatly reduces the
search space at each iteration. In particular, using this constraint, the last iterations are
solved in few seconds, whereas CPU times are higher than 100 seconds if it is not used.
We also solved the SSBSP18 instance with exactly 5 priority-slots using the full MOS
model without and with symmetry-breaking constraints, which corresponds to the first iter-
ation of the additive algorithm for this problem. Without symmetry-breaking constraints,
the problem was solved in 115 seconds (124788 nodes) as opposed to less than 2 seconds
(513 nodes) if symmetry-breaking constraints (2.29)-(2.32) are used. This shows why us-
ing symmetry-breaking concepts is crucial to solve scheduling problems as they tend to be
highly degenerate (Kallrath, 2002).
We performed another experiment which consisted of solving the full MOS model by using
the original assignment constraint (2.4) instead of the strengthened assignment constraint

Table 2.5: MOS computational results for single-stage batch scheduling problems.
Pb n LP MILP Nb of nodes CPU Cumulative CPU
3 189.000 189.000 3 0.81s 0.81s
4 189.000 no solution 0 0.77s 1.58
5 189.000 no solution 0 0.80s 2.38s
SSBSP8
6 189.000 no solution 0 0.76s 3.14s
7 189.000 no solution 0 0.86s 4.00s
8 188.035 no solution 0 0.90s 4.90s
3 298.517 296.543 0 0.83s 0.83s
4 299.000 297.974 604 1.67s 2.50s
5 299.000 no solution 207 1.38s 3.88s
6 299.000 no solution 198 1.71s 5.59s
SSBSP12 7 299.000 no solution 0 1.17s 6.76s
8 299.000 no solution 0 1.02s 7.78s
9 no solution 0.95s 8.73s
5 461.563 450.760 513 1.83s 1.83s
6 463.993 450.982 1952 5.71s 7.54s
7 467.196 451.504 2924 8.37s 15.91s
8 467.966 no solution 4600 16.84s 32.75s
9 467.966 no solution 4600 18.47s 51.22
SSBSP18
10 459.877 no solution 0 1.35s 52.57
7 593.615 575.929 3444 17.24s 17.24s
8 601.884 579.089 18261 69.90s 87.14s
9 607.713 579.570 55046 189.66s 276.80s
10 609.000 no solution 77800 408.97s 685.77s
SSBSP25
11 609.000 no solution 22400 113.60s 799.37s
12 597.717 no solution 0 3.61s 802.98s
13 583.289 no solution 0 2.31s 805.29s
8 649.531 628.413 1583 12.97s 12.97s
9 656.403 634.964 10870 56.16s 69.13s
10 662.720 635.104 42249 206.66s 275.79s
SSBSP29 11 666.512 no solution 22400 110.83s 386.62s
12 666.552 no solution 3700 32.72s 419.34s
13 668.164 no solution 1800 28.33s 447.67s
14 653.051 no solution 0 4.72s 452.39s

10000
MOS w/ minimum priority-slot usage
MOS w/o minimum priority-slot usage
1000
CPU time (s)

100
10
1
6 8 10 12 14
Number of priority-slots
Figure 2.7: Effect of the minimum priority-slot usage constraint.
(2.17). Instance SSBSP18 was solved in 3.43 seconds (587 nodes) instead of 1.83 seconds
(513 nodes) for the strengthened constraint. Thus, this constraint slightly helps to improve
the solution time although CPLEX is able to generate cuts to tighten the LP relaxation
accordingly. However, we also solved the full MOS model replacing the strengthened non-
overlapping constraints (2.20) and (2.32) by the following non-overlapping constraint with
unit-dependent transition time, which is based on constraint (2.8).
Ei1 v1 + Ei1 v2 + T RWu · (Zi1 v1 + Zi1 v2 )

i1 , i2 ∈ T, i1 < i2 , u ∈ U, v1 , v2 ∈ Wu
≤ Si2 v1 + Si2 v2 + (H + T RWu ) · (1 − Zi2 v1 − Zi2 v2 )
(2.33)
The model was solved in 3,288 seconds and 1,078,034 nodes (vs 1.83 seconds and 513 nodes)
which proves that CPLEX was not successful at improving the LP relaxation of the model
for this mixed-integer constraint. In general, MILP solvers are very efficient at solving IPs
as they are able to exploit the structure of the model in order generate very tight cuts (such
as clique constraints, Nemhauser and Wolsey, 1999) that can lead to large improvements of
the branch & bound search. However, it is much harder to generate tightening cuts from
constraint (2.33) using the clique structure. Therefore, using strengthened formulations for
mixed-integer constraints is very important as it greatly improves the performance of MILP
solvers.

2.7.2 MOS-SST Model
The MOS-SST model for the single-stage batch scheduling problem is derived from con-
straints (2.1)-(2.3), (2.12), (2.17), (2.20), (2.21), and (2.27).
XX
max Eiv
i∈T v∈W
X X
Ziv = 1 o∈O
i∈T v∈Wo
X X
Nu ≤ Ziv ≤ Nu u∈U
i∈T v∈Wu
X
Ziv ≤ 1 i ∈ T, u ∈ U
v∈Wu
X
v∈Wu
X X
i∈T v∈Wu
i1 , i2 ∈ T, i1 < i2 , u ∈ U
i1 <i<i2
X X
≤ Si2 v + (H + T RWu ) · (1 − Z i2 v )
v∈Wu v∈Wu
ti−1 ≤ ti i ∈ T, i 6= 1
X
Siv ≤ ti i ∈ T, u ∈ U
v∈Wu
X X
Siv ≥ ti − H · (1 − Ziv ) i ∈ T, u ∈ U
v∈Wu v∈W 0
X
Ziv ≥ 1 i∈T
v∈W
Ziv ∈ {0, 1} i ∈ T, v ∈ W
ti ∈ [0, H] i∈T
Due to the use of time-points, the symmetry-breaking constraints developed in section 2.7.1
cannot be applied to the MOS-SST model. Similarly to the MOS model, the additive ap-
proach is used and the initial number of priority-slots is set to

Table 2.6: MOS-SST computational results for single-stage batch scheduling problems.
3 189.000 177.001 1,220 1.47s 1.47s
4 189.000 183.149 6,079 3.48s 4.95s
5 189.000 185.567 16,339 7.44s 12.39s
SSBSP8
6 189.000 187.815 8,127 10.96s 23.35s
7 189.000 188.823 25,640 18.38s 41.73s
8 189.000 189.000 649 2.64s 44.37s
3 297.770 272.902 33 0.86s 0.86s
4 299.000 288.031 24,452 19.76s 20.62s
SSBSP12 5 299.000 290.640 195,160 145.87s 166.49s
6 299.000 292.697 507,940 394.78s 561.27s
7 299.000 294.289 815,078 716.90s 1,278.17s
5 468.000 428.544 256,375 476.74s 476.74s
SSBSP18
6 468.000 433.583 360,510 +1,000s +1,476.74s
SSBSP25 7 609.000 538.538 103,912 +1,000s +1,000s
SSBSP29 8 638.975 569.580 140,638 +1,000s +1,000s

n0 = O {u|Wu 6= ∅} . To solve the problem to global optimality, we use the sec-
ond criterion (n > |O|). However, all instances except the first one are very expensive to
solve, so we used a time limit of 1,000 seconds. Computational results are given in Ta-
ble 2.6. For each problem, the MOS-SST solution obtained with n priority-slots has lower
objective value than the MOS solution obtained with the same number of priority-slots as
stated previously. Overall results show that the MOS-SST time representation is much less
efficient than the MOS time representation.
2.7.3 MOS-FST Model
The MOS-FST model for the single-stage batch scheduling problem is derived from con-
straints (2.1)-(2.3), (2.15), and (2.25). Note that non-overlapping constraint (2.20) is not
included in the model as assignment constraint (2.25) is sufficient to enforce non-overlapping
constraints when processing times are constant (Dv = Dv ). Also, constraint (2.27) is not
included in the model as valid schedules containing time periods with no activity would
become infeasible.

XX
max Eiv
i∈T v∈W
X X
Ziv = 1 o∈O
i∈T v∈Wo
X X
i∈T v∈Wu
Siv = ti · Ziv i ∈ T, v ∈ W
 
 
X   X 

Dv + T Ru

Ziv + Zjv ≤1 i ∈ T, u ∈ U, δu (v) =
  H/n
v∈Wu 
 j∈T 

i−δu (v)<j<i

Ziv ∈ {0, 1} i ∈ T, v ∈ W
i−1
ti = ·H i∈T
n
Similarly to the MOS-SST model, it is not possible to develop symmetry-breaking con-

straints based on the isolated cliques of GN O for this model. To solve it, the multiplicative
H
approach is used and the initial number of priority-slots is set to n0 = 2 = 15. Due to
memory limitations we used the second stopping criterion (n > 1920) which does not guar-
antee global optimality of the solution. Computational results are given in Table 2.7. The
first iterations are always computationally inexpensive and lead to good feasible solutions,
which are improved in the next iterations. Although near-optimal solutions are obtained
quickly, the gap is never entirely closed due to the time discretization. It is also interesting
to note that most problems are solved at the root node.
2.7.4 SOS Model
The SOS time representation has not been studied in the case for batch scheduling prob-
lems as it requires to develop a problem-specific symmetry-breaking sequencing rule (see

Table 2.7: MOS-FST computational results for single-stage batch scheduling problems.
15 182.581 182.581 0 0.80s 0.80s
30 186.013 186.013 0 1.07s 1.87s
60 187.456 187.456 0 1.56s 3.43s
120 188.594 188.594 0 2.77s 6.20s
SSBSP8
240 188.719 188.719 0 6.07s 12.27s
480 188.813 188.813 0 16.89s 29.16s
960 188.907 188.907 0 52.92s 82.08s
1,920 188.954 188.954 0 183.32s 265.40s
15 288.482 288.482 0 1.02s 1.02s
30 292.671 292.671 0 1.43s 2.45s
60 295.466 295.466 0 2.05s 4.50s
120 296.393 296.393 0 3.70s 8.20s
SSBSP12
240 297.268 297.268 0 8.83s 17.03s
480 297.675 297.675 0 25.44s 42.47s
960 297.772 297.772 0 80.46s 122.93s
1,920 297.878 297.878 0 282.79s 405.72s
15 429.377 429.377 0 1.31s 1.31s
30 438.507 438.507 0 1.74s 3.05s
60 445.652 445.182 0 2.67s 5.72s
120 448.808 448.107 0 5.30s 11.02s
SSBSP18
240 450.308 449.607 0 14.15s 25.17s
480 451.339 450.782 0 39.90s 65.07s
960 451.652 451.157 0 138.20s 203.27s
1,920 451.800 451.297 0 459.25s 662.52s
15 537.852 536.490 0 1.39s 1.39s
30 554.266 536.490 0 2.05s 3.44s
60 566.344 566.085 0 3.36s 6.80s
120 573.141 573.091 0 7.07s 13.87s
SSBSP25
240 576.435 576.435 0 17.66s 31.53s
480 578.464 578.373 0 50.89s 82.42s
960 578.996 578.904 0 168.31s 250.73s
1,920 579.300 579.185 0 628.46s 879.19s
15 570.475 567.485 0 1.47s 1.47s
30 597.973 595.846 0 2.18s 3.65s
60 616.779 613.610 8 4.33s 7.98s
120 627.674 625.258 5 10.02s 18.00s
SSBSP29
240 631.863 630.988 0 22.60s 40.60s
480 634.336 633.426 0 75.74s 116.34s
960 635.119 634.019 512 389.57s 505.91s
1,920 635.741 no solution 400 +1,000s +1,505.91s

Priority-slots 1 2 3 Priority-slots 1 2 3 4 5 6 7 8 9 10 11 12
Unit 1 a b c Unit 1 a b c
Unit 2 d e f Unit 2 d e f
Unit 3 g h i Unit 3 g h i
Unit 4 j k l Unit 4 j k l
MOS Assignment SOS Assignment
Figure 2.8: Equivalent MOS and SOS assignments for SSBSP12.
chapter 4). However, the following comments can be made. If the number of batches to
be processed on each unit is known a priori, the problem reduces to a set of independent
sequencing problem for each unit. This is also true for in the MOS time representation. In
fact, it is possible to break symmetries in the SOS model in order to make it equivalent to
the MOS model as shown in Figure 2.8. In the MOS model, the 3 priority-slots are assigned
in parallel as each priority-slot is assigned to several operations, at most one for each unit.
In the SOS model, the 12 priority-slots are assigned in sequence as each priority-slot can
be assigned to at most one operation (i.e. at most one unit). Consider a sequencing rule
for the SOS model that reserves priority-slots 1, 2, and 3 to unit 1, priority-slots 4, 5,
and 6 to unit 2, and so on. Grey cells in Figure 2.8 are used to represent the assignments
that are forbidden by such a sequencing rule. In this case, there exists a bijective trans-
formation between MOS and SOS assignments as outlined by the lower-case letters. Both
assignments lead to identical precedence relations between assigned operations. A similar
idea can be used when the number of batches to be processed on each unit in not known
a priori. It should be noted that there are no non-trivial bicliques in the non-overlapping
graph GN O . Furthermore, the same comments apply to the multi-stage batch scheduling
problem studied in section 2.8.
2.7.5 Models Comparison
Figure 2.9 shows how the objective value of the best incumbent varies over time. A step
increase of the objective value corresponds to a new solution found during the solution
algorithm. Note that we do not include intermediate solutions found during each branch &

SSBSP12 SSBSP18
300
450
295
445
290
Objective Value
Objective Value
285 440
280
435
MOS MOS
MOS-SST MOS-SST
275 MOS-FST MOS-FST
Optimum 430 Optimum
270
1 10 100 1000 1 10 100 1000
CPU Time (s) CPU Time (s)
SSBSP25 SSBSP29
580
630
575 620
Objective Value
Objective Value
570 610
600
565
590
560 MOS MOS
MOS-FST 580 MOS-FST
Optimum Optimum
555
570
1 10 100 1000 1 10 100 1000

Figure 2.9: Comparison of time representations for single-stage batch scheduling problems.
bound search. Empty squares represent the time when the first stopping criterion is satisfied
while plain squares represent the time when the second stopping criterion is satisfied. This
applies only to the MOS model. From this figure, it is clear that the MOS model is superior
to the other models as it finds first feasible solutions very fast, and is able to find optimal
solutions and prove their optimality in reasonable time. The MOS-FST model compares
well in particular for finding first feasible solutions. It also finds near-optimal solutions in
reasonable time although it was slower than the MOS model for the instances that were
considered. Also, it is clear that the MOS-SST representation is not well suited to solve
these problems. For a more extended review of the performance of the different models
available to solve single-stage batch scheduling problems, the reader may refer to Méndez
and Cerdá (2003) and Castro and Grossmann (2006).

2.8 Multi-Stage Batch Scheduling Problem
We now extend the previous study to multi-stage batch scheduling problems from Pinto
and Grossmann (1995). The largest instance has 10 orders to be processed sequentially in 5
consecutive stages before identical due dates t = 500 hr. Twenty-five units are available to
process each order. Each unit is associated to one stage and can process only one order at
a time. A minimum unit-specific set-up time is required between any two orders processed
on the unit. Each order can only be processed on a subset of all units with order-and-unit-
specific processing times. Table 2.8 displays all required data. For each order, units that do
not have a corresponding processing time cannot be selected for this order. The objective,
as introduced by Pinto and Grossmann (1995), is to minimize total weighted earliness which
corresponds to maximizing a weighted summation of the end time of all orders in all stages.
The weights are stage dependent and defined as wl = 0.2 · l. Three instances have been
studied with 5, 8 and 10 orders, which we denote MSBSP5, MSBSP8 and MSBSP10. We
introduce the following specific sets.
• O ⊂ {o1 , . . . , o1 0} is the set of orders (O = {o1 , . . . , o5 } for MSBSP5).

• L = {l1 , . . . , l5 } is the set of stages
• U = {u1 , . . . , u25 } is the set of units.
• Uo is the set of units on which order o ∈ O can be processed. For instance, Uo1 =
{u1 , . . . , u17 , u20 , u21 , u23 , u24 }.
• Ou is the set of orders that can be processed on unit u ∈ U . For instance, Ou20 =
{o1 , o3 , . . . , o9 }.
• lu is the stage in which unit u ∈ U can be used. For instance, lu7 = l2
• Ul is the set of units which can be used in stage l ∈ L. For instance, Ul5 =
{u23 , u24 , u25 }
The set of operations can then be defined as follows. Exactly one operation is defined for
each order, each stage and each unit on which the order can be processed. This reduces
the number of indices from 3 to 1, although the combinatorial size of the problem remains

Table 2.8: Data for multi-stage batch scheduling problems.

Transition Order processing time (hr)
Stage Unit
Time (hr) 1 2 3 4 5 6 7 8 9 10
1 8 18.1 23 18.1 20 17 15 31 12 13 12
2 8 18.1 23 18.1 20 17 14 30 12 7 4
3 8 18.1 23 18.1 20 17 13 34 14 8 23
1
4 8 18.1 23 18.1 20 17 12 32 15 9 12
5 8 18.1 23 18.1 20 17 18 31
6 8 18.1 23 18.1 20 17 15 16 16 14
7 8 14 14 14 11 14 15 31 15 13
2 8 1 5 5 5 5 5 7 31 16 15 13
9 1 5 5 5 5 5 7 31 15 11 13
10 2.5 12 12 24 12 12 13 14 12 13 12
11 2.5 12 12 24 12 12 12 15 13 7 4
12 2.5 12 12 24 12 12 15 16 14 8 23
13 2.5 12 12 24 12 12 17 41 14 9 12
14 2.5 12 12 24 12 12 17 15 14
3
15 2.5 12 12 24 12 12 18 81 14 16 14
16 2.5 12 12 24 12 12 19 14 15 13
17 2.5 12 12 24 12 12 16 14 15 13
18 2.5 12 16 16 14 11 13
19 2.5 12 13 21 10 12 6
20 6 9.5 9.3 7.9 12.5 13.5 12 10 15
4 21 6 9.5 9.3 7.9 12.5 14 13 9 17 12
22 24 100 14.5 11 8 17 23
23 4 24 24 24 24 12 11 22 12
5 24 4 24 24 24 24 12 11 7 21 22
25 5 48 23 11 7 12
identical.
W = {(o, l, u) ∈ O × L × U, u ∈ Uo ∩ Ul }
= {o1 l1 u1 , . . . , o1 l1 u6 , o1 l2 u7 , . . . , o1 l2 u9 , o1 l3 u10 , . . .}
Figure 2.10 partially depicts the non-overlapping graph of MSBSP5. It contains 25 isolated
cliques each corresponding to one unit. We will denote Wu = {v = (o, l, u), o ∈ Ou , l = lu },
where u ∈ U , the set of operations executed on unit u, Wu is an isolated clique of GN O .
Given an order o ∈ O and a stage l ∈ L, exactly one execution of order o in stage l must be

o5 l1 u1
o4 l1 u1 o1 l4 u20 o5 l4 u20
o1 l1 u1 o2 l3 u18
o3 l1 u1 o3 l4 u20 o4 l4 u20
o2 l1 u1
Figure 2.10: Partial non-overlapping graph with isolated cliques for MSBSP5.
performed. Therefore, the set Wol = {v = (o, l, u), u ∈ Uo ∩Ul } has an associated cardinality
constraint of 1, that is NWol = NWol = 1. Furthermore, due to the multi-stage feature of
this problem, precedence constraints exist between operations belonging to different stages
of the same order. More precisely, for each order o ∈ O and each stage l ∈ L \ {l1 }, sets
Wo(l−1) and Wol must satisfy a precedence constraint PWo(l−1) Wol = 1.
2.8.1 MOS Model
The MOS model for the multi-stage batch scheduling problem is derived from constraints
(2.1)-(2.3), (2.10), (2.17), (2.20), (2.27), and (2.29)-(2.32).

X X
max 0.2 · l · Eiv
i∈T v∈W
v=(o,l,u)

X X
Ziv = 1 o ∈ O, l ∈ L
i∈T v∈Wol
X X
i∈T v∈Wu
X X X X
Eiv ≤ Siv o ∈ O, l ∈ L, l 6= 1
i∈T v∈Wo(l−1) i∈T v∈Wol
X
Ziv ≤ 1 i ∈ T, u ∈ U
v∈Wu
X
v∈Wu
X X
i∈T v∈Wu
i1 , i2 ∈ T, i1 < i2 , i2 > NWu , u ∈ U
i1 <i<i2
X X
≤ Si2 v + (H + T RWu ) · (1 − Z i2 v )
v∈Wu v∈Wu
X
Ziv ≥ 1 i∈T
v∈W
X X
Ziv ≤ Z(i−1)v i ∈ T, i 6= 1, u ∈ U
v∈Wu v∈Wu
X
Ziv = 0 i ∈ T, i > NWu , u ∈ U
v∈Wu
X
Ziv = 1 i ∈ T, i ≤ NWu , u ∈ U
v∈Wu
X
v∈Wu
X X
i∈T v∈Wu
i1 , i2 ∈ T, i1 < i2 , i2 ≤ NWu , u ∈ U
i1 <i<i2
X
≤ Si2 v
v∈Wu

Ziv ∈ {0, 1} i ∈ T, v ∈ W
Cardinality bounds for units are defined similarly to the single-stage case. The minimum
cardinality bound is obtained by considering the number of orders executed on units different

Table 2.9: MOS computational results for multi-stage batch scheduling problems.
2 6,827.56 6,827.56 0 0.98s 0.98s
3 6,828.76 6,828.76 58 1.66s 2.64
MSBSP5
4 6,996.76 no solution 0 2.26s 4.90s
5 6,996.76 no solution 0 3.23s 8.13s
3 10,985.61 10,985.16 837 16.70s 16.70s
4 11,364.67 10,986.36 1,149 59.87s 76.57s
5 11,364.67 no solution 500 30.34s 106.91s
MSBSP8
6 11,364.67 no solution 2,200 156.14s 263.05s
7 11,364.67 no solution 1,500 243.65s 506.70s
8 11,364.67 no solution 2,200 331.18s 837.88s
MSBSP10 4 13,627.41 13,581.16 34,340 +1,000s +1,000s
than u ∈ U that belong to the same stage lu .

( )
X X
NWu = max γv γv ∈ {0, 1},
(Dv + T RWu ) · γv ≤ H + T RWu u∈U
v∈Wu v∈Wu
 

 


 


 X 

1

NWu = max Wu , O −
min{n, NWu0 } u∈U

u0 ∈U
 

 0

u 6
= u

 

 
lu0 =lu
The multi-stage batch scheduling problems are solved using the additive approach with the
second stopping criterion (n > maxu Nu ) which guarantees global optimality of the solution.

The initial number of priority-slots is set to n0 = O minl∈L {u|u ∈ Ul } , which is the
minimum number of orders at least one unit has to process in each stage, more precisely
in bottleneck stages. Computational results are given in Table 2.9. Similarly to the single-
stage case, the results show that the first stopping criterion leads to significant reduction
of CPU times compared to the second. The instance with 10 orders cannot be solved to
global optimality although feasible solutions are obtained quickly using 4 priority-slots. The
branch & bound search is stopped after 1,000 seconds with a remaining 0.09% optimality
gap. A 1% optimality gap can be achieved in 219.08 seconds.

2.8.2 MOS-SST Model
The MOS-SST model for the multi-stage batch scheduling problem is derived from con-
straints (2.1)-(2.3), (2.10), (2.17), (2.20)-(2.22), and (2.27).
X X
max 0.2 · l · Eiv
i∈T v∈W
v=(o,l,u)

X X
Ziv = 1 o ∈ O, l ∈ L
i∈T v∈Wol
X X
i∈T v∈Wu
X X X X
Eiv ≤ Siv o ∈ O, l ∈ L, l 6= 1
X
Ziv ≤ 1 i ∈ T, u ∈ U
v∈Wu
X
v∈Wu
X X
i∈T v∈Wu
i1 , i2 ∈ T, i1 < i2 , u ∈ U
i1 <i<i2
X X
≤ Si2 v + (H + T RWu ) · (1 − Z i2 v )
v∈Wu v∈Wu
ti−1 ≤ ti i ∈ T, i 6= 1
X
Siv ≤ ti i ∈ T, u ∈ U
v∈Wu
X X
Siv ≥ ti − H · (1 − Ziv ) i ∈ T, u ∈ U
v∈Wu v∈W 0
X X X X
Zjv ≥ Zjv i ∈ T, o ∈ O, l ∈ L, l 6= 1
j∈T v∈Wo(l−1) j∈T v∈Wol
j<i j≤i
X
Ziv ≥ 1 i∈T
v∈W
Ziv ∈ {0, 1} i ∈ T, v ∈ W
ti ∈ [0, H] i∈T

Table 2.10: MOS-SST computational results for multi-stage batch scheduling problems.
MSBSP5 6 6,617.96 6,161.40 21,915 21.89s 23.34s
7 6,996.76 6,448.62 285,564 +1,000s +1,023.34s
MSBSP8 6 no solution 2.07s 3.88s
7 11,363.98 99,79.40 21,491 +1,000s +1,003.88s
MSBSP10
8 14,255.76 12,560.20 2,172 +1,000s +1,007.92s
The additive approach is used to solve the model and the initial number of priority-slots is
set to n0 = |L|, the number of stages, as for each order stage processing operations must
start at different dates. No problem instance was solved to global optimality, so we use
a time limit of 1,000 seconds. Computational results are given in Table 2.10. As for the
single-stage case, the MOS-SST time representation is much less efficient than the MOS
time representation. For the instance with 10 orders the MOS-SST model returns a worse
solution than the MOS model in 1,000 seconds (13.31% optimality gap remaining).
2.8.3 MOS-FST Model
The MOS-FST model for the multi-stage batch scheduling problem is derived from con-
straints (2.1)-(2.3), (2.10), (2.15), and (2.25).

X X
max 0.2 · l · Eiv
i∈T v∈W
v=(o,l,u)

X X
Ziv = 1 o∈O
i∈T v∈Wo
X X
i∈T v∈Wu
X X X X
Eiv ≤ Siv o ∈ O, l ∈ L, l 6= 1
 

 

X  X  Dv + T Ru
Z + Zjv ≤1 i ∈ T, u ∈ U, δu (v) =
 iv  H/n
v∈Wu 
 j∈T 

i−δu (v)<j<i

Ziv ∈ {0, 1} i ∈ T, v ∈ W
i−1
ti = ·H i∈T
n
To solve this model, the multiplicative approach is used and the initial number of priority-

slots is set to n0 = H 4 = 125. We use a time limit of 1,000s as a stopping criterion.
Computational results are given in Table 2.11. As for the single-stage case, the MOS-FST
time representation quickly finds near-optimal solutions in the first iterations. The instance
with 10 orders was not solved due to memory limitations.
2.8.4 Models Comparison
Figure 2.11 shows how the objective value of the best incumbent varies over time. From this
figure, it is clear that the MOS model is superior to the other models as it finds first feasible
solutions very fast and is able to find optimal solutions in reasonable time. The MOS-FST
model is significantly more expensive than previous single-stage cases as the scheduling

2.9 Conclusion
Table 2.11: MOS-FST computational results for multi-stage batch scheduling problems.
125 6,820.61 6,800.60 0 16.84s 16.84s
250 6,826.21 6,819.00 0 39.32s 56.16s
MSBSP5 500 6,827.76 6,824.80 499 155.83s 211.99s
1,000 6,828.76 6,828.00 509 619.69s 831.68s
2,000 6,828.76 6,828.10 0 +1,000s +1,831.68s
125 10,951.49 10,925.80 503 71.54s 71.54s
MSBSP8
250 10,965.66 10,956.60 31,692 +1,000s +1,071.54s
MSBSP5 MSBSP8
11000
6800
10950
6600
Objective Value
Objective Value
6400
10900
6200
MOS MOS
10850
MOS-SST MOS-FST
6000 MOS-FST Optimum
Optimum
5800 10800
1 10 100 1000 10 100 1000 10000
Figure 2.11: Comparison of time representations for multi-stage batch scheduling problems.
horizon is much longer (500 hours as opposed to 30 hours). Therefore, more priority-slots
need to be postulated which makes the model size grow significantly. Janak et al. (2004),
Castro et al. (2006) and Liu and Karimi (2007) have developed and studied several models
that can be applied to different types of multi-stage batch scheduling problems. Their
computational results show similar trends.
2.9 Conclusion
In this chapter, we have presented four different time representations that in different forms
have been previously introduced to solve process scheduling problems. Using the common
concept of priority-slot, it was shown that it is possible to derive relationship results between
these time representations. Additionally, generic scheduling constraints were presented with

2.9 Conclusion
corresponding strengthened formulations that rely on exploiting the non-overlapping graph

structure of these problems through maximum cliques and bicliques. These formulations
were developed as well as specific solution algorithms for each model. Given the unifying
framework of these models, we were able to apply them to batch scheduling of single and
multi-stage plants.
Intuitively, the representation that requires the fewest number priority-slots is the most
computationally efficient. In practice, the MOS time representation proved to be superior
to the other time representations on the two types of problem tested due to the smaller
number of priority-slots and symmetry-breaking constraints used.
In terms of computational performance, the discrete-time MOS-FST representation is
comparable to the MOS time representation due to the high efficiency of constraint (2.25)
for problems with constant processing times and time horizon of reasonable length.
On the other hand, the MOS-SST time representation performs poorly as it requires many
more priority-slots than the MOS representation but does not improve the LP relaxation
of the MILP model.
Finally, it should be noted that precedence-based formulations (Méndez et al., 2001) are
very good alternatives to priority-slot based representations. These formulations tend to
be limited to the context of batch scheduling problems without inventory tracking (Méndez
et al., 2006a). The corresponding models often have weak LP relaxations as they require a
large proportion of big-M constraints, but they are very compact and therefore easily solved
by current commercial MILP solvers.

Chapter 3
Short-Term Scheduling of Crude-Oil
Operations
3.1 Introduction
The optimal scheduling of crude-oil operations has been studied since the 90’s and has been
shown to lead to multimillion dollar benefits by Kelly and Mann (2003) as it is the first
stage of the oil refining process. It involves crude-oil unloading from crude marine vessels
(at berths or jetties) or from a pipeline to storage tanks, transfers from storage tanks to
charging tanks and atmospheric distillations of crude-oil mixtures from charging tanks.
The crude is then processed in order to produce basic products which are then blended
into gasoline, diesel, and other final products. Assuming that the schedule of crude supply
and production demands are determined by the long-term refinery planning, this chapter
studies the short-term scheduling problem maximizing gross margins of crude-oil mixtures.
Shah (1996) proposed to use mathematical programming techniques to find crude-oil
schedules exploiting opportunities to increase economic benefits. Lee et al. (1996) considered
a crude-oil scheduling problem involving crude unloading at berths, developed a discrete-
time MINLP model, and solved an MILP relaxation of the model. Later, Wenkay et al.
(2002) improved the model and proposed an iterative approach to solve the MINLP model,
taking into account the nonlinear blending constraints. Pinto et al. (2000), Moro and
Pinto (2004), and Reddy et al. (2004) used a global event formulation to model refinery
systems involving crude-oil unloading from pipeline or jetties. The scheduling horizon is
divided into fixed length sub-intervals, which are then divided in several variable length
Chapter 3. Short-Term Scheduling of Crude-Oil Operations 67

3.2 Problem Statement
time-slots. In parallel, Jia et al. (2003) developed an operation-specific event model and
applied it to the problems introduced by Lee et al. (1996) using a linear approximation
of storage costs. A comparison of computational performances between both continuous-
time and discrete-time models was given showing significant decreases in CPU time. Also,
solutions that are not guaranteed to be globally optimal were obtained using standard
MINLP algorithms. Recently, Furman et al. (2007) presented a more accurate version of the
event-point formulation, and Karuppiah et al. (2008) later addressed the global optimization
of this model using an outer-approximation algorithm where the MILP master problem is
solved by a Lagrangean decomposition. While rigorous, this method can be computationally
expensive.
The aim of this chapter is to apply different time representations and mathematical mod-
els to the crude-oil scheduling problem introduced by Lee et al. (1996). First, the problem
definition is given. Next, the MOS, MOS-SST, and MOS-FST models (see chapter 2) are
derived and a simple solution method to solve them is presented. Finally, computational
results are given to show the effectiveness of the proposed model and solution method. As
it requires additional work to break its inherent symmetries, the SOS model will be detailed
in chapter 4.
3.2.1 General Description
This work is aimed at solving the four examples of refinery crude-oil operations problems
introduced in Lee et al. (1996), respectively denoted COSP1, . . . , COSP4. Each crude-oil
operations system is composed of four types of resources: crude marine vessels, storage
tanks, charging tanks and crude distillation units (CDUs). Three types of operations, all
transfers between resources, are allowed: crude-oil unloadings from marine vessels to storage
tanks, transfers between tanks, and transfers from charging tanks to CDUs.
A crude-oil scheduling problem is defined by: (a) a time horizon, (b) arrival time of marine

vessels, (c) capacity limits of tanks, (d) transfer flowrate limitations, (e) initial composition
of vessels and tanks, (f) crude property specifications for distillations, (g) and demands for
each crude blend.
The logistics constraints of the problem are defined as follows.
(i) Only one berth is available at the docking station for vessel unloadings,
(ii) simultaneous inlet and outlet transfers on tanks are forbidden,
(iii) a tank may charge only one CDU at a time,
(iv) a CDU can be charged by only one tank at a time,
(v) and CDUs must be operated continuously throughout the scheduling horizon.
In this chapter, the goal is to determine how many times each operation will be executed,
when it will be performed (start time and duration), and the volume of crude to be trans-
ferred in order to maximize the gross margins of distilled mixed oil. The gross margin of
each crude can be estimated from the sales income of the final products minus its purchase
value and related refining operational costs. Depending on the market value of the different
crudes and final products, the model tries to process the most profitable crudes and to
store the other crudes. As opposed to the work of Lee et al. (1996), sea waiting, unloading,
storage and CDU switching costs are ignored.
As CDU switches between different crude blends are costly, it is considered that the
number of distillation operations is bounded. The bounds on the number of distillation
operations can be set by the user, or for the lower bound, it can be obtained by solving a
model minimizing the number of distillation operations. Typically, the number and type of
marine vessels is determined at the planning level as well as demands in each crude blend.
Due to the logistics constraint (ii), inventory tracking in tanks can be performed by
looking at the sequence of inlet and outlet operations only, and not at the sequence of
start and end events of such operations. Indeed, for each tank, the scheduling solution
can be decomposed into successions of inlet and outlet states during which one or several
operations are performed. Therefore, it is only necessary to enforce capacity limitations

1 4
Outlet
4
3
Inlet
2
Time
Inventory
Time
Inventory
1 2 3 4 Priority-slots
Figure 3.1: Example of tank schedule.
at the transitions between these states. Figure 3.1 depicts an example of tank schedule
with inlet and outlet states, time-based and priority-slot-based inventory profiles. Each
transfer activity is represented by a horizontal bar labeled with the corresponding priority-
slot. Note that the two inventory profiles do not coincide due to the use of different x-axes.
In particular, in the priority-slot-based profile, it is not possible to distinguish the two
consecutive inventory decreases corresponding to the two operations assigned to priority-
slot 4. Under assumption (ii), it is sufficient to enforce tank capacity limitations just before
transition slots as it corresponds to inventory upper and lower peaks. In practice, they are
enforced just before all priority-slots.
3.2.2 Case Study
Figure 1.4 depicts the refinery configuration for COSP1. The scheduling horizon is composed
of 8 days, and two marine vessels are scheduled to arrive at the beginning of day 1 (t = 0)
and day 5 (t = 4), and contain 1 million bbl of crude-oil A and B, respectively. There is one
CDU which has to process 1 million bbl of each crude-oil mixture, X and Y. The property

Table 3.1: Data for COSP1.

Scheduling horizon 8 days
Vessels Arrival time Composition Amount of crude (Mbbl)
Vessel 1 0 100% A 1,000
Vessel 2 4 100% B 1,000
Storage tanks Capacity (Mbbl) Initial composition Initial amount of crude (Mbbl)
Tank 1 [0, 1,000] 100% A 250
Tank 2 [0, 1,000] 100% B 750
Charging tanks Capacity (Mbbl) Initial composition Initial amount of crude (Mbbl)
Tank 1 (mix X) [0, 1,000] 100% C 500
Tank 2 (mix Y) [0, 1,000] 100% D 500
Crudes Property 1 (sulfur concentration) Gross margin ($/bbl)
Crude A 0.01 9
Crude B 0.06 4
Crude C 0.02 8
Crude D 0.05 5
Crude mixtures Property 1 (sulfur concentration) Demand (Mbbl)
Crude mix X [0.015, 0.025] [1,000, 1,000]
Crude mix Y [0.045, 0.055] [1,000, 1,000]
Unloading flowrate [0, 500] Transfer flowrate [0, 500]
Distillation flowrate [50, 500] Number of distillations 3
that is being tracked is the weight fraction of sulfur, 0.01 for crude A and 0.06 for crude B.
The sulfur concentration of crude mix X and Y should be in the ranges [0.015, 0.025] and
[0.045, 0.055], respectively. The two storage tanks initially contain 250,000 bbl of crude A
and 750,000 bbl of crude B. The two charging tanks initially contain 500,000 bbl of crude
C and D, with sulfur concentrations of 0.02 and 0.05, respectively. Crudes C and D are in
fact blends of crudes A and B that match the sulfur concentration specifications for crude
mix X and Y. The gross margins of crudes A, B, C, and D are 9 $/bbl, 4 $/bbl, 8 $/bbl,
5 $/bbl, respectively. Gross margins for crudes C and D have been calculated from gross
margins of crudes A and B. The data for COSP1 is given in Table 3.1. Flowrate limitations
are expressed in Mbbl/day.
Figure 3.2 depicts the Gantt chart of a sub-optimal solution for COSP1 with a profit
of $6,925,000 that might be obtained by using heuristics. Each task is represented by a
horizontal bar and each row corresponds to a specific operation. Tank inventories are also
displayed, showing that tank capacity limits are satisfied. Figure 3.3 shows in contrast the
Gantt chart of the optimal solution obtained, and proved optimal (i.e. 0% optimality gap),

Operation
Unloading 1
Unloading 2
Transfer 3
Transfer 4
Transfer 5
Transfer 6
Distillation 7
Distillation 8
0 1 2 3 4 5 6 7 8
Storage tank 1
Storage tank 2
Charging tank 1
Charging tank 2
Figure 3.2: Sub-optimal schedule for COSP1 (profit: $6,925,000).
with the proposed approach as shown in section 3.5. It corresponds to a profit of $7,975,000,
which represents a 13.2% increase. Clearly, finding such a solution is non-trivial.
In this section, we present three mathematical models based on the MOS, MOS-SST, and
MOS-FST time representations (see chapter 2). They are all based on the same sets,
parameters and variables.
3.3.1 Sets
The following sets will be used in the model.
• T = {1, . . . , n} is the set of priority-slots

• W is the set of all operations: W = WU ∪ WT ∪ WD (W = {1..8} for COSP1,
see Figure 1.4)

Operation
Unloading 1
Unloading 2
Transfer 3
Transfer 4
Transfer 5
Transfer 6
Distillation 7
Distillation 8
0 1 2 3 4 5 6 7 8
Storage tank 1
Storage tank 2
Charging tank 1
Charging tank 2
Figure 3.3: Optimal schedule for COSP1 (profit: $7,975,000).
• WU ⊂ W is the set of unloading operations (WU = {1, 2} for COSP1)

• WT ⊂ W is the set of tank-to-tank transfer operations (WT = {3, 4, 5, 6} for COSP1)
• WD ⊂ W is the set of distillation operations (WD = {7, 8} for COSP1)
• R is the set of resources (i.e. tanks, units): R = RV ∪ RS ∪ RC ∪ RD
• RV ⊂ R is the set of vessels
• RS ⊂ R is the set of storage tanks
• RC ⊂ R is the set of charging tanks
• RD ⊂ R is the set of distillation units
• Ir ⊂ W is the set of inlet transfer operations on resource r
• Or ⊂ W is the set of outlet transfer operations on resource r
• C is the set of products (i.e. crudes)
• K is the set of product properties (e.g. crude sulfur concentration)

3.3.2 Parameters
The following parameters are given.
• H is the scheduling horizon

• [Vvt , Vvt ] are bounds on the total volume transferred during transfer operation v ; in
all instances, Vvt = 0 for all operations except unloadings for which Vvt = Vvt is the
volume of crude in the marine vessel
• [ND , ND ] are the bounds on the number of distillations
• [F Rv , F Rv ] are flowrate limitations for transfer operation v
• Sv is the minimum start time of unloading operation v ∈ WU (i.e. arrival time of the
corresponding vessel)
• [xvk , xvk ] are the limits of property k of the blended products transferred during
operation v
• xck is the value of the property k of crude c
• [Ltr , Ltr ] are the capacity limits of tank r
• Lt0r is the initial total level in tank r
• L0rc is the initial crude level in tank r for crude c
• [Dr , Dr ] are the bounds of the demand on products to be transferred out of the
charging tank r during the scheduling horizon
• Gc is the individual gross margin of crude c
3.3.3 Variables
The variables used in the model are composed of binary assignment variables, and contin-
uous time, operation and resource variables.
• Assignment variables Ziv ∈ {0, 1} i ∈ T, v ∈ W

Ziv = 1 if operation v is assigned to priority-slot i, Ziv = 0 otherwise.
• Time variables Siv ≥ 0, Div ≥ 0, Eiv ≥ 0 i ∈ T, v ∈ W
Siv is the start time of operation v if it is assigned to priority-slot i, Siv = 0 otherwise.

Div is the duration of operation v if it is assigned to priority-slot i, Div = 0 otherwise.

Eiv is the end time of operation v if it is assigned to priority-slot i, Eiv = 0 otherwise.
• Operation variables Vivt ≥ 0 and Vivc ≥ 0 i ∈ T, v ∈ W, c ∈ C
Vivt is the total volume of crude transferred during operation v if it is assigned to
priority-slot i, Vivt = 0 otherwise.
Vivc is the volume of crude c transferred during operation v if it is assigned to priority-
slot i, Vivc = 0 otherwise.
• Resource variables Ltir and Lirc i ∈ T, r ∈ R, c ∈ C
Ltir is the total accumulated level of crude in tank r ∈ RS ∪ RC before the operation
assigned to priority-slot i.
Lirc is the accumulated level of crude c in tank r ∈ RS ∪ RC before the operation
assigned to priority-slot i.
It should be noted that the crude composition of blends in tanks is tracked instead of
their properties. The distillation specifications are later enforced by calculating a posteriori
the properties of the blend in terms of its composition. For instance, in problem COSP1, a
blend composed of 50% of crude A and 50% of crude B has a sulfur concentration of 0.035
which does not meet the specification for crude mix X nor for crude mix Y.
3.3.4 Objective Function
The objective is to maximize the gross margins of the distilled crude blends. Using the
individual gross margins Gc , it is written as follows.
X X XX
max Gc · Vivc
i∈T r∈RD v∈Ir c∈C
3.3.5 General Constraints
In this section, we present a set of general constraints which can be used in any of the time
representations studied in chapter 2.

The following variable bound and time constraints (3.1) are used.
Siv ≥ Sv · Ziv i ∈ T, v ∈ WU (3.1a)
Eiv ≤ H · Ziv i ∈ T, v ∈ W (3.1b)
Eiv = Siv + Div i ∈ T, v ∈ W (3.1c)
The following unloading and distillation cardinality constraints (3.2) are used.
XX
Ziv = 1 r ∈ RV (3.2a)
i∈T v∈Or
X X
ND ≤ Ziv ≤ ND (3.2b)
i∈T v∈WD
The following unloading precedence constraints (3.3) are used to make sure that crude
vessels unload their content according to their respective order of arrival at the refinery.
The notation r1 < r2 denotes that vessel r1 is scheduled to arrive at the refinery before
vessel r2 .
X X X X
Eiv ≤ Siv r1 , r2 ∈ RV , r1 < r2 (3.3a)
i∈T v∈Or1 i∈T v∈Or2
X X X X
Zjv ≥ Zjv i ∈ T, r1 , r2 ∈ RV , r1 < r2 (3.3b)
j∈T v∈Or1 j∈T v∈Or2
j<i j≤i
The following constraint (3.4) states that each CDU must be operated without inter-
ruption throughout the scheduling horizon. As CDUs perform only one operation at a
time, the continuous operation constraint is defined by equating the sum of the duration of
distillations to the time horizon.
XX
Div = H r ∈ RD (3.4)
i∈T v∈Ir
The following variable constraints (3.5) are directly derived from the definition of volume
and level variables.

Vivt ≤ Vvt · Ziv i ∈ T, v ∈ W (3.5a)
Vivt ≥ Vvt · Ziv i ∈ T, v ∈ W (3.5b)

X
Vivt = Vivc i ∈ T, v ∈ W (3.5c)
c∈C
X X X X
Ltir = Lt0r + Vivt − Vivt i ∈ T, r ∈ R (3.5d)
j∈T,j<i v∈Ir j∈T,j<i v∈Or
X X X X
Lirc = L0rc + Vivc − Vivc i ∈ T, r ∈ R, c ∈ C (3.5e)
j∈T,j<i v∈Ir j∈T,j<i v∈Or
X
Ltir = Lirc i ∈ T, r ∈ R (3.5f)
c∈C
The following operation constraints (3.6) include:
1. flowrate limitations that link volume and duration variables

2. property specifications, assuming that the mixing rule is linear
3. composition constraints, which are nonlinear
F Rv · Div ≤ Vivt ≤ F Rv · Div i ∈ T, v ∈ W (3.6a)

X
xvk · Vivt ≤ xck Vivc ≤ xvk · Vivt i ∈ T, v ∈ W, k ∈ K (3.6b)
c∈C
Vivc · Ltir = Lirc · Vivt i ∈ T, r ∈ R, v ∈ Or , c ∈ C (3.6c)
It has been shown (Quesada and Grossmann, 1995b) that processes including both mixing
and splitting of streams cannot be expressed as a linear model. Mixing occurs when two
streams are used to fill a tank and is expressed linearly in constraints (3.5d-3.5e). Splitting
occurs when partially discharging a tank, resulting in two parts: the remaining content of
the tank and the transferred products. This constraint is nonlinear. The composition of
the products transferred during a transfer operation must be identical to the composition
of the origin tank. Note that constraint (3.6c) is a bilinear reformulation of the original
constraint (3.7) and is correct even when operation v is not assigned to priority-slot i, as

then Vivt = Vivc = 0.
Lirc Vivc
= t i ∈ T, r ∈ R, v ∈ Or , c ∈ C (3.7)
Ltir Viv
The following resource constraints (3.8) models inventory capacity limitations. As si-
multaneous charging and discharging of tanks is forbidden, these constraints are sufficient.
Ltr ≤ Ltir ≤ Ltr i ∈ T, r ∈ RS ∪ RC (3.8a)
0 ≤ Lirc ≤ Ltr i ∈ T, r ∈ RS ∪ RC , c ∈ C (3.8b)

XX XX
Ltr ≤ Lt0r + Vivt − Vivt ≤ Ltr r ∈ RS ∪ RC (3.8c)
i∈T v∈Ir i∈T v∈Or
XX XX
0 ≤ L0rc + Vivc − Vivc ≤ Ltr r ∈ RS ∪ RC , c ∈ C (3.8d)
i∈T v∈Ir i∈T v∈Or
The following demand constraints (3.9) defines lower and upper limits, Dr and Dr , on
the total volume of products transferred out of each charging tank r during the scheduling
horizon.
XX
Dr ≤ Vivt ≤ Dr r ∈ RC (3.9)
i∈T v∈Or
3.3.6 Strengthened Constraints
Figure 3.4 displays the refinery system for problems COSP2 and COSP3. Each labelled arc
corresponds to a transfer operation. Figure 3.5 displays the corresponding non-overlapping
graph. It contains 4 maximal cliques of 3 operations: {1, 2, 3}, {5, 12, 13}, {7, 12, 13}, and
{9, 12, 13}. In particular, maximal clique {5, 12, 13} is due to non-overlapping constraints
(ii) and (iii) which makes it non trivial to detect. Therefore, a generic maximal clique
finding algorithm allows generating non-trivial strengthened non-overlapping constraints,
which is the main objective of the non-overlapping graph representation.
The maximum cliques of GN O are used to derive the following assignment and scheduling

Figure 3.4: Refinery crude-oil scheduling system for problem COSP2 and COSP3.
Crude Vessels Storage Tanks Charging Tanks CDUs
1 4
5 11
6 12
2 7
8 13
9 14
3 10
constraints, which are identical to constraints (2.17) and (2.19).

X
Ziv ≤ 1 i ∈ T, W 0 ∈ clique(GN O ) (3.10)
v∈W 0
X X X
Ei1 v + Div
X X
≤ Si2 v + H · (1 − Z i2 v )
v∈W 0 v∈W 0
3.3.7 Symmetry-Breaking Constraint for MOS Models
As mentioned by Kallrath (2002), degeneracies and symmetries often cause scheduling prob-
lems to be difficult to solve. The author suggests that nonlinearities involved in refinery
scheduling problems may reduce these effects. In this work, we consider the concept of
static symmetry-breaking constraints as presented by Margot (2008) in order to break the
symmetries that can be detected in the MOS model. In particular, the following generic
constraint is used. It states that an operation v cannot be assigned to priority-slot i if
no other non-overlapping operation is assigned to priority-slot i − 1. Indeed, in this case
operation v can be assigned to slot i − 1 instead of slot i with no impact on the scheduling
solution.
X
Ziv ≤ Z(i−1)v0 i ∈ T, i > 1, v ∈ W (3.12)
v 0 ∈W
N Ovv0 =1

3.4 Solution Method
4
1
5
11
6
12
2 7
13
8
14
9
3
10
Figure 3.5: Non-overlapping graph for crude-oil examples 2 and 3.
This constraint does not make any assumptions on the characteristics of the problem as op-
posed to symmetry-breaking constraint (2.29) which is specific to the case of batch schedul-
ing problems.
3.3.8 Full Models
The full mathematical formulations for each time representation presented in chapter 2 are
described in appendix C. The details for the SOS model are given in chapter 4.
3.4 Solution Method
The non-convex MINLP models described in the previous section can be solved using any
generic MINLP solver such as DICOPT (outer-approximation method) or BARON (global
solver using a spatial branch-and-bound search). The former code may converge to a poor

3.4 Solution Method
MILP
maximize
objective
s.t.
all constraints
except composition constraints
i
Fix assignment variables Zv
NLP
maximize
objective
s.t.
all constraints
Figure 3.6: Two step decomposition strategy.
sub-optimal solution, while the latter can be prohibitively expensive to use. Therefore, a
simple two-step procedure has been implemented (see Figure 3.6), leading to local optimal
solutions with an estimation of the optimality gap.
In the first step, a linear MILP relaxation of the model obtained after dropping the
nonlinear composition constraint (3.6c) is solved. It should be noted that material balances
are still enforced in this MILP relaxation through constraints (3.5c-3.5f). Furthermore,
the number of priority-slots n is selected at this stage using the additive or multiplicative
approach as described in section 2.6.
The solution returned by the MILP solver may violate the bilinear composition constraint
(3.6c). In this case, the binary variables Ziv are fixed, which means that the sequence of
operations is fixed, and the resulting nonlinear programming (NLP) model is solved using
the solution of the MILP as a starting point. This NLP model contains the same constraints
as in the MILP model plus the nonlinear constraint (3.6c). The solution obtained at this
stage might not be the optimum of the full model, but the optimality gap can be estimated
from the upper bound given by the MILP solution and the lower bound given by the NLP

3.5 Computational Results
solution.
The NLP model can be solved using either a global NLP solver (e.g. BARON) or a local
NLP solver (e.g. CONOPT). In the former case, the solver will return the global optimum
of the NLP if it is feasible. One could then add integer cuts to the MILP in order to close
the gap between the upper level and the lower level of the decomposition. This method will
return the global optimum of the MINLP. In the latter case, the NLP being non-convex,
several different locally optimal solution may be obtained depending on the starting point.
Thus, there is no proof that the iterative procedure will return the global optimum or even
a feasible solution to the MINLP, unless the NLP returns an objective value identical to the
MILP’s in which case global optimality is proved. As shown in section 3.5, the solutions
obtained at the first iteration are near-optimal so the procedure has been stopped at this
stage.
All experiments have been performed on an Intel Xeon 1.86GHz processor with GAMS as
modeling language, CPLEX 11 as MILP solver and CONOPT 3 as local NLP solver. The
CPU limit for solving the MILP has been set to 1,000s. The scheduling systems and data
for the four crude-oil operations scheduling problems from Lee et al. (1996) are given in
appendix B.
3.5.1 Scheduling Results
Figures 3.7 and 3.8 depicts the solutions obtained for problems COSP2 and COSP3 using
the proposed approach. The solution for COSP2 is proved optimal while the solution for
COSP3 is found within a 2.3% optimality gap. As there is no incentive to keep inventory
in the charging tanks, their final level is in general close to zero. However, the crude-oil
that arrived late to the refinery is mostly kept in the storage tanks. This leads to higher
transfer activity at the beginning than at the end of the scheduling horizon.

Operation
Unloading 1
Unloading 2
Unloading 3
Transfer 4
Transfer 5
Transfer 6
Transfer 7
Transfer 8
Transfer 9
Transfer 10
Distillation 11
Distillation 12
Distillation 13
Distillation 14
0 1 2 3 4 5 6 7 8 9 10
Storage tank 1
Storage tank 2
Storage tank 3
Charging tank 1
Charging tank 2
Charging tank 3
Figure 3.7: Optimal schedule for COSP2 (profit: $10,117,000).
The main uncertain parameter in refinery crude-oil scheduling problems is the arrival time
of tankers. The expected date of arrival of these marine vessels is known long in advance,
but is also subject to many changes before it actually arrives at the refinery. Figure 3.9
shows the optimal schedule for COSP2 when vessels are scheduled to arrive one day later,
leading to a 3.4% profit decrease (from $10,117,00 to $9,775,000). It should be noted that
the sequence of distillation operations is different in this case, and that there is a higher
transfer activity in the very beginning of the scheduling horizon. However, this schedule
assumes that the late arrival of vessels is known at time t = 0.

Operation
Unloading 1
Unloading 2
Unloading 3
Transfer 4
Transfer 5
Transfer 6
Transfer 7
Transfer 8
Transfer 9
Transfer 10
Distillation 11
Distillation 12
Distillation 13
Distillation 14
0 1 2 3 4 5 6 7 8 9 10 11 12
Storage tank 1
Storage tank 2
Storage tank 3
Charging tank 1
Charging tank 2
Charging tank 3
Figure 3.8: Schedule obtained for COSP3 within 2.3% optimality gap (profit: $8,540,000).
Assuming that the exact arrival dates are known slightly later (t = ε 1), the initial
decisions are fixed, so that operations 12, 14, 4, and 6 (which start at time t = 0 in the
original solution) are forced to start at time t = 0, but with variable durations. In this
case, the optimal solution is depicted in Figure 3.10. This solution is very similar to the
original optimal solution (Figure 3.7) as most transfers are simply delayed, although the
profit is reduced to $9,609,000. Only transfer operation 10 from t = 8 to t = 8.5 is no
longer used. The fact that the exact arrival of vessels is known only at t = ε leads to a 1.7%
profit decrease (from $9,775,000 to $9,609,000) compared to the case where it is known at
t = 0. This result shows that the initial decisions taken at t = 0 do not lead to a large
under-optimization in case of late vessel arrivals.

Operation
Unloading 1
Unloading 2
Unloading 3
Transfer 4
Transfer 5
Transfer 6
Transfer 7
Transfer 8
Transfer 9
Transfer 10
Distillation 11
Distillation 12
Distillation 13
Distillation 14
0 1 2 3 4 5 6 7 8 9 10
Storage tank 1
Storage tank 2
Storage tank 3
Charging tank 1
Charging tank 2
Charging tank 3
Figure 3.9: Optimal schedule for COSP2 with late vessel arrivals (profit: $9,775,000).
3.5.2 Performance of the MOS Model
The MILP relaxation of the MOS model is solved with the additive approach (see chapter 2)
using the second stop criterion (∆ ≤ 0). The initial number of priority-slots is set to n0 = 1.
Computational results are given in Table 3.2. All instances are solved within 20 seconds
with a few priority-slots. The NLP is modeled using the optimal number of priority-slots
determined at the MILP stage and is always solved in less than one second. The most
difficult instance, that is COSP2, is the one requiring the most priority-slots although it is
not the largest in size. The optimality gap is always small (less than 3%) even though the

Operation
Unloading 1
Unloading 2
Unloading 3
Transfer 4
Transfer 5
Transfer 6
Transfer 7
Transfer 8
Transfer 9
Transfer 10
Distillation 11
Distillation 12
Distillation 13
Distillation 14
0 1 2 3 4 5 6 7 8 9 10
Storage tank 1
Storage tank 2
Storage tank 3
Charging tank 1
Charging tank 2
Charging tank 3
Figure 3.10: Optimal schedule for COSP2 with late vessel arrivals and fixed initial decisions
(profit: $9,609,000).
solution strategy does not necessarily converge to a global optimum. This is explained by
the fact that the composition constraints are always satisfied, even if dropped, under specific
conditions. For each transfer v assigned to priority-slot i, the composition constraint (3.6c)
is satisfied if either of the following conditions holds true.
a) the origin tank contains only one crude before the transfer
b) the origin tank is fully discharged during the transfer

Table 3.2: MOS computational results for crude-oil scheduling problems.

Number Cumulative
Pb n LP MILP of nodes CPU CPU NLP Gap
1-4 infeas 3.08s 3.08s
COSP1 5 80.000 79.750 48 1.16s 4.24s 79.750 0%
6 80.000 no solution 0 1.56s 5.80s
1-3 infeas 2.22s 2.22s
4 103.000 90.000 0 1.05s 3.27s
COSP2 5 103.000 96.170 122 2.30s 5.57s
6 103.000 101.175 225 3.64s 9.21s 101.175 0%
7 103.000 no solution 335 8.61s 17.82s
1-2 infeas 1.43s 1.43s
3 84.905 82.500 0 0.85s 2.28s
COSP3 4 100.000 84.500 16 1.26s 3.54s
5 100.000 87.400 63 1.74s 5.28s 85.400 2.3%
6 100.000 no solution 300 3.57s 8.85s
1-3 infeas 2.59s 2.59s
COSP4 4 132.585 132.548 21 1.58s 4.17s 132.548 0%
5 132.585 no solution 0 1.72s 5.89s
These conditions are always met in the optimal solutions of all problems except COSP3.
They do not always hold true in the optimal solution of COSP3, which explains the positive
gap obtained.
3.5.3 Performance of the MOS-SST Model
The MILP relaxation of the MOS-SST model is solved with the additive approach using
the second stop criterion (∆ ≤ 0). The initial number of priority-slots is set to n0 = 1.
Computational results are given in Table 3.3. All instances are solved within 400 seconds
with slightly more priority-slots than for the MOS model. The most difficult instance is still
COSP2. It is interesting to note that the solution obtained for this instance is not actually
globally optimal as its objective value (101.174) is slightly lower than 101.175 (see Table 3.2).
The actual global optimal solution can be obtained with at least 10 priority-slots. It clearly
shows that the second stopping criterion does not guarantee global optimality, although it
is very efficient in practice.

Table 3.3: MOS-SST computational results for crude-oil scheduling problems.

Number Cumulative
Pb n LP MILP of nodes CPU CPU
1-4 infeas 2.97s 2.97s
5 80.000 79.722 144 1.25s 4.22s
COSP1
6 80.000 79.750 288 1.94s 6.16s
7 80.000 no solution 968 6.30s 12.46s
1-4 infeas 3.30s 3.30s
5 103.000 90.000 327 4.59s 7.89s
6 103.000 97.726 2316 33.51s 41.40s
COSP2
7 103.000 97.751 4672 70.26s 111.66s
8 103.000 101.174 4545 88.45s 200.11s
9 103.000 no solution 7007 185.29s 385.40s
1-3 infeas 2.52s 2.52s
4 100.000 82.500 48 1.43s 3.95s
5 100.000 84.500 333 3.42s 7.37s
COSP3
6 100.000 87.000 931 14.83s 22.20s
7 100.000 87.400 2101 38.02s 60.22s
8 100.000 no solution 3900 99.03s 159.25s
1-5 infeas 24.84s 24.84s
COSP4 6 132.585 132.548 1293 28.62s 53.46s
7 132.585 no solution 500 20.74s 74.20s
3.5.4 Performance of the MOS-FST Model
The MILP relaxation of the MOS-FST model is solved with the multiplicative approach
using the second stop criterion (∆ ≤ 0). The initial number of priority-slots is set to n0 = 4.
Computational results are given in Table 3.4. Only COSP1 is solved to global optimality
although feasible solutions are obtained quickly. Within the 1,000 second time limit, near-
optimal solutions are obtained for instance COSP3 and COSP4, but the solution obtained
for COSP3 shows a 4.8% gap with the best known solution of the MILP relaxation. The
MOS-FST discrete-time representation is not efficient at solving the crude-oil scheduling
problem. Indeed, as variable processing times are used, constraint (2.24) does not help
strengthening the model. Instead, if one wishes to use a discrete-time approach, it would
be preferable to consider the work of Lee et al. (1996). Indeed, in their formulation, each
operation that is executed over several consecutive time intervals is actually split into several

Table 3.4: MOS-FST computational results for crude-oil scheduling problems.

Number Cumulative
Pb n LP MILP of nodes CPU CPU
4 80.000 infeas 0 0.87s 0.87s
COSP1 8 80.000 79.750 74 1.56s 1.56s
16 80.000 no solution 100 15.17s 17.60s
4 98.000 infeas 0 0.90s 0.90s
COSP2 8 103.000 90.000 2114 28.76s 29.66s
16 103.000 96.250 11852 +1,000s +1,029.66s
4 84.929 82.500 0 0.93s 0.93s
8 99.120 84.250 22 2.55s 3.48s
COSP3
16 100.000 87.156 563 29.97s 33.45s
32 100.000 no solution 300 +1,000s +1,033.45s
4 132.585 infeas 0 1.20s 1.20s
COSP4 8 132.585 132.266 574 12.82s 14.02s
16 132.585 132.362 2769 376.84s 390.86s
32 132.585 no solution 0 +1,000s +1,390.86s
smaller operations, one for each time interval.
3.5.5 Performance of the MILP-NLP Decomposition Strategy
Table 3.5 shows the performance of different solution methods on all four problems using
the MOS time representation. MINLP solvers are given as a starting point the solution
of the LP obtained by removing the nonlinear constraints of the model and relaxing the
integrality constraints on binary variables. Furthermore, the MINLPs are solved with the
direct approach using the optimal number of priority-slots. The results show that, while
requiring lower CPU times than other solvers, the solution obtained with the two-step
procedure is optimal for all problems but COSP3 for which it is very close to the best
known solution. The best alternative is to use DICOPT as the iterative outer-approximation
algorithm behaves similarly to the MILP-NLP decomposition. Indeed, during the first
iteration, DICOPT solves an MILP and then solves the NLP obtained after fixing the
binary variables. The main difference comes from the fact that in the first MILP, nonlinear
constraints are linearized at the solution of the initial NLP relaxation instead of being

3.6 Conclusion
Table 3.5: Performance of different MINLP algorithms for crude-oil scheduling problems.
Solution method Problem Slots Solution CPU time Optimality gap

COSP1 1-6 79.750 6s 0%
COSP2 1-7 101.175 18s 0%
MILP-NLP decompositiona
COSP3 1-6 85.400 9s 2.3%
COSP4 1-5 132.548 6s 0%
COSP1 5 79.350 11s -
COSP2 6 101.175 9s -
DICOPTb
COSP3 5 85.449 8s -
COSP4 4 132.548 8s -
COSP1 5 79.350 77s -
COSP2 6 101.175 891s -
SBBb
COSP3 5 85.449 778s -
COSP4 4 132.548 154s -
COSP1 5 79.750 76s -
COSP2 6 101.175 25s -
AlphaECPb
COSP3 5 82.904 234s -
COSP4 4 132.548 171s -
COSP1 5 79.750 69s 0%
COSP2 6 100.896 +1,000s 2.0%
BARONb
COSP3 5 60.000 +1,000s 40.0%
COSP4 4 132.548 645s 0%
a
MILP solved using the additive procedure and NLP solves with optimal number of priority-slots
b
MINLP solved with optimal number of priority-slots
dropped. This explains why DICOPT does not find the optimal solution of COSP1. All
the MILPs solved in DICOPT, including the first one, are not relaxations of the non-convex
MINLP.
3.6 Conclusion
In this chapter, the crude-oil scheduling problem has been stated and the corresponding
MOS, MOS-SST, and MOS-FST mathematical formulations have been derived. The uni-
fied modeling approach developed in chapter 2 proved to be very effective as most of the
constraints used in all models are identical. These models have been strengthened using

3.6 Conclusion
the non-overlapping graph which displays non-trivial cliques. These cliques can be used
to generate tightening constraints that might have not been determined without using this
systematic approach.
As the models contain nonlinear composition constraints, a two-step MILP-NLP decom-
position procedure has been developed in order to solve the MINLP models. The number
of priority-slots and the sequence of operations is determined at the first MILP stage and
a feasible solution is obtained by solving the second-stage NLP. The approach has the ad-
vantage of providing optimal or near-optimal solutions with an estimate of the optimality
gap. It performs better than other solvers both in terms of computational performance and
optimality of the solution. DICOPT is the best alternative for solving the MINLP model
directly.

Chapter 4
Single-Operation Sequencing Model for
Crude-Oil Operations Scheduling
4.1 Introduction
In this chapter, we develop the SOS model for the crude-oil operations scheduling problems
COSP1, . . . , COSP4. It is based on the same sets, parameters and variables as the MOS,
MOS-SST, and MOS-FST models derived in chapter 3. The objective function (see sec-
tion 3.3.4) and general constraints (see section 3.3.5) are identical. We first derive strength-
ened assignment and non-overlapping constraints by selecting both cliques and bicliques
from the non-overlapping graph. Next, we introduce symmetry-breaking constraints for the
SOS model that are based on a sequencing rule expressed using a regular language and its
corresponding deterministic finite automaton. Various computational results demonstrat-
ing the impact of such symmetry-breaking techniques and of the number of priority-slots
are presented. Finally, a comparison of all mathematical models for the crude-oil schedul-
ing operations problem is provided. The full SOS mathematical model for the crude-oil
operations scheduling problem is given in appendix C.
4.2 Strengthened Constraints
As mentioned in chapter 2, both cliques and bicliques of GN O can be used to generate non-
overlapping constraints. Table 4.1 displays all maximal cliques and all maximal bicliques
that are not derived from cliques for problems COSP2 and COSP3 (see Fig. 3.5). They
are 15 maximal cliques and 15 maximal bicliques which corresponds to 15 + 2 · 15 = 45
Chapter 4. Single-Operation Sequencing Model for Crude-Oil Scheduling 92

Table 4.1: Maximal cliques and bicliques for COSP2 and COSP3.
Nb of vertices Maximal cliques
{1, 4}, {1, 5}, {2, 6},
{2, 7}, {2, 8}, {3, 9},
{3, 10}, {4, 11},
2
{6, 11}, {8, 14},
{10, 14}, {11, 12},
{13, 14}
{1, 2, 3}, {5, 12, 13},
3
{7, 12, 13}, {9, 12, 13}
Nb of vertices Maximal bicliques
({4}; {1, 4, 11}), ({6}; {2, 6, 11}),
3
({8}; {2, 8, 14}), ({10}; {3, 10, 14})
({5}; {1, 5, 12, 13}), ({7}; {2, 7, 12, 13}),
4 ({9}; {3, 9, 12, 13}), ({11}; {4, 6, 11, 12}),
({14}; {8, 10, 13, 14})
({1}; {1, 2, 3, 4, 5}), ({3}; {1, 2, 3, 9, 10}),
5
({5, 7, 9, 12, 13}; {12, 13})
({2}; {1, 2, 3, 6, 7, 8}),
6 ({12}; {5, 7, 9, 11, 12, 13}),
({13}; {5, 7, 9, 12, 13, 14})
constraints (4.1) and (4.2) which are identical to the constraints (2.19) and (2.26) developed
in chapter 2.
X X X
Ei1 v + Div
X X
≤ Si2 v + H · (1 − Z i2 v )
v∈W 0 v∈W 0
X X X i1 , i2 ∈ T, i1 < i2 ,
Ei1 v ≤ Si2 v + H · (1 − Z i2 v ) (4.2a)
X X X i1 , i2 ∈ T, i1 < i2 ,
Ei1 v ≤ Si2 v + H · (1 − Z i2 v ) (4.2b)

Table 4.2: Cliques and bicliques selections a, b, and c for COSP2 and COSP3.
Cst No. Selection a Selection b Selection c
(i) 1 {1, 2, 3} {1, 2, 3} implied by 2, 3, 4
2 {1, 4}, {1, 5} ({1}; {4, 5}) ({1}; {1, 2, 3, 4, 5})
3 {2, 6}, {2, 7}, {2, 8} ({2}; {6, 7, 8}) ({2}; {1, 2, 3, 6, 7, 8})
4 {3, 9}, {3, 10} ({3}; {9, 10}) ({3}; {1, 2, 3, 9, 10})
(ii)
5 {4, 11}, {6, 11} ({4, 6}; {11}) ({4, 6, 11, 12}; {11})
6 {5, 12, 13}, {7, 12, 13}, {9, 12, 13} ({5, 7, 9}; {12, 13}) ({5, 7, 9, 12, 13}; {12, 13})
7 {8, 14}, {10, 14} ({8, 10}; {14}) ({8, 10, 13, 14}; {14})
(iii) 8 implied by 6 {12, 13} implied by 6
9 {11, 12} {11, 12} implied by 5
(iv)
10 {13, 14} {13, 14} implied by 7
There is a clear trade-off between the tightness of the LP relaxation and the size of the
model. Therefore, it is important to carefully select the cliques and bicliques that are used to
enforce non-overlapping constraints. We introduce 3 different selection heuristics. The first
selection strategy, denoted by a, consists of selecting all maximal cliques only, leading to 15
constraints (4.1). The second selection strategy, denoted by b, consists of deriving cliques
and bicliques directly from the non-overlapping constraint definitions. The third selection
strategy, denoted by c, consists of improving selection b by extending it to maximal cliques
and bicliques and removing unnecessary elements from the selection. Table 4.2 displays
selections a, b and c. The first column indicates the descriptive constraint (as introduced
in section 3.2.1) that generates the corresponding cliques or bicliques. For example, non-
overlapping constraint (ii) for the first storage tank is intuitively represented by the clique
({1}; {4, 5}) in selection b but it is extended to the maximal biclique ({1}; {1, 2, 3, 4, 5})
in selection c. Also, non-overlapping constraint (i) is represented by the maximal clique
{1, 2, 3} in selection b, but is removed in selection c as it is implied by bicliques 2, 3, and
4. Therefore, selection b leads to 16 non-overlapping constraints (4 of type (4.1) and 12 of
type (4.2)) while selection c leads to 12 non-overlapping constraints (4.2) only.

4.3 Symmetry-Breaking Constraints
In section 3.3.7, a generic symmetry-breaking constraint was proposed in order to reduce

the search space of the MOS model. Here, a problem-specific approach is developed in order
to break the symmetries that arise in the SOS model. The methodology is based on the
discrete representation of an SOS solution as a sequence of operations.
4.3.1 Symmetric Sequences of Operations
It is possible to determine different sequences of operations leading to the same schedule.

For the optimal solution of COSP1 (Figure 3.3), one can show that there are 120 symmetric
sequences of operations, 4 of which are displayed in Figure 4.1. These sequences are obtained
from the optimal one by moving operations in the sequence in such a way that the same
precedence constraints are active. As there are many different optimal discrete solutions, the
branch & bound algorithm will explore many redundant nodes of the search tree. Therefore,
an efficient symmetry-breaking tool is needed to avoid searching irrelevant solutions.
4.3.2 A Sequencing Rule Based on a Regular Language
A sequencing rule is defined in order to select the sequences of operations to be explored.

This rule is expressed as a regular language, which can be recognized by a deterministic finite
automaton (DFA). A regular language is a set of words (sequences of letters) defined from
an alphabet (i.e. the set of operations), the empty word ε, and the operations concatenation
‘·‘ (symbol usually omitted), union ‘+‘, and Kleene star ‘*‘. Given two languages L1 and
L2 , these operations are defined by the following formulas.
L1 · L2 = {w = w1 · w2 s.t. w1 ∈ L1 and w2 ∈ L2 }
L1 + L2 = {w s.t. w ∈ L1 or w ∈ L2 }
L1 ∗ = {ε} ∪ L1 ∪ L1 · L1 ∪ L1 · L1 · L1 ∪ . . .

(a) Sequence 76-83513-762.

1 6
2 10
3 4 7
4
5 5
6 2 9
7 1 8
8 3
0 1 2 3 4 5 6 7 8
(b) Sequence 67-83513-762.
1 6
2 10
3 4 7
4
5 5
6 1 9
7 2 8
8 3
0 1 2 3 4 5 6 7 8
(c) Sequence 76-83135-762.
1 5
2 10
3 4 6
4
5 7
6 2 9
7 1 8
8 3
0 1 2 3 4 5 6 7 8
(d) Sequence 76-83513-627.
1 6
2 9
3 4 7
4
5 5
6 2 8
7 1 10
8 3
0 1 2 3 4 5 6 7 8
Figure 4.1: Symmetric sequences of operations for COSP1.
The reader may refer to Hopcroft and Ullman (1979) for a complete definition of regular
languages.

4.3.3 Rule Derivation for COSP1
The rule presented in this section has been designed in order to remove as many symmetric
sequences of operations as possible, regardless of its complexity. Notice that there is a
trade-off between the symmetry-breaking capabilities of the sequencing rule (number of
symmetric solutions it eliminates) and its complexity.
In COSP1, two distillation states exist as either distillation 7 or distillation 8 is being ex-
ecuted at any time. Thus, the sequence of operations can be decomposed into subsequences,
each corresponding to one distillation state, as shown in Figure 4.1.
Let L7 (resp. L8 ) be the regular language describing the possible sequences of operations
during distillation state 7 (resp. distillation state 8). Note that only transfer operations 1,
2, 4 and 6 are allowed to be executed during this state due to non-overlapping constraints.
If no unloading operation is performed, operations 4 and 6 need to be executed at most
once. Thus, in that case, we choose to define the regular language L7 so that a subsequence
corresponding to the distillation state 7 starts with distillation 7 and may follow by at most
one occurrence of transfer operations 4 and 6, in this order.
L7 = {7, 74, 76, 746} = 7(ε + 4)(ε + 6)
If unloading operation 1 is allowed to be executed during distillation state 7, then it can

be executed at most once. Also, it might be necessary to perform transfer operation 4 before
and after this unloading. Thus, in that case, we define the regular language L7 as follows.
L7 = 7(ε + 4)(ε + 6)(ε + 1 + 14)
If both unloading operations 1 and 2 are allowed to be executed during distillation state
7, then it might also be necessary to perform transfer operation 6 before and after this
unloading. Also, unloading 1 has to be sequenced before operation 2 due to the prece-
dence constraint between unloading operations (crude-oil vessels have to unload in order of
arrival). Thus, in this general case, we choose to define the regular language L7 as follows.
L7 = 7(ε + 4)(ε + 6)(ε + 1 + 14)(ε + 2 + 26)

Table 4.3: List of sequences belonging to regular language L7 .

Length Sequences belonging to L7
1 7
2 71, 72, 74, 76
3 712, 714, 726, 741, 742, 746, 761, 762
4 7126, 7142, 7412, 7414, 7426, 7461, 7462, 7612, 7614, 7626
5 71426, 74126, 74142, 74612, 74614, 74626, 76126, 76142
6 741426, 746126, 746142, 761426
7 7461426
1 1 4
6
2
7 6 2 2
1
4
2 2 6
Figure 4.2: Automaton DFA7 recognizing regular language L7 .
The set of all sequences of operations belonging to the regular language L7 is displayed
in Table 4.3. It can be represented by the DFA depicted in Figure 4.2, noted DFA7 . This
DFA reads a sequence starting with operation 7, and then reads the following operation in
the sequence by moving though the corresponding labeled arc. A sequence is accepted if it
can be entirely read by DFA7 .
Similarly, the regular language L8 , represented by DFA8 , is defined as follows.
L8 = 8(ε + 3)(ε + 5)(ε + 1 + 13)(ε + 2 + 25)
Finally, the regular language L describing an efficient sequencing rule for COSP1 can be
defined using language L7 and L8 . Figure 4.3 shows a scheme of the DFA recognizing the
regular language L.
L = (ε + L7 )(L8 · L7 )∗ (ε + L8 )

DFA7
7
78
8
DFA8
Figure 4.3: Automaton recognizing the regular language L.
It should be noted that this symmetry-breaking rule captures all possible schedules and
removes many redundant sequences of operations. In Figure 4.1, schedule (a) is accepted
by the sequencing rule while the other schedules are rejected, and thus not explored during
the search. However, there are some symmetric sequences that remain accepted such as
78132 and 71832. Indeed, in these two sequences belonging to the language L, exchanging
operations 1 and 8 does not change the active precedence constraints.
4.3.4 Regular Constraint
Once the regular language L and its corresponding DFA have been defined, it is possible
to include the sequencing rule using the linear system of equations proposed by Côté et al.
(2007). The equations represent a unit network flow through a directed layered graph
initially introduced by Pesant (2004) for constraint programming.
Let M = (Q, Σ, δ, q1 , F ), where Q is the set of states (i.e. nodes), Σ is the alphabet,
δ is the transition function, q1 is the initial state, and F is the set of final states, be a
DFA recognizing the regular language L. The idea is to unfold the automaton states on
n + 1 layers (where n is the length of the sequence), the first layer corresponding to the
initial state, the last layer corresponding to the possible final states, and the transition
between layers corresponding to the automaton transitions defined by δ. Then, a sequence
is recognized by the DFA M if and only if there is a flow of unit 1 from the initial state in
the first layer to a final state in the last layer such that the transition taken between any
two adjacent layers i and i + 1 corresponds to the ith letter of the sequence.
The regular language membership constraint makes use of flow binary variables Sivq ,

where q ∈ Q. Sivq = 1 if the automaton is in state q when operation v is assigned to

priority-slot i, Sivq = 0 otherwise. The network flow problem is defined with the following
constraints.
X
Sivq = Ziv i ∈ T, v ∈ W linking constraint (4.3a)
q
X
S1vq1 = 1 initial unit flow (4.3b)
v
X X
S(i−1)vq0 − Sivq = 0 i ∈ T, i 6= 1, q ∈ Q flow conservation (4.3c)
v,q 0 ,q=δ(q 0 ,v) v
X
Snvq = 1 final unit flow (4.3d)
v,q,δ(q,v)∈F
Constraint (4.3a) links the Ziv variable to the Sivq variables (q ∈ Q) ensuring that the
assignment of operation v to priority-slot i is equivalent to a unit flow traversing one of
the corresponding arcs in the network. Constraint (4.3b) sets the flow leaving the network
source node (1, q1 ) to 1. Constraint (4.3c) ensures that the flow entering network node (i, q)
is equal to the flow leaving it. Constraint (4.3d) sets the flow leaving the network nodes
(n, q) through transitions δ(q, v) ∈ F to 1.
Although O(|T | · |W | · |Q|) new variables and O(|T | · |W | + |T | · |Q|) new constraints are
added to the model, the search space is substantially reduced as only sequences belonging
to the regular language L are explored. It should be noted that it is not necessary to
declare the variable Sivq as binary. Indeed, if all variables Ziv are fixed, then the sequence
of operations is fixed to v1 ...vi ...vn . As the automaton M is deterministic, there is a unique
sequence of states q1 ...qi ...qn qn+1 (where qn+1 ∈ F ) corresponding to the states visited upon
processing the sequence v1 ...vi ...vn . Therefore, the network flow problem stated above has
a unique solution defined by Sivq = 1 if v = vi and q = qi , Sivq = 0 otherwise.
In this section, we present experimental results obtained on an Intel Xeon 1.86GHz processor
with GAMS as modeling language, CPLEX 11 as MILP solver and CONOPT 3 as local

Table 4.4: SOS computational results for crude-oil operations scheduling problems.
Pb n LP MILP Nb of nodes CPU
COSP1a 13 80.000 79.750 18 5.88s
COSP1b 13 80.000 79.750 19 4.45s
COSP1c 13 80.000 79.750 21 4.92s
COSP2a 21 103.000 101.175 36 120.42s
COSP2b 21 103.000 101.175 19 55.57s
COSP2c 21 103.000 101.175 25 60.50s
COSP3a 21 100.000 87.400 28 191.47s
COSP3b 21 100.000 87.400 33 97.70s
COSP3c 21 100.000 87.400 31 64.46s
COSP4a 26 132.585 132.548 16 606.86s
COSP4b 26 132.585 132.548 37 574.95s
COSP4c 26 132.585 132.548 32 308.43s
NLP solver. The DFAs for problems COSP2, COSP3, and COSP4 have been defined in a
similar way as for COSP1, adapted to each refinery configuration.
4.4.1 Performance of the SOS Model
The MILP relaxation (defined in section 3.5.5) of the SOS model is solved using the direct
approach (see section 2.6). The number of priority-slots is set to its maximum value as deter-
mined in section 4.4.2. Computational results are given in Table 4.4 for each clique/biclique
selection heuristic. The behavior of this time representation is quite different than for the
MOS and MOS-SST representations as COSP4 is now the most difficult while COSP2 is
quite easy to solve. Indeed, COSP4 is solved with the largest number of priority-slots as
it is the largest instance in terms of operations and resources, which makes it the hardest
instance. The results obtained for selection a shows that using only maximal cliques is not
the most efficient in the SOS model. It is rather preferable to combine it with bicliques as
in selections b and c. Additionally, it is clear that the selection improvements in selection c
lead to a significant decrease in CPU time for COSP3 and COSP4. This is due to the fact
that non-overlapping constraints are strengthened and the model size is reduced.

4.4.2 Effect of the Number of Priority-Slots
It is critical to postulate an appropriate number of priority-slots to be used in the model.

Indeed, a small number of priority-slots may lead to infeasibility, while a large number of
priority-slots may result in an unsolvable model. The strategy used to determine the number
of priority-slots is dependent on the problem. We introduce the concepts of minimum
sequence and minimum schedule.
Definition 1 (minimum sequence). We denote N Ov the set of operations that must
not overlap with operation v ∈ W . A minimum sequence s = s1 s2 . . . sn is a sequence of
operations that satisfies the following property:
∀i, j ∈ T, i < j, ∀v ∈ W, (si = sj = v) ⇒ (∃k ∈ T, i < k < j, ∃w ∈ N Ov \ {v}, sk = w)
Definition 2 (minimum schedule). A minimum schedule is a feasible schedule that can

be generated by a minimum sequence s.
In other words, a minimum schedule is a schedule such that between any two executions
of an operation v, another non-overlapping operation w ∈ N Ov \ {v} is performed between
these two executions. By merging any two executions of an operation v which do not
respect the previous property, it can be shown that any optimal schedule for the crude-oil
scheduling problem can be transformed into a minimum schedule. Therefore, it is sufficient
to explore the space of minimum schedule. The symmetry-breaking rule presented earlier
only accepts minimum sequences.
In the case of the refinery problems introduced by Lee et al. (1996), it can be shown
that the number of operations in a minimum schedule is bounded as long as the number of
distillation operations is bounded. We use the terminology maximum number of operations
to represent the maximum length of a minimum sequence. For example, if COSP1 is
constrained to be solved with a maximum of 3 distillation operations, the maximum number
of operations is 13 as in the sequence 7461483525746. Any sequence of 14 operations is
not minimum, and thus will be rejected by the sequencing rule in combination with the
distillation cardinality constraints. Therefore, it is unnecessary to postulate a number of

priority-slots greater than 13, whether the symmetry-breaking rule is used or not. The
maximum number of operations is 21 for COSP2 and COSP3 (with at most 5 distillation
operations), 26 for COSP4 (with at most 7 distillation operations).
Figure 4.4 shows the evolution of the number of nodes explored, the CPU time, the MILP
and NLP solutions objective value ($100,000 unit), and the optimality gap with respect to
the number of priority-slots when solving COSP1 to COSP4. Xpress was used to solve
the MILP relaxation. The grey area represents the gap between MILP and NLP solutions.
For all problems, the computational expense is small when the number of priority-slots is
small, as the size of the problem and the feasible space are small. Also, the computational
expense is small when the number of priority-slots is large (close to its maximum), as the
solver must assign one operation to each priority-slot while satisfying both the cardinality
constraints and the symmetry-breaking rule, thus reducing the size of the feasible space. In
between, the number of nodes explored and the CPU time reach a maximum, although not
at the same number of priority-slots. It can also be observed that the objective value of
the optimal solutions of the MILP relaxation increases with the number of priority-slots as
more flexibility is given to the solver to find feasible solution. The optimality gap tends to
decrease.
4.4.3 Remark on the Optimality of the Solution
When no symmetry-breaking rule is used, the global optimal solution of the SOS model
with maximum number of priority-slots leads to the best possible schedule. Indeed, the
optimal sequence of operations can be extended with unused operations, for which the vol-
ume transferred is set to 0. Thus, it can be obtained even if the user postulated more
priority-slots than needed. However, in general, if a sequencing rule is used, the global
optimal solution might not be obtained with maximum number of priority-slots, but it will
be obtained with a number of slots identical to the number of operations actually performed
in this solution. For example, consider the sequence 71287 and assume it is optimal. In
this case, unloading operations 1 and 2 may each overlap with two consecutive distillations.

COSP1 COSP2
200 4 1200 80
No. of Nodes
1000
Number of Nodes
Number of Nodes
150 3 60
CPU Time (s)
CPU Time (s)

800
100 2 600 40
CPU Time 400

50 1 20
200
0 0 0 0
6 7 8 9 10 11 12 13 9 10 11 12 13 14 15 16 17 18 19 20 21
Number of Slots Number of Slots
80 5 5
MILP 102
4 4
Optimality Gap (%)
Optimality Gap (%)

79.5
100
3 3
Solution
Solution
79 NLP
2 98 2
78.5 1 1
Gap 96
0 0
78 94
6 7 8 9 10 11 12 13 9 10 11 12 13 14 15 16 17 18 19 20 21
COSP3 COSP4
5000 400 2500 300
250
4000 2000
Number of Nodes
Number of Nodes
300
CPU Time (s)
CPU Time (s)

200
3000 1500
200 150
2000 1000
100
100
1000 500
50
0 0 0 0
9 10 11 12 13 14 15 16 17 18 19 20 21 13 14 15 16 17 18 19 20 21 22 23 24 25 26
100 5 5
95 4 4
Optimality Gap (%)
Optimality Gap (%)
3 132 3
Solution
Solution
90
2 2
85
1 1
80
0 0
75 131
9 10 11 12 13 14 15 16 17 18 19 20 21 13 14 15 16 17 18 19 20 21 22 23 24 25 26
Figure 4.4: Performance of the SOS model on crude-oil scheduling problems (MILP solver:
Xpress).

However, if the model is solved with 13 slots using the sequencing rule, unloading opera-
tions cannot overlap with two consecutive distillations. Indeed, if unloading 1 is sequenced
after a distillation 7 and before a distillation 8, transfer 4 will necessarily be sequenced
before and after unloading 1, within the current subsequence for distillation state 7 (i.e.
74614 or 7461426). Therefore, due to precedence constraints, unloading 1 is forced to be
executed entirely during distillation 7, and cannot overlap with next or previous distilla-
tions. Nonetheless, in our experiments, the best solutions were always obtained with the
maximum number of priority-slots which tends to justify this choice for the parameter n.
4.4.4 Effect of Symmetry-Breaking Constraints
In this section, we study the effect of the symmetry-breaking rule on solving the MILP
relaxation of the SOS model, which is the first step of the MILP-NLP decomposition pro-
cedure presented in section 3.5.5. Two models will be compared: the basic model that
includes all linear constraints of the MILP except the symmetry-breaking constraints, and
the extended model that includes all linear constraints as well as the symmetry-breaking
constraints.
Consider a single instance of COSP1 (Figure B.1, Table B.1) for which 13 priority-slots
are postulated. As shown in Table 4.5, although the extended model contains many more
new variables and relatively fewer additional constraints than the basic model, the number
of nodes explored is greatly reduced (from more than one million to 63), which leads to very
low CPU time (from more than 3600s to 2s). This is due to the removal of many symmetric
solutions from the search space by using the symmetry-breaking rule. It should be noted
that both models have the same LP relaxation. The optimal solution of the basic model is
found early during the search, but a large amount of time is used to prove its optimality.
Figure 4.5 shows how both models perform on COSP1 when varying the number of
priority-slots from 6 to 13. Clearly, the computational expense needed to solve the basic
model grows exponentially with the number of priority-slots. However, both the number of
nodes and the CPU time remain stable when solving the extended model.

4.5 Comparison of Crude-Oil Scheduling Models
Table 4.5: Size and performance of Basic and Extended models on COSP1 (13 slots).
Binary CPU
Variables Variables Constraints MILP Nodes Time
Basic 1,392 104 2,515 79.75 +180,000 +3,600s
Extended 2,801 104 2,952 79.75 21 5s
1x107 No. of nodes (Extended)

10000
CPU time (Extended)
1x106 No. of nodes (Basic)
CPU time (Basic) 1000
100000
Number of nodes
CPU Time (s)

10000 100
1000
100 10
10
1
1
0.1 0.1
6 7 8 9 10 11 12 13
Number of time-slots
Figure 4.5: Performance of the Basic and Extended models on COSP1 (6 to 13 slots).
In this section, we compare the four time representations applied to the crude-oil operations
scheduling problem in chapter 3 and 4. Figure 4.6 shows how the objective value of the best
incumbent varies over time depending on the time representation used. Results for the SOS
model are presented for clique/biclique selection heuristic c. As for the batch scheduling
problems the MOS model proves to be superior. However, the MOS-SST model behaves
better than the MOS-FST model for these problems due to non-constant processing times.
The SOS model performs better than the MOS-SST model for instance COSP2, worse for
instance COSP4, and similarly for the other instances. From these results, it seems that
the MOS-SST model scales better with the size of the problem than the SOS model as it
requires fewer priority-slots. Finally, in Table 4.6 we present the model sizes of the MILPs

COSP1 COSP2
79.8 105
100
79.75
Objective Value
Objective Value
95
79.7
90
MOS MOS
79.65
MOS-SST MOS-SST
85
MOS-FST MOS-FST
SOS SOS
Optimum Optimum
79.6 80
1 10 100 1 10 100 1000
COSP3 COSP4
88
134
86
132
Objective Value
Objective Value
84 130
128
MOS MOS
82
MOS-SST MOS-SST
MOS-FST MOS-FST
SOS SOS
126
Optimum Optimum
80
1 10 100 1000 1 10 100 1000
Figure 4.6: Comparison of time representations for crude-oil scheduling problems.
corresponding to all time representations using the number of priority-slots leading to the
best solution. It is clear that the MOS time representation leads to the most compact
models which partly explains its efficiency. The MOS-FST time representation is much
larger due to the higher number of priority-slots required. Furthermore, it is interesting
to note that the SOS time representation requires many more continuous variables. These
variables are introduced to represent the network flow used to break symmetries as described
in section 4.3. Overall, it is also interesting to note that all four models provide the same
LP relaxation for each instance. Indeed, the objective function is purely economic and not
directly linked to scheduling issues, as opposed to the previous batch scheduling examples.
In the case of crude-oil operations scheduling, solving the LP relaxation corresponds to
generating a solution that only satisfies overall material balances, distillation demands and

4.6 Conclusion
Table 4.6: Size of MOS, MOS-SST, MOS-FST, and SOS models for crude-oil scheduling
problems.
Problem Model n Binary Vars Continuous Vars Constraints
MOS 5 40 496 1,086
MOS-SST 6 48 601 1,403
COSP1
MOS-FST 8 64 793 1,804
SOS 13 104 2,801 2,952
MOS 6 84 1,303 2,620
MOS-SST 8 112 1,745 3,758
COSP2
MOS-FST 16 224 3,473 8,051
SOS 21 294 13,084 7,835
MOS 5 70 1,211 2,302
MOS-SST 7 98 1,702 3,431
COSP3
MOS-FST 16 224 3,873 8,439
SOS 21 294 13,609 7,909
MOS 4 76 1,489 2,806
MOS-SST 6 114 2,239 4,338
COSP4
MOS-FST 16 304 5,953 12,269
SOS 26 494 28,445 14,351
specifications, and capacity limits at the end of the scheduling horizon.
4.6 Conclusion
In this chapter, the SOS time representation has been applied to crude-oil scheduling prob-
lems with the objective of maximizing gross margins. The corresponding mathematical
model is strengthened using the non-overlapping graph of the scheduling system. It is
possible to represent a solution schedule as a single sequence of operations that can be re-
stricted with respect to a problem-specific DFA-based symmetry-breaking sequencing rule.
Computational results show that using symmetry-breaking constraints is critical in order
to solve the SOS model in reasonable times. Similarly to the batch scheduling problems
studied in chapter 2, the MOS time representation is the most efficient approach. The SOS
model compares well with the MOS-SST model, while the discrete-time MOS-FST time
representation is clearly not suitable as the problem involves variable processing times.

Chapter 5
Tightening the Linear Relaxation of a
Crude-Oil Operations Scheduling MINLP
Using Constraint Programming
5.1 Introduction
In this chapter, we consider the optimization of the crude-oil operations scheduling problem
with the objective of minimizing the total logistics costs, as studied in Lee et al. (1996).
Logistics costs include sea waiting and unloading costs for marine vessels, storage costs
in tanks, and CDU switching costs. In earlier work, an operation specific event point
continuous-time formulation has been developed (see Jia et al., 2003) and applied to the
problems from Lee et al. (1996) (COSP1, . . . , COSP4) using a linear approximation of stor-
age costs. However, no proof of optimality or estimate of optimality gap is given with this
approximation. The global optimization of this model was later addressed by Karuppiah
et al. (2008) using an outer-approximation algorithm where the MILP master problem is
solved by a Lagrangian decomposition. While rigorous, this method is still computationally
expensive.
In chapter 3, we presented a solution procedure for solving non-convex crude-oil schedul-
ing MINLPs that returns a suboptimal solution with an estimate of the gap with respect
to the optimal solution. The main contribution of this chapter is to reduce the optimality
gap by tightening the linear relaxation of the MINLP using McCormick convex envelopes
for bilinear products of continuous variables (see McCormick, 1976). This can be done by
using discretization techniques and applying McCormick convex and concave estimators on
Chapter 5. Tightening the Linear Relaxation of an MINLP using CP 109

5.2 MINLP Model
each discrete element of the continuous space (see Bergamini et al., 2008; Wicaksono and
Karimi, 2008; Pham et al., 2009). On the other hand, it has been shown how Constraint
Programming (CP) techniques can be effectively integrated with branch & bound proce-
dures for the global optimization of MINLPs (see Sahinidis, 2003). In this chapter, CP is
used during the search to tighten variable bounds and generate new McCormick cuts, thus
effectively tightening the mixed-integer linear program (MILP) relaxation. When applied to
the sequential MILP-NLP approach, this extension of the branch & cut algorithm leads to
reduced optimality gaps and allows finding better suboptimal solutions with lower logistics
cost. The concepts developed in this work can be applied to a generic solver as in the global
MINLP solver BARON, or can be used to effectively solve other optimization problems to
global optimality by exploiting their specific structure.
This chapter is organized as follows. First we present the MINLP mathematical formula-
tion with the nonlinear objective function. The model is then reformulated in order to get
a linear objective function and obtain a simple MILP relaxation using McCormick convex
envelopes. Next, this MILP relaxation is improved by adding tighter McCormick cuts for
some subproblems explored during the branch & cut procedure. Finally, we provide com-
putational results showing the impact of this approach in terms of relaxation tightness and
CPU time.
5.2 MINLP Model
In this section, the SOS time representation is used to derive the MINLP model for the
crude-oil scheduling problem. It differs from the SOS model presented in appendix C by a
different objective function and an additional constraint that defines the sea waiting time
for each crude vessel.

5.2 MINLP Model
The objective function minimizing total logistics costs can be expressed as follows:
 
X X
min SW IT CHIN GCOST  Ziv − 1
r∈RD i,v∈Ir
X X
+ U N LOADIN GCOST Div
i∈T v∈WU
X X
+ SEAW AIT IN GCOST Wiv
i∈T v∈WU
+ F IXEDST ORAGECOST

X X Div
+ (SCr2 − SCr1 ) H − Siv − · Viv
2
i∈T r1 ,r2 ∈R,v∈Or1 ∩Ir2
where:
• Wiv is the variable waiting time before unloading v if it is assigned to slot i, Wiv = 0
otherwise
• SW IT CHIN GCOST is the cost associated with each CDU switch
• U N LOADIN GCOST is the unloading cost rate
• SEAW AIT IN GCOST is the sea waiting cost rate
• F IXEDST ORAGECOST is the total cost for storing the initial refinery inventory
during the horizon H if no crude transfer is performed
• SCr is the storage cost rate in resource r
The last term, which involves bilinearities making the model nonconvex, evaluates the vari-
ation of the total storage cost. For each transfer operation v between resources r1 (storage
cost rate SC1 ) and r2 (storage cost rate SC2 ) assigned to priority-slot i, the corresponding
variation of storage cost is calculated as follows:
Z t=Siv +Div Z t=H
Viv
∆COSTiv = (SCr2 − SCr1 ) · (t − Siv )dt + Viv dt
t=Siv Div t=Siv +Div

1
= (SCr2 − SCr1 ) Div · Viv + (H − Siv − Div ) · Viv
2

Div
= (SCr2 − SCr1 ) H − Siv − · Viv
2

5.2 MINLP Model
The variable Wiv is defined by the following constraint:
Wiv = Siv − Sv · Ziv i ∈ T, v ∈ WU (5.1)
The full model is written as follows:

 
X X
r∈RD i,v∈Ir
X X
+ U N LOADIN GCOST Div
i∈T v∈WU
X X
+ SEAW AIT IN GCOST Wiv
i∈T v∈WU

X X Div
+ (SCr2 − SCr1 ) H − Siv − · Viv
2
i∈T r1 ,r2 ∈R,v∈Or1 ∩Ir2
s.t. Wiv = Siv − Sv · Ziv i ∈ T, v ∈ WU
Constraints (3.1)-(3.6), (3.8)-(3.10), (4.1)-(4.3), and (2.27)
Table 5.1 displays the corresponding cost data for problems COSP1, . . . , COSP4. The fixed
t
P
storage cost is calculated as r SCr · L0r . The additional parameters are given in ap-
pendix B. Note that constraint (3.6) contains the nonlinear composition constraints dis-
cussed in chapter 3. The model is intended to be solved with the MILP-NLP decomposition
procedure introduced in section 3.5.5. In order to obtain near-optimal solutions, it is critical
to derive a tight linear relaxation of the full MINLP. In particular, the objective function
needs to be reformulated as explained in the following section. The nonlinear constraints
(3.6c) are simply dropped in the MILP, as justified by the results obtain in chapter 3 in
terms of optimality gap. However, these constraints are included in the second stage NLP.

5.3 Reformulation and Linear Relaxation
Table 5.1: Cost data for crude-oil operations scheduling problems.

COSP1 COSP2 COSP3 COSP4
Switching cost 50 50 50 30
Unloading cost 8 8 10 7
Sea waiting cost 5 5 5 5
Storage cost in storage tanks 0.05 0.05 0.04 0.05
Storage cost in charging tanks 0.08 0.08 0.08 0.06
Fixed storage cost 104 158 134.4 273
5.3 Reformulation and Linear Relaxation
It is common to reformulate MINLP problems by introducing new variables to represent

nonlinear terms. The MINLP model can then be rewritten as follows:
 
X X
r∈RD i,v∈Ir
XX
+ U N LOADIN GCOST v ∈ WU Div
i∈T
XX
+ SEAW AIT IN GCOST v ∈ WU Wiv
i∈T
X
+ (SCr2 − SCr1 ) Xiv
i,r1 ,r2 ,v∈Or1 ∩Ir2
s.t. Xiv = Aiv Viv i ∈ T, w ∈ W

Div
Aiv = H − Siv − i ∈ T, w ∈ W
2
Wiv = Siv − Sv · Ziv i ∈ T, v ∈ WU
Constraints (3.1)-(3.6), (3.8), (3.10), (4.1)-(4.3), and (2.27)
where variables Xiv are used to represent the bilinear terms Aiv · Viv . Using McCormick
convex envelopes introduced in McCormick (1976), this MINLP can be linearly relaxed by

5.4 McCormick Cuts
replacing the constraint Xiv = Aiv · Viv by:
Xiv ≥ AL L L L
iv Viv + Aiv Viv − Aiv Viv
Xiv ≥ AU U U U
Xiv ≤ AL U L U
Xiv ≤ AU L U L
The terms AL U L U
iv , Aiv , Viv , and Viv represent the lower and upper bounds of variables Aiv and
Viv . It is important to note that the tightness of this linear relaxation strongly depends on
the bounds of the variables Aiv and Viv .
5.4 McCormick Cuts
Once the linear relaxation of the MINLP has been defined, the corresponding MILP can be
solved by a branch & cut algorithm as implemented in the commercial tool CPLEX. This
approach consists of successively solving many subproblems (nodes) of the original MILP
by solving its continuous relaxation and generating new subproblems when needed. User-
defined constraints can also be added to each subproblem during the search using callbacks
(see ILOG Inc., 2007).
A subproblem of the MILP is obtained by fixing some of the discrete variables Ziv to
either 0 or 1. Each of these subproblems correspond to a linear relaxation of the MINLP
subproblem obtained by fixing the same discrete variables to the same values. If the LP
relaxation of an MILP subproblem is integer feasible (i.e. all discrete variables take an
integer value, whether it is fixed or not), it might still not satisfy all MINLP constraints. In
such cases, it is possible to infer stronger McCormick constraints by contracting variables
bounds for the current discrete solution.
We denote p a discrete solution defined by the discrete values taken by the variables Ziv .
The discrete solution p corresponds to a unique sequence of operations (v1p , . . . , vnp ). We

5.4 McCormick Cuts
denote (CP )p the following CP model:




 MINLP constraints


(CP )p Ziv = 1 ∀(i, v), v = vip



 Z = 0 ∀(i, v), v 6= v p

iv i
When a discrete solution p is obtained at a given node of the search tree, ILOG CP is used
p
to perform constraint propagation on the model (CP )p leading to variable bounds AL iv ,
p L p U p that are possibly tighter than the bounds defined at the root
AU

iv , Viv , and Viv
node (modeling stage). The McCormick constraints derived from these bounds are valid for
the discrete solution p but are not valid for the current node as some variables Ziv might
not have been fixed yet. The NLP subproblem corresponding to the discrete solution p is
in fact a restriction of the NLP subproblem corresponding to the current node. Therefore,
the following modified big-M McCormick constraints are defined to be added to the current
node subproblem:
p p p p X
Xiv ≥ AL
iv Viv + Aiv VivL − AL
iv VivL − M1 · (n − Zivip )
i
U p U p U p U p
X
Xiv ≥ Aiv Viv + Aiv Viv − Aiv Viv − M2 · (n − Zivip )
i
p p p p X
Xiv ≤ AL
iv Viv + Aiv VivU − AL
iv VivU + M3 · (n − Zivip )
i
p p p p X
Xiv ≤ AU
iv Viv + Aiv VivL − AU
iv VivL + M4 · (n − Zivip )
i
where:
p p p p
M1 = AL
iv VivU + AU L
iv Viv − AL
iv VivL
p U p
p U p
M 2 = AU VivU + AU − AU

iv iv Viv iv Viv
n o
L p L U p L p U p
M 3 = AU U L

V
iv iv − A iv Viv + A iv V iv − A iv V iv
n o
U p L L p U p L p
M 4 = AU U L

V
iv iv − A iv Viv + A iv V iv − A iv Viv
A lazy constraint callback (see ILOG Inc., 2007) has been implemented in order to gen-
erate these local McCormick cuts at each node where an integer feasible solution is found.

Select next node
Solve LP relaxation
Infeasible ?
Prune node
Suboptimal ? YES
NO
Integer Generate
feasible ? YES McCormick cuts
NO
Generate new
Cuts added ?
nodes (branch) YES
NO: accept current

solution
Figure 5.1: Branch & cut algorithm with McCormick cuts.
If at least one McCormick cut is violated by the current discrete solution, it is added to the
node subproblem and the LP relaxation is resolved as depicted in Fig. 5.1. This cut gener-
ation procedure can also be executed at nodes where no integer feasible solution is found.
However, in order to reduce the computational expense corresponding to the generation of
variable bounds and cuts, only integer feasible nodes are processed.
The four problems introduced in Lee et al. (1996) have been solved using two algorithms.
The BasicRelaxation algorithm consists of initially adding McCormick constraints to the
MILP model without generating new cuts during the search. The ExtendedRelaxation al-
gorithm consists of adding McCormick constraints to the MILP and new cuts during the
search, as explained in the previous section. Both approaches have been developed in

C++ using CPLEX 11 (MILP) and ILOG CP 6.5 (CP). The NLPs have been solved using
CONOPT 3. Experiments have been run on an Intel Core 2 Duo 2.16GHz processor.
In both basic and extended cases, Table 5.2 gives the solution of the MILP and the NLP,
the corresponding optimality gap, and the total CPU time and the number of nodes explored
during the search. In the ExtendedRelaxation case, the computational time for generating
McCormick cuts using CP is also displayed in column ”CP”, it is already included in the
total CPU time. The computational time for solving NLPs is not reported as it is always
lower than 5s.
The results show that using the approach developed in this paper leads to important
reductions of the optimality gap compared to the BasicRelaxation algorithm (3.48% vs
14.83% average gap). Besides, a better feasible solution (3.3% cost reduction) has been
found for COSP2 using the two-step MILP-NLP procedure. This demonstrates how the
linear relaxation of the MINLP can been tightened by adding McCormick constraints to
the subproblems of the MILP leading to integer feasible solutions. More precisely, at the
node where the optimal MILP solution of COSP1 is found, two rounds of cuts are added.
The first round of cuts leads to an increase of the objective value from 199.1 to 212.4 (6.7%
increase), while the second round increases the objective value to 213.7 (0.6% increase).
This indicates that few rounds of cuts are necessary in this case and that the first round of
cuts is the most important in order to tighten the MILP relaxation.
The optimality gaps obtained in the BasicRelaxation case are much larger than the opti-
mality gaps obtained in chapters 3 and 4. This is due to the fact that the objective function
considered in this previous work is linear, thus leading to a tighter MILP relaxation.
In terms of computational expense, the required CPU time increases by 9.5% for the
ExtendedRelaxation procedure. This is due to the increase of the number of nodes explored
in some cases (COSP2 and COSP3), the increase of the model size for subproblems for which
McCormick constraints have been generated, and the time used for performing constraint
propagation on the CP model (4.7% of total CPU time). This last point can be improved
with a better CP model and a more efficient implementation.

5.6 Conclusion
Table 5.2: Results obtained with BasicRelaxation and ExtendedRelaxation algorithms.

BasicRelaxation
MILP
Pb NLP Gap
Solution CPU Nodes
COSP1 199.1 9s 22 222.3 11.7%
COSP2 297.8 215s 55 362.9 21.9%
COSP3 254.6 224s 73 287.6 13.0%
COSP4 331.8 600s 20 374.0 12.7%
ExtendedRelaxation
MILP
Pb NLP Gap
Solution CPU CP Nodes
COSP1 213.7 11s 1s 20 222.3 4.0%
COSP2 343.1 246s 19s 57 351.2 2.4%
COSP3 269.2 337s 16s 95 287.6 6.8%
COSP4 371.3 554s 18s 19 374.0 0.7%
Table 5.3 shows computational results obtained with different MINLP algorithms on
COSP1. Local optimizers DICOPT, SBB, Bonmin and AlphaECP have been tested as well
as the global solver BARON. DICOPT and SBB did not return the best known solution,
SBB failed to find any feasible solution. However, the corresponding CPU times are similar
to the BasicRelaxation and ExtendedRelaxation procedures. Bonmin-OA found the best-
known solution in reasonable time. AlphaECP and BARON also found the best known
solution, but the former requires one order of magnitude increase in CPU time and does
not give any optimality gap estimate, while the latter requires more than two orders of
magnitude increase in CPU time (optimization is stopped after 1 hour) but has the smallest
optimality gap.
5.6 Conclusion
In this chapter, a new approach for handling bilinear terms in MINLPs has been presented.
It involves using CP bound contraction techniques during the branch & cut search in order
to generate cuts based on McCormick convex envelopes for products of continuous variables.
The procedure has the advantage of using the complementary strengths for CPLEX and

5.6 Conclusion
Table 5.3: Results obtained with diferrent MINLP algorithms on COSP1 and COSP2.
COSP1 COSP2
Algorithm Solution CPU Gap Solution CPU Gap
BasicRelaxation 222.3 9s 11.7% 362.9 215s 21.9%
ExtendedRelaxation 222.3 11s 4.0% 351.2 246s 2.4%
DICOPT 233.5 14s - 351.2 1235s -
SBB Local infeas. 14s - Local infeas. 697s -
Bonmin-OA 222.3 27s - No solution +3,600s -
AlphaECP 222.3 260s - 358.0 +3,600s -
BARON 222.3 +3,600s 4.1% No solution +3,600s -
ILOG CP to obtain better solutions and reduce the optimality gap. Due to the tight
integration of both tools using the modeling technology ILOG CONCERT, the increase in
solution times remains small. Although the proposed approach has been applied to a crude-
oil scheduling SOS MINLP model, it can also be applied to other time representations, such
as MOS, with no major modifications.
This approach may be improved by extending the interaction between MILP and CP as
in the programming system SCIP (see Achterberg, 2004) or in the integrated solver SIMPL
(see Yunes et al., 2010), although a conceptual difference lies in the fact that CPLEX
is used as the main mixed-integer branch & cut algorithm. Also, MINLP cuts such as
McCormick cuts can be generated not only for integer feasible subproblems, but for other
subproblems, thus pruning additional nodes during the search. Finally, optimality-based
reduction techniques (see Sahinidis, 2003) can be used to add new constraints to the CP
model to remove feasible solutions that are not optimal. As a consequence variable bounds
may be further contracted leading to a tighter MILP relaxation.

Chapter 6
Integration of Refinery Planning and
Crude-Oil Scheduling using Lagrangian
Decomposition
6.1 Introduction
Although, integration of planning and scheduling has recently been addressed in the context
of multiproduct continuous and batch production plants (see Erdirik-Dogan and Grossmann,
2008; Maravelias and Sung, 2009), very little work has been done towards the integration of
planning and crude-oil scheduling problems in the context of refineries. This is due to the
fact that in this case, the planning model is not an aggregate scheduling model. Therefore,
the decomposition methods developed for batch and continuous plants are not directly
applicable to refineries. In particular, planning and scheduling correspond to two different
problems solely linked through CDU feedstocks. Therefore, instead of using a hierarchical
decomposition, a spatial Lagrangian decomposition is preferred. The reader may refer to
Fisher (1985) and Guignard (2003) for extensive reviews on Lagrangian relaxation and
decomposition techniques. These approches have been applied to many industrial problems
such as production planning and scheduling integration (see Li and Ierapetritou, 2010) or
multiperiod refinery planning (see Neiro and Pinto, 2006). Thus, it seems natural to apply
Lagrangian decomposition to solve the integrated refinery planning and crude-oil scheduling
problem.
The content of this chapter is organized as follows. First, the planning and scheduling
problems are stated as well as the full-space integrated problem. Next, a Lagrangian decom-
Chapter 6. Integration of Refinery Planning and Crude-Oil Scheduling 120

position scheme based on the dualization of CDU feedstock linking constraints is presented
and a new hybrid method is introduced to solve the Lagrangian relaxation. A heuristic
algorithm is then developed to obtain good feasible solutions for the integrated full-space
problem. After some technical and practical remarks, we present numerical illustrations of
the proposed approach on a small case study and a larger refinery problem.
6.2.1 Refinery Planning Problem
The refinery planning problem can be regarded as a flowsheet optimization problem with
multiple periods during which the refinery system is assumed to operate in steady-state. Due
to extensive stream mixing, the model for each period is based on a pooling problem that
is extended in order to include process models for each refining unit. The different periods
in the model are connected through many material inventories. In this work, we consider
a single-period planning model based on a pooling problem inspired from the literature
(see for instance Adhya et al., 1999). A basic refinery planning system is represented in
Figure 6.1. A set of crudes i ∈ I are to be mixed in different types of crude-oil blends
j ∈ J (e.g. low-sulfur and high-sulfur blends), each associated to a specific CDU operating
mode. For each mode and each crude, several distillation cuts are obtained with different
yields. These crude cuts are then blended into intermediate pools which are used to prepare
several final products. Therefore, the refinery planning system is composed of the following
elements:
• One input stream for each selected crude i ∈ I and each type of crude blend j ∈ J
• One CDU with fixed yields for each distillation cut
• Set of distillation cuts k ∈ K
• One pool for each type of crude blend j ∈ J and each cut k ∈ K
• One intermediate stream between each pool (j, k) ∈ J × K and each final product
l∈L

Crude Distillation Intermediate Product

Feed Cuts Streams Sales
xij
F xijk ijkp
1 , q1 xjkl jkp
2 , q2 xlS , pl
CDU Pooljk Prodl
≤ Ci ≤ Dl
≤ Z lp
∈ [−F R, F R] · H
Figure 6.1: Basic refinery planning system.
• Set of final products l ∈ L

• One sales stream for each final product l ∈ L
• Set of qualities p ∈ P
The yield of crude i ∈ I in distillation cut k ∈ K when processed in crude blend j ∈ J is

assumed to be fixed and is denoted by αijk . In terms of stream qualities, it is assumed that
distillation cuts have fixed qualities while pool qualities are calculated by bilinear quality
balance constraints. A pure flow-based model is used to formulate the pooling problem as
shown below. CDU flowrate limitations are considered independent of the operating mode
and are enforced globally for all crudes processed during the period. The nomenclature
used is as follows:
• Variables:
– xij
F is a variable representing the amount of crude i selected for CDU distillation in
blend j
– xijk
1 is a variable representing the amount of cut k extracted from crude i in blend j
– xjkl
2 is a variable representing the flow of material between pool (j, k) and product l
– q2jkp is a variable representing the quality p of pool (j, k)

– xlS is a variable representing the amount of final product l sold

• Parameters:
– q1ijkp is the (fixed) quality p of cut k extracted from crude i in blend j

– p is the market value of final products
– C i is the amount of available crude i
– αijk is the yield of cut k extracted from crude i in blend j
– H is the planning horizon
– [F R, F R] is the bounds on CDU flowrate
– Dl is the maximum demand in product l
– Z lp is the maximum specification for property p of product l
The NLP model is expressed as follows:

X
max pl xlS (sales revenue maximization)
l∈L
xij
X
i
s.t. 0≤ F ≤C i∈I (crude availability)
j∈J
xij
XX
FR · H ≤ F ≤ FR · H (CDU flowrate limitations)
i∈I j∈J
xijk
1 =α
ijk
· xij
F (i, j, k) ∈ I × J × K (CDU yield calculation)
X ijk X jkl
x1 = x2 (j, k) ∈ J × K (pool mass balance)
i∈I l∈L
(pool quality balance)

q1ijkp xijk jkp
xjkl
X X
1 = q2 2 (j, k, p) ∈ J × K × P
i∈I l∈L nonlinear
xjkl
XX
l
2 = xS l∈L (product mass balance)
j∈J k∈K
xlS ≤ Dl l∈L (maximum product demand)
(product quality requirement)

q2jkp xjkl
XX
lp l
2 ≤ Z xS (l, p) ∈ L × P
j∈J k∈K nonlinear
xij ijk jkl l jkp

F , x1 , x2 , xS ≥ 0, q2 ∈R
Figure 6.2 displays the pooling structure of a case study with corresponding data for
crudes (A, B), blends (X, Y ), distillation cuts (M, N ), and final products (P, Q, R, S). Two
different qualities are considered in this case.

(0.07, 1.10)
(X, A) 0.6 12
(X, M ) P
≤ 250 0.4 (0.08, 1.05) ≤ 120
≤ (0.08, 0.90)
(X, B) 0.5 (0.11, 0.40) 10

(X, N ) Q
≤ 150 0.5 ≤ 80
(0.11, 0.90) ≤ (0.09, 1.00)
CDU
(0.07, 1.10)
(Y, A) 0.4 9
(Y, M ) R
≤ 250 0.6 (0.08, 1.40) ≤ 90
≤ (0.10, 1.10)
(Y, B) 0.3 (0.12, 0.80) 7

(Y, N ) S
≤ 150 0.7 ≤ 110
(0.13, 0.90) ≤ (0.13, 1.30)
∈ [5, 50] · 8
Figure 6.2: Refinery planning case study.
In the remainder of the chapter, we consider the following NLP, which is a simplified

version of the planning model PP .

 max VPT xS





 s.t. fP (xF , xI , xS ) ≤ 0

PP



 gP (xI ) ≤ 0


xF ∈ R|F | , xI ∈ R|I| , xS ∈ R|S|


The nomenclature used is as follows:
• VP is the market value of final products

• xF is a set of continuous variables representing CDU feedstock quantities over the
single planning period
• xS is a set of continuous variables representing final products sales
• xI is a set of intermediate continuous variables (e.g. pool quantity and property
variables)


1 3
4 7
5 8
2 6
Figure 6.3: Refinery crude-oil scheduling system for COSP1.
• fP (xF , xI , xS ) ≤ 0 is the set of linear constraints (e.g. material balance constraints)

• gP (xI ) ≤ 0 is the set of nonlinear constraints (e.g. quality balance constraints)
6.2.2 Crude-Oil Scheduling Problem
The crude-oil scheduling problem deals with the unloading, transfer and blending operations
executed on crude-oil tankers and crude-oil inventories. The goal is to sequentially prepare
multiple crude blends, which are defined by specific property requirements. Each type of
crude blend corresponds to a specific CDU operating mode. Different objectives have been
studied, namely minimization of logistics costs (see chapter 5 and Lee et al., 1996) or
maximization of profit (see chapters 3 and 4). In this work, the objective is to minimize the
total replacement cost of the crudes that are selected for distillation. The replacement cost
is the cost of replacing the crude once it has been processed. The crude-oil schedule must
satisfy inventory capacity limitations, crude tankers arrival dates as well as the logistics
constraints described in chapter 3. Figure 6.3 shows the refinery system corresponding to
problem 1 introduced in Lee et al. (1996). Table 6.1 displays the dimensionless data for
this example. Besides a different objective function, the example is modified by introducing
a minimum duration of one day for distillation operations. Therefore, due to crude blend
alternative sequencing, at most 4 batches of each crude mix can be processed in 8 days.
The scheduling problem is formulated using the MOS time representation as introduced in
chapter 2 (see appendix C).
In the remainder of the chapter, we consider the following MINLP, which is a simplified

Table 6.1: Crude-oil scheduling data for case study.

Vessels Arrival time Composition Amount of crude
Vessel 1 0 100% A 100
Vessel 2 4 100% B 100
Storage tanks Capacity Initial composition Initial amount of crude
Tank 1 [0, 100] 100% A 25
Tank 2 [0, 100] 100% B 75
Charging tanks Capacity Initial composition Initial amount of crude
Tank 1 (mix X) [0, 100] 80% A, 20% B 50
Tank 2 (mix Y) [0, 100] 20% A, 80% B 50
Crudes Property 1 (sulfur concentration) Crude unit cost
Crude A 0.01 7
Crude B 0.06 6
Crude mixtures Property 1 (sulfur concentration) Maximum number of batches
Crude blend X [0.015, 0.025] 4
Crude blend Y [0.045, 0.055] 4
Distillation flowrate [5, 50] Minimum duration of distillations 1 day

version of the scheduling model PS .




 max −VCT yF


 s.t. fS (yB , yC , yF ) ≤ 0

PS



 gS (yC ) ≤ 0


yB ∈ {0, 1}|B| , yC ∈ R|C| , yF ∈ R|F |


The nomenclature used is as follows:
• VC is the replacement cost of crude-oils (usually based on market value)

• yF is a set of continuous variables representing total CDU feedstock quantities over
the scheduling horizon
• yC is a set of continuous variables representing other continuous decisions (e.g. timing
decisions)
• yB is a set of binary variables representing the sequencing decisions (i.e. variables
Ziv )
• fS (yB , yC , yF ) ≤ 0 is the set of linear constraints (e.g. scheduling constraints)
• gS (yC ) ≤ 0 is the set of nonlinear stream composition constraints (i.e. constraint (3.6c))

6.3 Lagrangian Decomposition Scheme
6.2.3 Full-Space Problem
Given the importance of crude selection for refinery optimization, the refinery planning
problem and the crude-oil scheduling problem should ideally be optimized simultaneously.
This can only be done by solving an integrated full-space MINLP problem, denoted (P ),
which aims at optimizing all refinery decisions subject to planning, scheduling, and linking
constraints.



 max VPT xS − VCT yF


s.t. fP (xF , xI , xS ) ≤ 0










 gP (xI ) ≤ 0


fS (yB , yC , yF ) ≤ 0


(P )



 gS (yC ) ≤ 0


yF − x F = 0






xF ∈ R|F | , xI ∈ R|I| , xS ∈ R|S|






yB ∈ {0, 1}|B| , yC ∈ R|C| , yF ∈ R|F |


The integrated objective is to maximize profit defined by final product sales revenues minus
crude-oil replacement costs. The linking constraint yF − xF = 0 ensures consistency of
planning and scheduling decisions in terms of CDU feedstock quantities. More precisely, it
ensures that the amounts of crudes selected for distillation are identical in the planning and
scheduling solutions. Also, to be consistent in time, it is considered that the planning and
scheduling horizons have identical lengths.
In the remainder of the chapter, we use the notation v (P ) to denote the optimal objective

value for problem P .
The full-space problem (P ) is a large-scale MINLP, which contains many binary variables
from the crude-oil scheduling problem and many non-convex constraints from the refinery
planning model. Due to convergence issues and the presence of many potential local optima,

standard MINLP solvers for convex optimization, such as AlphaECP, Bonmin, DICOPT,
KNITRO, or SBB, may fail solving the model or may return poor solutions. Global MINLP
solvers, such as BARON, Couenne, or LINDOGLOBAL, are in principle able to solve the
problem but they may require prohibitive computational times. Therefore, a specific solu-
tion strategy needs to be developed to address this problems.
Robertson et al. (2010) proposed a multi-level approach consisting of approximating the
refinery planning model by multiple linear regressions that are then used in the crude-
oil scheduling problem for the minimization of the total logistics and production costs.
The method is applied to a case study comprising two different crudes. Although com-
putationally effective, the use of linear regressions may not be sufficient for the the global
optimization of highly nonlinear refinery planning model.
In this work, we present an integration approach based on Lagrangian decomposition,
which is a special case of Lagrangian relaxation (Guignard, 2003). The idea is to build
a relaxed version of the full-space problem, which is decomposable, and therefore much
easier to solve. In particular, the decomposition procedure is based on the dualization

of the linking constraint yF − xF = 0. The relaxed problem PR (λ) , composed of NLP
and MINLP models, is defined by removing this constraint and penalizing its violations by
adding the term λT (yF − xF ) to the objective function. The parameter λ is a Lagrange
multiplier whose value is fixed prior to solving the model and adjusted iteratively.



 max VPT xS − λT xF + (λ − VC )T yF






 s.t. fP (xF , xI , xS ) ≤ 0


gP (xI ) ≤ 0





PR (λ) fS (yB , yC , yF ) ≤ 0



gS (yC ) ≤ 0






xF ∈ R|F | , xI ∈ R|I| , xS ∈ R|S|






yB ∈ {0, 1}|B| , yC ∈ R|C| , yF ∈ R|F |



As already mentioned, problem PR (λ) is easier to solve as it can be decomposed into two


subproblems PP (λ) and PS (λ) .
v (PR (λ)) = v (PP (λ)) + v (PS (λ))

The subproblem PP (λ) , an NLP, is a modification of the original refinery planning

problem PP as it consists of assigning crude costs λ to the CDU feedstock variables
xF . For a given crude i, increasing λi will decrease the incentive to select this crude for
distillation processing. On the other hand, decreasing λi will increase the incentive to select
it.




 max VPT xS − λT xF


 s.t. fP (xF , xI , xS ) ≤ 0

PP (λ)



 gP (xI ) ≤ 0


xF ∈ R|F | , xI ∈ R|I| , xS ∈ R|S|



The subproblem PS (λ) , an MINLP, is a modification of the original crude-oil scheduling
problem (PS ) as it consists of assigning crude values λ to the CDU feedstock variables yF .
For a given crude i, increasing λi will increase the incentive to select this crude for blending
and distillation processing. On the other hand, decreasing λi will decrease the incentive to
select it.




 max (λ − VC )T yF


 s.t. fS (yB , yC , yF ) ≤ 0

PS (λ)



 gS (yC ) ≤ 0


yB ∈ {0, 1}|B| , yC ∈ R|C| , yF ∈ R|F |


On the whole, the Lagrange multiplier λ can be seen as a crude purchase cost for the
planning system, and as a crude sales value for the scheduling system. Therefore, the spatial
Lagrange decomposition procedure applied to this problem can be seen as introducing a
crude market between the planning and scheduling systems (see Fig. 6.4). The planning
system acts as a consumer who buys crude from the market, while the scheduling system
acts as a producer who sells it to the market. It is clear that, for fixed prices, both actors

6.4 Solution of the Dual Problem
Crude
Market
(λ) xF
yF
Scheduling Planning
System yF = xF System
Figure 6.4: Economic interpretation of the Lagrangian decomposition.
are independent, which explains why the two corresponding subproblems can be solved in
parallel.
Although computationally convenient, this decomposition procedure does not solve the

original full-space problem. In particular, it is well-known that PR (λ) is a relaxation of
the full-space problem, thus v (PR (λ)) > v (P ). However, one can search for a Lagrange
multiplier λ that minimizes v (PR (λ)) in order to get as close as possible to v (P ). This
problem is called the dual problem and the function λ 7→ v (PR (λ)) is often called the
Lagrangian function.

PD min v (PR (λ))
λ
Duality theory establishes that v (PD ) − v (P ) ≥ 0 (see Guignard, 2003). In some cases
(e.g. nonconvex models), we may have v (PD ) − v (P ) > 0 and this difference is called dual
gap. Our hope is that this dual gap is small enough so that valuable information can be
inferred to generate near-optimal heuristic solutions (see section 6.5).
Several approaches have been proposed in the literature in order to solve the dual problem
associated with the Lagrangian relaxation. A classical approach is the subgradient method
proposed by Held and Karp (1971) and Held and Karp (1974). Many researchers have used
and improved this technique over the years (see Camerini et al., 1975; Bazaraa and Sherali,

1981; Fisher, 1981). This approach is preferred as it usually predicts very good Lagrange
multiplier updates. However, special care must be taken in order to ensure convergence and
it requires a good strategy for defining and updating the subgradient step size. Another
approach that theoretically displays better convergence properties has been introduced by
Cheney and Goldstein (1959) and Kelley (1960). It is often denoted as the cutting plane
method. In practice, this approach usually takes a long time to converge as many iterations
are necessary in order to obtain good Lagrange multiplier updates. A refinement of this
approach is the boxstep method presented in Marsten et al. (1975). It allows obtaining
better updates for the Lagrange multiplier during early iterations while keeping the same
convergence properties. Other refinements of previous approaches include the bundle method
(Lemaréchal, 1974), the volume algorithm (Barahona and Anbil, 2000), and the analytic
center cutting plane method (Goffin et al., 1992).
All the above approaches are based on an iterative solution procedure between the primal
and the dual world. Figure 6.5 gives a schematic description of this algorithm. The first
step consists of initializing the Lagrange multipliers. Problem-specific strategies, often based
on the economic interpretation of λ, exist in order to provide good initial values. Then,
at each iteration the relaxed problem is solved and a primal solution is obtained. If a
stopping criterion is satisfied, the algorithm converges. Otherwise, the Lagrange multipliers
are updated for the next iteration. The definition of the stopping criterion depends on the
approach used and can, in certain cases, ensure convergence to an optimal dual solution λ∗ .
In this work, we introduce a new hybrid method to update the Lagrange multipliers. It is
based on the three concepts of cutting planes, subgradient and boxstep. Cutting planes are
valid constraints for the dual problem that are generated at each iteration. They are used to
record valuable dual information to be used during later iterations. A subgradient provides
a descent direction for the dual problem, while a boxstep is defined to allow deviations from
this direction within a specified domain. The combination of these techniques ensures good
convergence properties while providing efficient Lagrange multiplier updates. At iteration
K + 1, the Lagrange multiplier is updated to the solution of the following restricted LP

Update λK Dual World
Solve PR (λK )

K = 1
Primal World
Initialize λ1 Get xK
F and yF
K
Return λ∗ and yes no

Converged ? K = K +1
v (PR (λ∗ ))
Figure 6.5: General iterative primal-dual algorithm.
dual problem with subgradient-based boxstep.




 min η


η ≥ VPT xkS − VCT yFk + λT (yFk − xkF ) ∀k = 1 . . . K (CPk )



 s.t.

v(PD )−v (PR (λK )) K

P̂DK+1 λ = λK + α (yF − xK

2 F)+δ (SG+BS)

 ||yFK −xK
F ||

η ∈ R, λ ∈ R|F |






α ∈] − ∞, α], δ ∈ [−δ, δ]|F |


The variables λ and η are classically used in the pure cutting plane approach. The
pure cutting plane restricted LP dual problem PDK+1 consists of minimizing η subject to

constraints (CPk ), k = 1 . . . K only:




 min η

PDK+1

s.t. η ≥ VPT xkS − VCT yFk + λT (yFk − xkF ) ∀k = 1 . . . K (CPk )


η ∈ R, λ ∈ R|F |


This problem is always unbounded during early iterations (i.e. v PDK = −∞) so it cannot

be used directly to update the Lagrange multipliers. This issue can be solved by defining a
bounded feasible set for the multipliers based on their interpretation (e.g. lower and upper
bounds). However, in this work, an iteratively updated boxstep based on the subgradient
step is used so that the restricted dual problem P̂DK+1 is bounded.

The subgradient step is classically defined as

2 K
λ = λK + α v (PD ) − v PR (λK )
K
/ yF − xK
F
(yF − xK
F)

λ2 η
subgradient step
λK
2 PR (λK K
1 , λ2 )
ep
st
t
en
di
ra
bg
CPK−1
su
CPK
λ1 λ1
λ1 λK
1
λ1 λK
1
(a) (b)
Figure 6.6: Plots of the feasible space of (P̂DK+1 ).
where v (PD ) can be estimated using a heuristic solution for (P ). However, instead of
heuristically updating the step size, it is optimized using variable α, which is bounded by
the parameter α > 0. Note that α is allowed to take negative values. Variable δ defines a
deviation from the subgradient step. The parameter δ > 0 is the maximum deviation in each
multiplier direction and defines a full-dimensional boxstep. Both subgradient and boxstep
concepts are simultaneously embedded in constraint (SG+BS). Note that the parameters
α and δ can be heuristically updated at each iteration. In practice, our computational
experiments have shown that using fixed values for α and δ is a reasonable choice.
Figure 6.6 displays the feasible space (grey area) of (P̂DK+1 ). The projection on the space
of Lagrange multipliers (λ1 , λ2 ) depicts the shape of the subgradient-based boxstep. In the
space of (λ1 , η), CPk represents the projection of the cutting plane generated at iteration k.
Note that the feasible space of P̂DK+1 contains λ1 = λK K K K

1 , λ2 = λ2 , η = v PR (λ1 , λ2 ) .
In both plots (a) and (b), λ1 corresponds to a lower bound on multiplier λ1 induced by the
boxstep constraints.
The stopping criterion for this hybrid strategy is identical to the pure cutting plane
method and is based on the Lagrangian gap between the relaxed primal problem and the

6.5 Heuristic Solutions
restricted dual problem:
v PR (λK ) − v PDK ≤ ε

(6.1)
In the pure cutting plane approach (without constraint (SG+BS)), the optimal value of the
restricted dual problem v PDK iteratively approximates v (PD ) from below since it involves

the minimization of a relaxation of v (PD ).
v PDK ≤ v (PD )

∀K (6.2)
Therefore, the stopping criterion (6.1) ensures convergence to an ε-optimal solution of the

dual problem PD as:
v PR (λK ) − v PDK ≤ ε ⇒ v PR (λK ) − v (PD ) ≤ ε

(6.3)
In the proposed approach, when the pure restricted dual problem PDK becomes bounded,

the stopping criterion (6.1) is used to check convergence while the hybrid restricted dual
problem P̂DK is used to update the Lagrange multipliers. Finite convergence properties

for the pure cutting plane and boxstep methods in the context of mixed-integer linear
programming can be obtained in Frangioni (2005) and Marsten et al. (1975), respectively.
For the rest of the chapter, it is assumed that practical convergence of the proposed hybrid
method can be achieved in the context of integrated refinery planning and scheduling.
In this section, a classical adaptation of the primal-dual algorithm is presented in order

to obtain solutions that satisfy all constraints of the full-space problem, including linking
constraints. As explained by Frangioni (2005), the solution of the Lagrangian dual problem

yields primal information that can be used to generate good heuristic solutions for P . In
this chapter, a heuristic step is introduced in the iterative algorithm to produce valid lower
bounds P LB (see Fig. 6.7). This induces a second stopping criterion based on the dual gap:

Update λK Dual World
K = 1
Solve PR (λK )

Primal World
Initialize λ1 ,
Get xK K K
F , yF , and yB Upper Bound
P LB = −∞
Solve Heuristic Primal World
Update P LB Lower Bound
yes no
Return P LB Converged ? K = K +1
Figure 6.7: Iterative primal-dual algorithm with heuristic step.
v PR (λK ) − P LB ≤ ε. If either of the two stopping criteria is satisfied, the algorithm

converges and returns P LB .

The heuristic algorithm consists of fixing binary variables yB from the crude-oil scheduling
K in the solution of the relaxed problem P (λK ) . As a

formulation to their values yB R
K) ,

consequence, the full-space problem P reduces to a continuous NLP, denoted PH (yB
and can then be effectively solved. If it is feasible and its (local) optimal solution is better
than the previous incumbent, P LB is updated. Otherwise, P LB is left unchanged.





s.t. fP (xF , xI , xS ) ≤ 0










 gP (xI ) ≤ 0


 K, y , y ) ≤ 0
fS (yB
K
 C F
PH (yB )



 gS (yC ) ≤ 0


yF − x F = 0






xF ∈ R|F | , xI ∈ R|I| , xS ∈ R|S|






yC ∈ R|C| , yF ∈ R|F |


K corresponds to fixing the

In this heuristic solution, fixing binary variables yB to yB
selection and sequencing of operations for the crude-oil scheduling system. Therefore, when

6.6 Remarks
K ) , the nonlinear solver has the opportunity to re-optimize all other

solving problem PH (yB
continuous decisions such as quantities, blend recipes, and timing of operations (start time,
duration, and end time). It is crucial to note that the timing of operations can only
be re-optimized if these decisions are handled by continuous variables. For instance, the
discrete-time MOS-FST representation (see chapter 2), uses binary variables to determine
the timing decisions whereas all other representations, denoted by MOS, MOS-SST, and
SOS, use continuous variables instead. This shows a clear advantage, although not intuitive,
of continuous-time scheduling formulations over discrete-time representations. In particular,
the latter might be inefficient in the context of this work as it would decrease the flexibility
of the heuristic algorithm to find good, or at least feasible solutions.
Overall, it is interesting to note that using this approach, the iterative primal-dual algo-
rithm that solves the Lagrangian dual problem acts as a discrete solution generator that
suggests potentially good discrete solutions for the full-space problem. In other words, it
searches the optimal selection and sequencing of operations for the crude-oil scheduling
system. Additionally, it provides an upper bound for the global optimal solution.
6.6 Remarks
6.6.1 CDU Feedstocks and Lagrange Multipliers
The number of Lagrange multipliers highly depends on the CDU feedstock possibilities.
Exactly one multiplier is needed for each feasible combination of crude and type of crude
blend (or corresponding CDU operating mode). In the case study, there are 2 different
crudes which can both be blended in any of the 2 different types of blends, so 4 Lagrange
multipliers are needed to solve the dual problem. Typical large-scale refineries may need 50
and up to 100 Lagrange multipliers.
The optimal value of the Lagrange multipliers correspond to the optimal marginal costs
of the linking constraints for the convexified problem (for more details in the case of MILP
models, see Frangioni, 2005). Therefore, the Lagrange multipliers can be seen as the optimal

6.6 Remarks
X
Y Y
A A
B B X A
Scheduling
8 days
A
X
Y B
A A B
B B
Planning
8 days 12 days
Figure 6.8: Crude-oil scheduling and multi-period refinery planning integration.
pricing strategy between the crude-oil scheduling system and the refinery planning system.
In other words, the optimal Lagrange multipliers are crude prices for which it is marginally
equivalent to either exchange crudes between the two systems or sell and buy to and from
the crude market (see Fig. 6.4). From this observation, it is natural to use the crude costs
defined in the crude-oil scheduling problem as initialization for the Lagrange multipliers.
For the case study, we use the following initial values:
λ1(X,A) = λ1(Y,A) = 7
λ1(X,B) = λ1(Y,B) = 6
6.6.2 Multi-Period Refinery Planning
In this work, the refinery planning problem is expressed over a single period for which CDU
feedstocks are synchronized with the crude-oil scheduling problem. Even though it is often
computationally critical to increase the time horizon for scheduling problem, this can easily
be done in refinery planning models by introducing additional time periods. Therefore,
one can define 2 or more time periods for the refinery planning model and synchronize
CDU feedstocks for the first period only, as shown in Figure 6.8. In this particular case, the

6.6 Remarks
X Y Y
A X
A A
B B X
B A
Scheduling
3 days 3 days 2 days
A
X Y B
A X
A A B
B
B B
Planning
3 days 3 days 2 days 12 days
Figure 6.9: Disaggregated CDU feedstocks synchronization.
refinery planning decisions for the second period, including CDU feedstock decisions xF , are
made without taking into account crude-oil scheduling constraints. This methodology allows
making short-term scheduling decision while considering the long-term economic impacts
of these decisions, which cannot be done with detailed long-term scheduling models due to
their computational complexity.
6.6.3 CDU Feedstocks Aggregation
An important issue with the proposed approach comes from the fact that the CDU feed-
stocks for the linking period are aggregated. In the optimal solution, the crude-oil operations
schedule prepares several batches for each type of crude blends. Then, for each of these blend
types, all the corresponding batches are accumulated and the refinery planning solution de-
termines the processing decisions for the aggregated batch. This approximation may lead to
sub-optimality or even technical infeasibility of the solutions obtained. This problem can
be solved by postulating exactly one period for each batch and synchronizing all the cor-
responding CDU feedstocks. Figure 6.9 depicts the synchronization of disaggregated CDU
batches.

6.6 Remarks
6.6.4 Handling Nonlinearities in Crude-Oil Scheduling Model

Although, the relaxed problem PR (λ) is decomposable, it is not easy to solve to global
optimality. In particular, two major issues arise. First, the crude-oil scheduling problem

PS (λ) corresponds to an MINLP due to the presence of nonlinear composition constraints.
In chapter 3, an MILP relaxation is derived by simply dropping these nonlinear constraints.
The solution can then be refined by fixing the binary variables and solving the reduced
NLP, similarly to the heuristic approach presented in section 6.5. Results show that the
solution obtained is close to the global optimum as it tends to satisfy the relaxed nonlin-
ear constraints. Therefore a similar methodology is used in this work. Instead of simply
dualizing the linking constraints, the nonlinear scheduling constraints gS (yC ) ≤ 0 are also
relaxed (i.e. dropped). The modified relaxed full-space MINLP problem (NLP + MILP) is

denoted by P̃R (λ) :



 max VPT xS − λT xF + (λ − VC )T yF






 s.t. fP (xF , xI , xS ) ≤ 0


gP (xI ) ≤ 0


P̃R (λ)



 fS (yB , yC , yF ) ≤ 0


xF ∈ R|F | , xI ∈ R|I| , xS ∈ R|S|






yB ∈ {0, 1}|B| , yC ∈ R|C| , yF ∈ R|F |



The corresponding modified crude-oil scheduling MILP subproblem is denoted by P̃S (λ) :



 max (λ − VC )T yF

P̃S (λ) s.t. fS (yB , yC , yF ) ≤ 0


yB ∈ {0, 1}|B| , yC ∈ R|C| , yF ∈ R|F |


The decomposability property is preserved:

v P̃R (λ) = v (PP (λ)) + v P̃S (λ)

Finally, the modified dual problem P̃D can be defined as:

P̃D min v P̃R (λ)
λ

6.6 Remarks
This modified dual problem still provides a valid upper bound for the original full-space
K ) is also defined. It is

problem P . The following modified heuristic problem P̃H (yB
K ) by dropping the nonlinear scheduling

obtained from the original heuristic problem PH (yB
constraints. It is used as a first heuristic step to get a good initial point before solving the
K) .

original heuristic NLP problem PH (yB









 s.t. fP (xF , xI , xS ) ≤ 0


gP (xI ) ≤ 0





K

P̃H (yB ) K, y , y ) ≤ 0
fS (yB C F



yF − x F = 0






xF ∈ R|F | , xI ∈ R|I| , xS ∈ R|S|






y ∈ R|C| , y ∈ R|F |


C F
6.6.5 Handling Nonlinearities in the Refinery Planning Model

In order to obtain valid upper bounds when solving PR (λ) or P̃R (λ) , the refinery plan-

ning problem PP (λ) has to be solved to global optimality. Although global optimization
of industrial large-scale refinery planning models is still unachievable, the refinery planning
case study presented in section 6.2 is solvable by the global NLP solver BARON in reason-
able time. However, it should be noted that it is critical to provide tight bounds for the
quality variables q2jkp . In particular, based on the structure of the pooling system, we use
the following bounds:
min q1ijkp ≤ q2jkp ≤ max q1ijkp

i∈I i∈I
6.6.6 Detailed Implementation
Based on previous remarks to handle nonlinearities, the complete heuristic algorithm is

developed as depicted in Figure 6.10. Although global optimality cannot be ensured, the

dual gap can be estimated using the upper bound provided by v P̃R (λ∗ ) . Local NLP

6.6 Remarks
Dual World
Solve PDK

CPLEX
Solve P̂DK

Dual World
Update λK CPLEX
K = 1 Primal World
Solve P̃R (λK )

Initialize λ1 , Upper Bound
Get xK K K
F , yF , and yB
P LB = −∞ BARON+CPLEX
K)
Primal World
Solve P̃H (yB
CONOPT
K)
Primal World
Solve PH (yB
Lower Bound
Update P LB
CONOPT
yes no
Return P LB Converged ? K = K +1
?

v (P̃R (λK ))−P LB


 v (P̃R (λK ))
≤ε (dual gap)

or
K K ?
 v(P̃R (λ ))−v(PD )


≤ ε (Lagrangian gap)

v (P̃R (λ ))
K
Figure 6.10: Complete algorithm implementation.
solvers, such as CONOPT, are used for the heuristic steps as the solution time is more
critical than global optimality for these problems. The hybrid restricted dual problem
P̂DK is solved using the best solution of the modified heuristic problems P̃H (yB
K ) to

estimate the optimal dual solution v (PD ). The stopping criterion is based on relative gaps
but can also be expressed in terms of absolute gaps. Converging on the Lagrangian gap
means that no further improvement of the upper bound can be achieved and the current
Lagrange multipliers are optimal.

6.7 Numerical Illustration
Table 6.2: Lagrangian iterations statistics (6 priority-slots, NLP=SNOPT).
Pure Hybrid Step Modified Modified Original CPU

Dual Dual
Size Relaxation
Heuristic
Heuristic Time
v PDK v P̂DK αK v P̃R (λK ) K) K)

Iteration v P̃H (yB v PH (yB
1 —a —b —b 645.000 400.942 —c 3s
2 —a 389.942 1 689.049 564.000 560.500 9s
3 —a 547.929 1 637.063 592.368 592.368 13s
4 —a 582.790 1 603.377 592.368 592.368 18s
5 —a 591.172 1 609.526 586.191 —c 24s
6 —a 590.629 0.851 598.380 562.881 —c 29s
7 580.717 591.324 -0.617 600.171 586.191 —c 33s
8 592.368 592.785 0.390 594.774 586.191 —c 46s
9 592.369 592.369 1 592.372 592.368 592.368 50s
a
unbounded LP
b
not available
c
locally infeasible NLP
In this section, several computational results are presented for three approaches: the di-
rect MINLP approach, a basic sequential scheduling-planning procedure and the proposed
Lagrangian decomposition method. Experiments have been performed on an Intel Xeon
1.86GHz processor using GAMS as the modeling and algorithmic language. A 1,000 sec-
onds time limit has been used for each run. The following local NLP solvers have been
used: CONOPT, SNOPT and IPOPT. In our experiments, the convergence tolerance ε is
set to 0.0001, the maximum step size parameter α is set to 1 and the step bound parameter
δ to 0.05.
The number of priority-slots for the crude-oil scheduling model is set to 6 and 7. Ta-
bles 6.2 and 6.3 show iteration statistics for the Lagrangian decomposition method using
SNOPT as the heuristic NLP solver. In particular, the optimal value of each problem solved
is given as well as the optimal step size calculated by the hybrid restrict dual problem and
the cumulative CPU time at the end of each iteration (e.g., for 6 priority-slots, the first
iteration took 3 seconds). Dashes are used when the information is not available (Hybrid

Table 6.3: Lagrangian iterations statistics (7 priority-slots, NLP=SNOPT).

Dual Dual
Size Relaxation
Heuristic
Heuristic Time

1 — — — 645.000 393.470 — 3s
2 — 381.562 1 751.494 568.077 568.077 10s
3 — 552.076 1 622.785 592.368 — 16s
4 — 574.735 1 614.178 592.368 592.368 22s
5 — 583.166 1 602.218 592.368 592.079 50s
6 — 588.279 1 617.268 592.368 592.079 56s
7 — 592.856 0.911 600.752 592.368 592.368 66s
8 — 592.835 0.517 595.264 592.368 — 81s
9 — 592.288 0.357 595.292 592.368 592.368 95s
10 592.368 592.369 1 592.369 592.368 592.079 101s
Dual and Step Size for the first iteration), when the problem is unbounded (Pure Dual
during the first few iterations), or when it is locally infeasible (Original Heuristic in some
iterations).
In each case, the global optimal solution is found (see underlined entries in the Original
Heuristic column) and proved optimal. It can be noted that in some iterations the step size
variable α is strictly lower than 1, which corresponds to cases where the pure subgradient
multiplier update would violate some cutting planes generated at previous iterations. The
proposed approach automatically overcomes this issue. The increase of CPU time between
6 and 7 priority-slots is mostly explained by the increase in size of the MILP scheduling
model. Figures 6.11 and 6.12 plot the evolution of the objective value of various problems
solved during the Lagrangian iterations.
The optimal value of the Lagrange multipliers is λ∗(X,A) = λ∗(X,B) = λ∗(Y,A) = λ∗(Y,B) = 7.
Figures 6.13 and 6.14 plot the evolution of the Lagrange multipliers during the Lagrangian
iterations. The proposed approach demonstrates its efficiency through stable updates and
fast convergence of the Lagrange multipliers.
A basic sequential procedure is introduced to compare with the Lagrangian decomposition

800
700
Objective Value
600
500 v PDK

v P̂DK
400
K

v P̃R (λ )
K)

v PH (yB
300
2 4 6 8
Iteration K
Figure 6.11: Lagrangian iteration objective values (6 priority-slots, NLP=SNOPT).
800
700
Objective Value
600
500 v PDK

v P̂DK
400
v P̃R (λK )
K)

v PH (yB
300
2 4 6 8 10
Iteration K
Figure 6.12: Lagrangian iteration objective values (7 priority-slots, NLP=SNOPT).

7.5
7
λK
6.5
(X, A)
(X, B)
(Y, A)
6 (Y, B)
2 4 6 8
Iteration K
Figure 6.13: Lagrange multiplier updates (6 priority-slots, NLP=SNOPT).
7.5
7
λK
6.5
(X, A)
(X, B)
(Y, A)
6 (Y, B)
2 4 6 8 10
Iteration K
Figure 6.14: Lagrange multiplier updates (7 priority-slots, NLP=SNOPT).

Table 6.4: Comparative performance of different MINLP algorithms.
6 priority-slots 7 priority-slots
MINLP Objective CPU Optimality Objective CPU Optimality
Solver Value Time Gap Value Time Gap
Proposed (CONOPT) 592.368 37s 0% 592.368 94s 0%
Proposed (SNOPT) 592.368 50s 0% 592.368 101s 0%
Proposed (IPOPT) 592.368 244s 0% 592.368 833s 0%
Sequential (BARON) 545.000 9s — 545.000 10s —
DICOPT (CONOPT) 545.000 5s — 592.368 7s —
DICOPT (SNOPT) 592.368 429s — 592.368 6s —
DICOPT (IPOPT) 568.077 54s — 592.368 44s —
AlphaECP (CONOPT) 512.324 67s — 545.000 120s —
AlphaECP (SNOPT) 512.324 67s — 545.000 395s —
AlphaECP (IPOPT) 512.324 69s — 545.000 175s —
SBB (CONOPT) 592.368 267s — 592.368 +1,000s —
SBB (SNOPT) — +1,000s — — +1,000s —
SBB (IPOPT) — +1,000s — 493.536 +1,000s —
LINDOGLOBAL 568.077 +1,000s 11.9% 532.857 +1,000s 17.1%
BARON 592.170 +1,000s 7.3% 400.000 +1,000s 37.5%

approach. First, the modified crude-oil scheduling problem P̃S is solved and the binary
variables are fixed to their solution value. Then, the modified and original heuristic problems
0 ) and P (y 0 ) are successively solved. This procedure is not computationally

P̃H (yB H B
expensive as it requires solving only one MILP, solved with CPLEX, and two NLPs, both
solved with BARON. However, the solution obtained, if any, might not be optimal.
Additionally, the Lagrangian decomposition approach is compared to the direct approach

which consists of solving the full-space problem P with various MINLP solvers. Table 6.4
shows the computational performance of these MINLP algorithms. The sequential approach
quickly provides a feasible solution which is 8.0% lower than the global optimum (545.000
against 592,368). The solvers DICOPT, AlphaECP and SBB cannot guarantee global op-
timality of the solution returned while LINDOGLOBAL and BARON are actual global
MINLP solvers. DICOPT is able to find the global optimal solution in many cases in rea-
sonable CPU times. AlphaECP never finds the global optimum. SBB seems to return the
best solutions when CONOPT is used as NLP solver but it is requires large CPU times.

Amount of crude in each blend

200 Crude A
Y Crude B
X
150
Y
Y Y
X X
100 X
50
0
592.368 568.077 545.000 512.324
Solution objective values
Figure 6.15: Blend compositions in solutions obtained with 6 priority-slots.
Neither LINDOGLOBAL or BARON have found the global optimum in the specified time
limit. In comparison to these solvers, the proposed Lagrangian decomposition approach
proves to be very effective for the following reasons:
• it is computationally effective (although DICOPT is faster);

• it always returns the global optimum;
• it is very robust with the choice of NLP solver (although IPOPT is significantly
slower).
Figure 6.15 depicts the composition of each crude blend in four different solutions. Crude
blend X is mostly composed of crude A while crude blend Y is mostly composed of crude
B. If all solutions except the second one (with objective value 568.077) are considered,
one may conclude that blend Y is ”more profitable” than blend X because the objective
value increases when processing larger amounts of blend Y and smaller amounts of blend
X. Additionally, one could say that increasing the amount of distilled crude increases the
overall profit (solution 1 processes 253.034 units of crude, solution 3, 224.49 units of crude,
and solution 4, 212.867 units of crude). Interestingly, the second solution does not follow
these observations. In this solution, the amount of blend X is larger the amount of blend Y .

6.8 Larger Refinery Problem
Besides, the total amount of crude processed is the largest (276.923 units of crude). This
shows how difficult it is to approximate the economic behavior of refinery operations, even
for such a small case study. Therefore, it is possible that linear approximations of refinery
operations as proposed by Robertson et al. (2010) might not be able to correctly evaluate
the economic value of some feasible solutions.
In this section, the proposed Lagrangian decomposition approach is applied on a larger

refinery example and compared to standard MINLP solvers and the basic sequential proce-
dure presented in section 6.7. The refinery planning problem is based on a crude distillation
simulation model developed by Gueddar and Dua (2010). This crude distillation model is
based on a layered artificial neural network (ANN). This ANN is generated by solving an
MINLP which aims at fitting empirical data for atmospheric distillation of several crudes
(crude assays) while simplifying the calculations. The model obtained is able to predict
cut yields and cut properties from the chosen cut points and properties of the inlet crude.
Figure 6.16 depicts the full nonlinear planning model. Three different types of crude blends
are processed in three different CDU operating modes and five crude cuts are produced:
LPG (liquefied petroleum gas), naphta, kerosene, diesel, and residue. Each discrete CDU
mode is defined by the distillation cut point between diesel and residue fractions: 340, 360,
or 380◦ C. The decision variables for each CDU mode are composed of individual crude
flows and distillation cut points between naphta, kerosene and diesel fractions. The three
streams produced for each fraction are then blended into cut pools. The bilinear pooling
constraints discussed in section 6.2 are classically used to calculate the properties of each
cut pool. Table 6.5 shows market prices for each distillation cut and corresponding property
specifications. Crude availability constraints are also included in the model in accordance
with the description of the crude-oil scheduling problem. This refinery planning model is
composed of 1,831 variables and 1,817 constraints. The full mathematical model is detailled

Distillation cut points:

Naphta/Kerosene ∈ [145, 175]◦ C
Kerosene/Diesel ∈ [220, 250]◦ C
Diesel/Residue = 340◦ C LPG
Naphta
Crude Blend 1
CDU Mode 1 Kerosene
Diesel
Naphta/Kerosene ∈ [145, 175]◦ C Residue
Diesel/Residue = 360◦ C LPG Pool LPG
Naphta Pool Gasoline
Crude Blend 2
CDU Mode 2 Kerosene Pool Kerosene
Diesel Pool Diesel
Naphta/Kerosene ∈ [145, 175]◦ C Residue Pool Residue
Diesel/Residue = 380◦ C LPG
Naphta
Crude Blend 3
CDU Mode 3 Kerosene
Diesel
Residue
Figure 6.16: Planning model for larger refinery problem.
in appendix D.
The crude-oil scheduling problem is based on example 3 from Lee et al. (1996) (see
Figure 6.17) with the parameters described in Table 6.6. It consists of three crude arrivals,
three storage tanks, three charging tanks (one for each type of crude mixture), and two
identical CDUs whose respective scheduled feedstocks are merged when linked to the refinery
planning problem. Seven different crudes are available, and fourteen transfer operations can
be executed to prepare the different crude blends. When 6 priority-slots are used, the crude-
oil scheduling problem is composed of 1,814 variables (84 binary) and 2,338 constraints.
Table 6.7 and Figure 6.18 show the iteration statistics for the Lagrangian decomposition
method using 6 priority-slots for the crude-oil scheduling model. The maximum step size
parameter α is set to 1 and the step bound parameter δ to 0.05. Because we were not

Table 6.5: Crude cut prices and specification for larger refinery problem.
Crude Cut Price Specifications
LPG 8.5 None
specific gravity ∈ [0.72, 0.775]
motor octane number ≥ 45
Naphta 8.0
research octane number ≥ 45
sulfur weight content ≤ 120ppm
Kerosene 7.0
freeze point ≤ −40◦ C
cetane number ≥ 48
Diesel 8.0
cloud point ≤ 4◦ C
sulfur weight content ≤ 2800ppm
Residue 6.5 None

1 4
5 11
6 12
2 7
8 13
9 14
3 10
Figure 6.17: Refinery crude-oil scheduling system for COSP3.
able to solve the refinery planning problem to global optimality, we used CONOPT as the
NLP solver. Other local NLP solvers as SNOPT and IPOPT did not perform as well (slow
convergence, poor solutions or local infeasibilities). As the refinery planning problem is not
solved to global optimality, it is not possible to rigorously estimate global optimality for
the solution obtained for the full-space problem. Therefore, the convergence tolerance ε is
set to 0.01 instead of 0.0001 as in the case study. Furthermore, many iterations are needed
to generate enough cutting planes and make the pure restricted dual problem bounded. In
order to achieve convergence after a few iterations, the Lagrangian gap is calculated using

Table 6.6: Crude-oil scheduling data for larger refinery problem.

Vessels Arrival time Composition Amount of crude
Vessel 1 0 100% A 500
Vessel 2 4 100% B 500
Vessel 3 8 100% C 500
Storage tanks Capacity Initial composition Initial amount of crude
Tank 1 [0, 1,000] 100% D 200
Tank 2 [0, 1,000] 100% E 200
Tank 3 [0, 1,000] 100% F 200
Charging tanks Capacity Initial composition Initial amount of crude
Tank 1 (mix 1) [0, 1,000] 100% G 300
Tank 2 (mix 2) [0, 1,000] 100% E 500
Tank 3 (mix 3) [0, 1,000] 100% F 300
Crudes Property 1 (sulfur concentration) Crude unit cost
Crude A 0.01 6
Crude B 0.085 6.5
Crude C 0.06 5.5
Crude D 0.02 7.2
Crude E 0.05 6.7
Crude F 0.08 6.2
Crude G 0.03 7.5
Crude mixtures Property 1 (sulfur concentration) Maximum number of batches
Crude mix 1 [0.025, 0.035] 6
Crude mix 2 [0.045, 0.065] 6
Crude mix 3 [0.075, 0.085] 6
Distillation flowrate [5, 50] Minimum duration of distillations 1 day
the objective value of the hybrid dual problem, which is always bounded, as follows:

v P̃R (λK ) − v P̂DK ?
≤ε
v P̃R (λK )
The proposed approach converges on the Lagrangian gap in 15 iterations. The final dual
gap, although it does not represent a valid optimality gap, is 3.8%. 83% of the time is spent
on solving the crude-oil scheduling MILPs. The optimal value of the Lagrange multiplier
is given in Table 6.8. With 7 priority-slots, the decomposition procedure converges in
18 iterations and 4,451 seconds, the increase of CPU time being explained mostly by the
increased number of priority-slots. The solution obtained is slightly lower than the previous
one because the algorithm did not fully converge within tight tolerances.
Table 6.9 presents computational performances of the different approaches. Clearly, the

Table 6.7: Lagrangian iterations statistics for larger refinery problem (6 priority-slots,
NLP=CONOPT).

Dual Dual
Size Relaxation
Heuristic
Heuristic Time

1 — — — 266.226 — — 43s
2 — -2.166 1 417.311 257.943 222.339 102s
3 — 253.546 1 411.657 258.488 245.306 173s
4 — 281.076 0.857 325.906 258.106 — 235s
5 — 260.746 1 288.587 258.104 — 298s
6 — 256.301 1 273.197 258.516 245.467 372s
7 — 256.385 1 271.415 258.027 250.989 459s
8 — 258.091 1 264.974 257.961 233.713 539s
9 — 258.847 0.861 265.149 257.961 233.713 610s
10 — 257.623 1 267.541 257.740 228.230 682s
11 — 259.500 0.868 261.237 257.750 247.971 769s
12 — 259.170 1 261.512 257.750 247.971 850s
13 — 259.126 0.759 264.109 258.027 250.989 913s
14 — 259.121 1 263.708 258.306 — 979s
15 — 259.592 0.774 260.857 258.516 238.763 1,045s
400
Objective Value
300
200

100 v P̂DK

v P̃R (λK )
0 v PH (yBK)

0 2 4 6 8 10 12 14 16
Iteration K
Figure 6.18: Lagrangian iteration objective values for larger refinery problem (6 priority-
slots, NLP=CONOPT).

Table 6.8: Optimal Lagrange multipliers for each crude and each CDU mode.
Crude CDU Mode Initial value (=Price) Optimal value
1 6.0 6.169
A 2 6.0 6.207
3 6.0 7.695
1 6.5 6.971
B 2 6.5 6.987
3 6.5 7.026
1 5.5 6.062
C 2 5.5 6.009
3 5.5 6.054
1 7.2 7.186
D 2 7.2 7.262
3 7.2 7.962
1 6.7 6.859
E 2 6.7 6.878
3 6.7 6.898
1 6.2 6.340
F 2 6.2 6.226
3 6.2 6.293
1 7.5 6.930
G 2 7.5 7.360
3 7.5 7.360
Table 6.9: Comparative performance of several MINLP algorithms for larger refinery prob-
lem (NLP solver: CONOPT).
6 priority-slots 7 priority-slots
MINLP Objective CPU Objective CPU
Solver Value Time Value Time
Proposed 250.989 1,045s 250.128 4,451s
Sequential 116.814 46s — 79s
DICOPT — +3,600s — +3,600s
AlphaECP — +3,600s — +3,600s
SBB — +3,600s — +3,600s

6.9 Conclusion
Table 6.10: Blend compositions in the optimal solution of larger refinery problem.
Crude Blend 1 Blend 2 Blend 3
A 0.164 49.480
B 33.293 16.707
C 50.000
D 0.066 19.792
E 3.500 55.943 2.982
F 50.000
G 9.480
Lagrangian decomposition procedure is much more robust than the other algorithms. Only
the sequential approach was able to deliver a solution with 6 priority-slots, but its objective
value is much lower (50% reduction) than the best known solution. Indeed, the scheduling
solution obtain during the first stage of the sequential procedure does not take into account
the economic impact on the refinery planning problem, which leads to poor decisions. The
standard local MINLP solvers were not able to find a solution within one hour because the
MINLP model is too large: 3,645 variables (84 binary) and 4,177 constraints.
Table 6.10 shows the crude compositions of the optimal solution with objective value
250.989. The two CDUs are mostly operated in mode 2 which corresponds to the average
cut point for diesel and residue cuts.
6.9 Conclusion
In this chapter, a novel approach towards the integration of planning and scheduling has
been developed in the context of oil refining. In particular, a precise crude-oil operations
scheduling model and a coarse refinery planning model were optimized simultaneously using
Lagrangian decomposition. It makes use of Lagrange multipliers as a way to communicate
economic information between the two subsystems. The methodology leads to a classical
primal-dual iterative algorithm to solve the Lagrangian dual problem. The critical multiplier
update step is performed by solving a new hybrid restricted dual problem. This approach
combines the strengths of cutting planes and subgradient steps and does not require to

6.9 Conclusion
define heuristic updates of parameters during iterations.

Although it is not guaranteed, our results achieved a 0% dual gap for the smaller case
study. It is well-known that augmented Lagrangian techniques (see Li and Ierapetritou,
2010) can ensure closing the dual gap for any instance. However, this would require to solve
the refinery planning subproblem to global optimality, which is not yet achievable in an
industrial context. It is therefore more practical to use standard duality in order to obtain
feasible heuristic solutions for the integrated problem.
The proposed Lagrangian decomposition procedure has been applied to a larger and more
complex refinery problem. The refinery planning problem is based on a crude distillation
simulation model that has been developed by Gueddar and Dua (2010) using an artificial
neural network to fit empirical data. The crude-oil scheduling problem is the third example
of Lee et al. (1996). The results have shown that the proposed approach remains very com-
petitive compared to other MINLP algorithms, namely the two-step MILP-NLP sequential
procedure and standard MINLP solvers.

Chapter 7
Conclusion
In this thesis we have developed models and solution methods to address the problem of
short-term scheduling of refinery crude-oil operations. In chapter 2, we proposed a unified
modeling approach and solution methods for solving process scheduling problems and ap-
plied it to single-stage and multi-stage batch scheduling problems. In chapter 3, we applied
this approach to the crude-oil operations scheduling problem and introduced a two-step
MILP-NLP procedure to solve the MINLP model. In chapter 4, we developed a symmetry-
breaking approach based on regular languages for the SOS time representation in order to
improve its computational performance. In chapter 5, we integrated CP bound contraction
techniques within CPLEX in order to tighten the linear relaxation of the SOS MINLP model
during the branch & bound search. Finally, in chapter 6 we used a Lagrangian decompo-
sition approach to integrate the refinery planning problem with the crude-oil operations
scheduling problem, and we introduced a new hybrid dual algorithm for determining the
optimal Lagrange multipliers.
7.1 Time Representations and Mathematical Models for Pro-

cess Scheduling Problems
In chapter 2, we presented a unified modeling approach for process scheduling problems

based on four different time representations. All time representations make use of the generic
concepts of priority-slots and operations. We established theoretical relations between these
time representations and showed that the MOS time representation can lead to the most
compact mathematical formulations as it requires fewer priority-slots.
Chapter 7. Conclusion 156

7.1 Time Representations and Mathematical Models
Using common sets of parameters, variables, and constraints, we developed MILP mod-
els for each time representation. These models only differ by a few representation-specific
constraints, namely constraints (2.12)-(2.16). The proposed models were automatically
strengthened using the non-overlapping graph of the scheduling system. Strengthened
constraints (2.17)-(2.21) and (2.23)-(2.26) were obtained using the maximal cliques and
maximal bicliques of the graph.
A typical issue with scheduling problems is to select the number of priority-slots to be
used. This issue can be avoided in specific cases as precedence-based scheduling model
for batch processes with fixed number of batches. In this chapter, we presented generic
solution methods for selecting the number of priority-slots. The performance of the additive
approach has been improved by using cutoffs and the new constraint (2.27), which reduces
the search space at each iteration.
The MOS, MOS-SST and MOS-FST time representations have been applied to single-
stage batch scheduling problems with up to 29 orders processed on 4 parallel units. Using
the available data for unit processing times, we derived tight cardinality bounds for the
number of orders to be processed on each unit. In the case of the MOS mode, we presented
a symmetry-breaking constraint (2.29) that was used with unit cardinality bounds in order
to derive the problem-specific strengthened constraints (2.30)-(2.32). The computational
results show that all instances can be solved to rigorous global optimality in less than
810s, the worst case being the instance with 25 orders. Furthermore, the results showed
that during the additive approach, the MOS model for instance SSBSP18 with 10 priority-
slots was solved in less than 2s with constraint (2.27) while the same iteration takes more
than 100s if it is not used. Similarly, the results showed than the MOS model for instance
SSBSP18 with 5 priority-slots was solved in less than 2s with symmetry-breaking constraints
(2.29)-(2.32), while it takes more than 100s if they are not used. Finally, the results showed
that the strengthened non-overlapping constraint (2.20) can lead to a decrease of CPU times
from 3,288s to less than 2s for the same model. The MOS-SST approach failed to solve
instances with 18, 25 and 29 orders in less than 1,000s. The MOS-FST approach only failed

7.1 Time Representations and Mathematical Models
to solve the instance with 29 orders in less than 1,000s.

Additionally, the MOS, MOS-SST and MOS-FST time representations have been applied
to multi-stage batch problems involving up to 10 orders, 5 stages and 25 units. The MOS
approach was able to solve instances with 5 and 8 orders to global optimality in less than
1,000s and failed on instance MSBSP10. No instance was solved to global optimality with
the MOS-SST and MOS-FST approaches in the given time limit.
The proposed unified modeling approach can be seen an integration of several approaches
studied over the past. The main challenge was to use identical concepts in order to derive
efficiently strengthened mathematical formulations for all time representations that can be
applied to a large variety of problems. The SOS time representation has been introduced in
this work and needs to be further developed in order to efficiently solve scheduling problems.
In chapter 4, we studied the application of this time representation in the context of crude-
oil scheduling problems. The main limitation of the proposed approach is its inability to
handle discrete resource constraints, or inventory transfers at start and/or end times of
process operations. In order to apply this approach to a larger class of problems including
these types of constraints (for example in STN and RTN scheduling problems), the proposed
time representations need to be extended to sequence start and end times of all operations
independently.
On the algorithmic part, the additive approach, although briefly cited in previous papers
(for example, see Maravelias and Grossmann, 2003), has been enhanced with the use of
cutoffs and constraint (2.27). To the best of our knowledge, the proposed multiplicative
approach has not been applied to discrete-time scheduling models in the past. It benefits
from the fact that any solution obtained at a given iteration is feasible for the next iteration,
which justifies objective cutoffs as well as stopping criteria based on objective improvement
tolerances. As the initial number of priority-slots may have a large impact on the global
solution time, this approach can benefit from the work of Maravelias and Papalamprou
(2009) who proposed a strategy for the selection of the length of each time period based on
polyhedral results for the discrete-time MILP.

7.2 Short-Term Scheduling of Crude-Oil Operations
7.2 Short-Term Scheduling of Crude-Oil Operations
In chapter 3, we adapted the MOS, MOS-SST, and MOS-FST models presented in chapter 2
to the crude-oil scheduling problem. We showed how the non-overlapping graph can be used
to derive non-trivial strengthened constraints and we introduced new general symmetry-
breaking constraints for the MOS time representation.
We presented a simple two-step MILP-NLP decomposition procedure that provides an
estimate of optimality gap for the returned solution. The first stage MILP is solved using
the additive approach for MOS and MOS-SST time representations and the multiplicative
approach for the MOS-FST time representation. Since the stopping criterion that is used
(no improvement of the objective value) cannot guarantee global optimality of the solution
returned for the MILP, the upper bound and therefore the estimate of optimality gap are not
rigorous. In practice, the upper bound was always valid except in the case of the MOS-SST
model applied to instance COSP2 (the upper bound found was 101.174 instead of 101.175).
All instances were solved within a 3% optimality gap using the MOS time representation.
An important challenge is to reduce this optimality gap for problems displaying stronger
non-convexities. This issue has been addressed in chapter 5 for minimizing total logistics
costs in the crude-oil operations scheduling problem.
In terms of computational performance, the proposed MINLP procedure behaves gener-
ally better than other MINLP solvers on the crude-oil scheduling problem. The approach
benefits from the fact that it is possible to derive a simple, yet very tight MILP relaxation
of the global MINLP. It is interesting to notice that DICOPT displays very similar results.
Indeed, the first iteration of the outer-approximation algorithm implemented in DICOPT
is very similar to the proposed two-step procedure. The only difference comes from the
fact that in DICOPT, the MILP contains first-order Taylor expansion of the nonlinear con-
straints around the solution of the relaxed MINLP (i.e. the NLP relaxation). However, the
issue of selecting the optimal number of priority-slot was not address while using DICOPT.
The results show that the MILP model was always solved within 20s for the MOS time

7.3 Single-Operation Sequencing Model for Crude-Oil Operations Scheduling
representation and within 400s for the MOS-SST time representation. In the MOS-FST
time representation, only the first instance COSP1 was solved within 1,000s. The poor
performance of this discrete-time formulation is due to fact that constraint (2.24) does not
handle effectively variable processing times. Instead, it is preferable to use the discrete-
time formulation proposed in Lee et al. (1996). This type of discrete-time representation
differs conceptually from the representations proposed in this thesis since one execution of
a given operation may be split into adjacent smaller executions, one per time period. This
type of formulation has many advantages when processing times are variables. However,
it may require additional binary variables to represent the priority-slots corresponding to
the actual start and end times of each operation executions (before splitting). For example,
this is required when the scheduling problem involves cardinality constraints.
The solutions obtained for the problem COSP1 showed that schedule optimization can
lead up to a 13.2% profit increase over manual or heuristic scheduling approaches. The
impact of uncertainty in terms of vessel arrival has been studied in the case of problem
COSP2 showing that a reactive scheduling approach may be sufficient in the case of profit
maximization. The incentive of using advanced strategies for handling uncertainties (e.g.
stochastic programming or robust optimization) has not been precisely evaluated.
7.3 Single-Operation Sequencing Model for Crude-Oil Oper-

ations Scheduling
In chapter 4, we studied the SOS time representation applied to the crude-oil operations
scheduling problem. As the non-overlapping graph can be used to generate many non-
redundant strengthened non-overlapping constraints, we developed three heuristic selection
rules for maximal cliques and bicliques that aim at reducing the number of such constraints.
It was shown that the SOS model may display many symmetric solutions. Therefore, we
developed symmetry-breaking constraints in order to reduce the size of the feasible space
of the MILP. The symmetry-breaking approach is based on a problem-specific sequencing

7.4 Tightening the Linear Relaxation of an MINLP Using CP
rule that needs to be derived by the modeler taking into account the scheduling constraints
involved in the system. We proposed to use regular languages and deterministic finite
automata to represent, store and model this sequencing rule. Additionally, we determined
the maximum number of operations needed to represent any solution for each instance. The
SOS model has been solved using the direct approach and fixing the number of priority-slots
to the maximum number of operations.
The four crude-oil scheduling instances were all solved in less than 310s using the SOS time
representation, showing that it is less efficient than the MOS time representation. This is
mostly due to the large size of the corresponding MILP. On average, the SOS model contains
329% more binary variables, 1188% more continuous variables, and 275% more constraints
than the MOS model. Although global optimality cannot be rigorously achieved with
the proposed approach, the MILP solution returned by the SOS model is always optimal.
Finally, we have shown that the symmetry-breaking sequencing rule reduces the CPU time
for solving problem COSP1 with 13 priority-slots from more than 1 hour to only 2s.
The comparative study showed that the SOS time representation is less promising than
the MOS time representation due to its larger model size. However, sequencing rules can be
used to enforce complex sequencing constraints instead of symmetry-breaking constraints.
In principle, they could also be applied to time representations other than SOS even though
multiple operations may be assigned to each priority-slot.
7.4 Tightening the Linear Relaxation of a Crude-Oil Opera-

tions Scheduling MINLP Using CP
In chapter 5, we developed an integrated MILP / CP algorithm in order to tighten the MILP

relaxation of a crude-oil scheduling problem modeled with the SOS time representation.
The objective function studied corresponds to the minimization of the logistics cost and
involves bilinear terms due to storage costs. As shown in Lee et al. (1996), this objective
function can be artificially linearized using a discrete-time representation. However, it is an

7.4 Tightening the Linear Relaxation of an MINLP Using CP
interesting case study for testing and enhancing the MILP-NLP decomposition approach.
In particular, the optimality gaps obtained with the basic two-step procedure range from
11.7% to 21.9%. Therefore there is a clear incentive to improve the upper bound obtained
at the MILP stage.
The hybrid search algorithm is based on deriving variable bounds for volume and time
variables in order to generate valid McCormick convex envelopes during the search. CPLEX
is used as the main search algorithm, thus benefiting from its advanced branching strategies,
heuristics, cutting planes, and efficient memory management. ILOG CP is used to update
variable bounds at each node and tightening McCormick cuts are generated. The optimality
gap for all instances was reduced to less than 7% for all instances. Additionally, a new
improved solution was obtained for problem COSP2. The comparison with other MINLP
solvers show that this approach outperforms other MINLP algorithms, both in terms of CPU
times and optimality of the solutions returned. As demonstrated in chapter 3, DICOPT is
the best MINLP alternative.
The proposed MILP / CP approach can be applied to any MINLP with bilinear terms,
or in general any MINLP with nonlinear terms that can be linearly relaxed using variable
bounds. The model used for CP constraint propagation is critical in the determination
of tight bounds at each node. In this work, the mathematical formulation has been used
directly as a CP model that proved to be sufficient in the case of the crude-oil operations
scheduling problem. Depending on the structure of the problem, the CP model may have to
be further enhanced. Also, the CP model can be used to perform logic inference in addition
to the classic LP relaxation. Finally, the approach would benefit from spatial branch &
bound techniques in order to further improve the upper bound and potentially select better
sequences of operations for the NLP stage. The work developed in this chapter is closely
related to the solvers SIMPL (see Yunes et al., 2010) and SCIP (for linear problems only,
see Achterberg, 2004).

7.5 Integration of Refinery Planning and Crude-Oil Scheduling
7.5 Integration of Refinery Planning and Crude-Oil Schedul-

ing using Lagrangian Decomposition
In chapter 6, we developed a Lagrangian decomposition approach to integrate refinery plan-

ning with the crude-oil operations scheduling problem modeled with the MOS time repre-
sentation. The two subproblems are linked through a single CDU feedstock synchronization
constraint, which is dualized for the purpose of the Lagrangian relaxation. The approach
benefits from the fact that the full-space MINLP is decomposed into an MILP and an NLP
which can be solved effectively by standard solvers. As the approach is based on relaxing
the full-space MINLP, the crude-oil scheduling MINLP subproblem is naturally replaced
by an MILP relaxation as described in chapters 3, 4 and 5. The Lagrange multipliers are
optimized using a hybrid dual problem composed of cutting planes and a subgradient-based
boxstep. Furthermore, we presented a simple heuristic procedure for obtaining a lower
bound that consists of solving the full-space NLP obtained after fixing the binary variables
to the values of the crude-oil scheduling solution obtained at each iteration.
The computational results show that the approach converges quickly in less than 10
iterations (less than 2 min) with 0% optimality gap for a small-scale example solved with
6 and 7 priority-slots. The heuristic procedure found a feasible solution in 58% of the
iterations although only 4 different solutions were found. The optimal solution was found
in both cases, with any NLP solver for the heuristic, and in reasonable times showing the
robustness of the Lagrangian decomposition approach. Among standard MINLP solvers,
DICOPT was the best alternative although it did not find the optimal solution in all cases.
On a larger example, 15 iterations and 1,045s were needed in order to solve the problem with
6 priority-slots. The main computational challenge is to solve the MILP relaxation of the
crude-oil scheduling problem in shorter CPU times. Furthermore, the estimate of optimality
gap for the best known solution is 3.8%. However, this estimate is not rigorous since the
upper bounds obtained at each iteration are not achieved through global optimization of
the relaxed primal problem.

7.6 Contributions of the Thesis
From a practical point of view, the integration of both subproblems may need to be further
improved in order to be applied to industrial problems. In section 6.6, we outlined several
ideas in order to improve this integration. The first one consists of using a multi-period
refinery planning problem in order to consider long-term economic impacts while optimizing
short-term scheduling solutions, thus avoiding solving computationally expensive long-term
scheduling models. Furthermore, the aggregation of crude-oil feedstock over the scheduling
period may lead to inconsistencies as individual CDU batches are not considered in the
refinery planning problem. This issue should be addressed by using exactly one period per
CDU batch.
7.6 Contributions of the Thesis
The main contributions of the thesis can be summarized as follows.
1. A unified modeling approach for process scheduling problems has been proposed in
chapter 2. The methodology can be applied to a large class of scheduling problems
including single-stage batch scheduling, multi-stage batch scheduling and crude-oil
operations scheduling. It is solely based on the general concepts of priority-slots
and operations, which are used to represent any feasible solution by a sequence of
multiple or single operations. The corresponding MILP models only differ by a few
representation-specific constraints.
2. A new time representation, called single operation sequencing (SOS), has been in-
troduced in chapter 2 and applied to the crude-oil operations scheduling problem in
chapters 4 and 5. As the corresponding MILP model displays many symmetric so-
lutions, we proposed to use a symmetry-breaking sequencing rule represented by a
deterministic finite automaton and included it in the model as a network flow formu-
lation. Sequencing rules can also be used to represent complex sequencing constraints.
3. A global representation of the non-overlapping constraints in the scheduling problem
has been developed in chapter 2. It can be seen as an extension of the disjunc-

7.7 Recommendations for Future Work
tive graph (Balas, 1969) for operations that may be executed several times. The
non-overlapping graph has been used to automatically generate strengthened non-
overlapping constraints using its maximal cliques and maximal bicliques.
4. A general symmetry-breaking constraint for MOS scheduling models has been intro-
duced in section 3.3.7. As it does not make any assumption on the structure of the
scheduling system, it can be applied to any problem.
5. A simple two-step MILP - NLP decomposition procedure has been proposed for solving
crude-oil scheduling MINLPs in chapter 3. It takes advantage of the fact that it is
possible to derive a tight MILP relaxation for the MINLP. After solving the MILP,
the binary variables are fixed and the full-space NLP is solved to obtain a feasible
solution with an estimate of the optimality gap.
6. A general integration approach for refinery planning and crude-oil operations schedul-
ing problems has been proposed in chapter 6. It based on a Lagrangian decomposition
of the full-space MINLP problem. The decomposability property is obtain by dual-
izing the linking constraints, which synchronizes CDU feedstocks decisions in both
subproblems.
7. We developed a new generic hybrid dual algorithm for optimizing the Lagrange multi-
pliers in the context of Lagrangian decomposition. This approach is based on solving
an LP dual problem composed of cutting planes and a subgradient-based boxstep.
Only two parameters have to be set up by the user: the maximum step size and the
maximum deviation from the subgradient step.
1. As mentioned in chapter 2, a major limitation of the proposed time representations is

that operations must be sequenced as a whole. However, in network-based scheduling
problems, reaction operations are often decomposed in 3 phases: reactor charging
(considered as instantaneous), reaction process, reactor discharging (also considered

as instantaneous). In order to track inventory levels, it is necessary to sequence start

and end events, independently. This can be accomplished by extending to proposed
time representations and assigning simultaneously two priority-slots to each operation
execution. Any solution can then be represented by a sequence of start and end events,
corresponding to all transfer events. Additionally, this concept can be used to model
discrete resource constraint as start and end events correspond to allocation and
release events.
2. The additive approach presented in chapter 2 has not been applied to the integrated
refinery planning and crude-oil scheduling problem studied in chapter 6. Therefore,
the automatic selection of the number of priority-slots has not been addressed yet.
The integration approach can be improved by integrating this algorithm to the La-
grangian decomposition procedure. In particular, two different approaches could be
investigated. The first one consists of executing the additive approach at each itera-
tion to solve the crude-oil scheduling problem. The second consists of executing the
additive approach outside of the primal-dual procedure using the optimal Lagrange
multiplier form each iteration as a starting point for the next iteration. Additionally,
cutting planes generated at previous iterations may still be re-used for later iteration,
thus improving the convergence of the entire algorithm.
3. The two-step MILP-NLP decomposition procedure proposed in chapter 2 can be ex-
tended in two different ways. For local optimization, the procedure may iterate by
adding outer-approximation cuts generated from the NLP steps (as in DICOPT). The
first MILP provides an upper bound, while successive iterations may provide addi-
tional feasible solutions, thus improving the lower bound. For global optimization,
the NLP problems may be solved with spatial branch and bound and the procedure
may iterate by adding Benders cuts to the MILP master problem (as in Benders de-
composition). In any case, the algorithm may be adapted to the specific structure of
the crude-oil operations scheduling problem.
4. The MILP relaxation tightening approach developed in chapter 5 has not been applied

towards the integration of refinery planning and crude-oil operations scheduling. The
results have shown that the best solution for the larger refinery example in chapter 6
has been obtained with an optimality gap of 3.8%. This optimality gap can be reduced
by improving the upper bound obtained by the Lagrangian relaxation and therefore
by improving the MILP relaxation of the crude-oil scheduling problem as studied in
chapter 5.

Chapter 8
Bibliography
Achterberg, T., 2004. SCIP - a framework to integrate constraint and mixed integer pro-
gramming - a framework to integrate constraint and mixed integer programming. Tech.
rep., Zuse Institue Berlin.
Adams, J., Balas, E., Zawack, D., 1988. The shifting bottleneck procedure for job shop
scheduling. Management Science 34 (3), 391–401.
Adhya, N., Tawarmalani, M., Sahinidis, N. V., 1999. A Lagrangian approach to the pooling
problem. Industrial and Engineering Chemistry Research 38 (5), 1956–1972.
Adjiman, C. S., Androulakis, I. P., Floudas, C. A., 2000. Global optimization of mixed-
integer nonlinear problems. AIChE Journal 46 (9), 1769–1797.
Alhajri, I., Elkamel, A., Albahri, T., Douglas, P. L., 2008. A nonlinear programming model
for refinery planning and optimisation with rigorous process models and product quality
specifications. International Journal of Oil, Gas and Coal Technology 1 (3), 283–307.
Andersen, H. R., Hadzic, T., Hooker, J. N., Tiedermann, P., 2007. A constraint store
based on multivalued decision diagrams. In: 13th International Conference on Principles
and Practice of Constraint Programming. Vol. 4741 of Lecture Notes in Computer Science.
Springer, pp. 118–132.
Audet, C., Brimberg, J., Hansen, P., Le Digabel, S., Mladenović, N., 2004. Pooling prob-
lem: Alternate formulations and solution methods. Management Science 50 (6), 761–776.
Audet, C., Hansen, P., Jaumard, B., Savard, G., 2000. A branch and cut algorithm for
nonconvex quadratically constrained quadratic programming. Mathematical Programming
87 (1), 131–152.
Balas, E., 1969. Machine sequencing via disjunctive graphs: An implicit enumeration
algorithm. Operations Research 17 (6), 941–957.
Balas, E., 1985. Disjunctive programming and a hierarchy of relaxations for discrete opti-
mization problems. SIAM Journal on Algebraic and Discrete Methods 6 (3), 466–486.
Baptiste, P., Le Pape, C., Nuijten, W., 2001. Constraint-based Scheduling: Applying Con-
straint Programming to Scheduling Problems. Vol. 39 of International Series in Operations
Research and Management Science. Kluwer Academic Publishers.
Chapter 8. Bibliography 168

Barahona, F., Anbil, R., 2000. The volume algorithm: producing primal solutions with a
subgradient method. Mathematical Programming 87 (3), 385–399.
Bazaraa, M. S., Sherali, H. D., 1981. On the choice of step size in subgradient optimization.
European Journal of Operational Research 7 (4), 380–388.
Bergamini, M. L., Grossmann, I. E., Scenna, N., Aguire, P., 2008. An improved piece-
wise outer-approximation algorithm for the global optimization of minlp models involving
concave and bilinear terms. Computers and Chemical Engineering 32, 477–493.
Bixby, R. E., Fenelon, M., Gu, Z., Rothberg, E., Wunderling, R., 1999. MIP: Theory
and practice - closing the gap. In: System Modelling and Optimization. IFIP Conference
Proceedings. pp. 19–50.
Bodington, C. E., Baker, T. E., 1990. A history of mathematical programming in the

petroleum industry. Interfaces 20 (4), 117–127.
Camerini, P. M., Fratta, L., Maffioli, F., 1975. On improving relaxation methods by mod-
ified gradient techniques. Mathematical Programming Studies 3, 26–34.
Casas-Liza, J., Pinto, J. M., 2005. Optimal scheduling of a lube oil and paraffin production
plant. Computers and Chemical Engineering 29 (6), 1329–1344.
Castro, P. M., Grossmann, I. E., 2005. New continuous-time milp model for the short-
term scheduling of multistage batch plants. Industrial and Engineering Chemistry Research
44 (24), 9175–9190.
Castro, P. M., Grossmann, I. E., 2006. An efficient milp model for the short-term scheduling
of single stage batch plants. Computers and Chemical Engineering 30 (6-7), 1003–1018.
Castro, P. M., Grossmann, I. E., Novais, A. Q., 2006. Two new continuous-time models for
the scheduling of multistage batch plants with sequence dependent changeovers. Industrial
and Engineering Chemistry Research 45 (18), 6210–6226.
Cerdá, J., Henning, G. P., Grossmann, I. E., 1997. A mixed-integer linear programming
model for short-term scheduling of single-stage multiproduct batch plants with parallel
lines. Industrial and Engineering Chemistry Research 36 (5), 1695–1707.
Cheney, E. W., Goldstein, A. A., 1959. Newton’s method for convex programming and
Tchebycheff approximation. Numerische Mathematik 1 (1), 253–268.
Côté, M. C., Gendron, B., Rousseau, L. M., 2007. Modeling the regular constraint with
integer programming. In: Integration of AI and OR Techniques in Constraint Programming
for Combinatorial Optimization Problems. Vol. 4510 of Lecture Notes in Computer Science.
Springer, pp. 29–43.
Dua, V., 2010. A mixed-integer programming approach for optimal configuration of arti-
ficial neural networks. Chemical Engineering Research and Design 88 (1), 55–60.

Duran, M. A., Grossmann, I. E., 1986. An outer-approximation algorithm for a class of
mixed-integer nonlinear programs. Mathematical Programming 36 (3), 307–339.
Erdirik-Dogan, M. E., Grossmann, I. E., 2008. Simultaneous planning and scheduling of

single-stage multi-product continuous plants with parallel lines. Computers and Chemical
Engineering 32 (11), 2664–2683.
Fisher, M. L., 1981. The Lagrangian relaxation method for solving integer programming
problems. Management Science 27 (1), 1–18.
Fisher, M. L., 1985. An application oriented guide to Lagrangian relaxation. Interfaces 15,
10–21.
Floudas, C. A., Aggarwal, A., 1990. A decomposition strategy for global optimum search
in the pooling problem. ORSA Journal on Computing 2 (3), 225–235.
Floudas, C. A., Lin, X., 2004. Continuous-time versus discrete-time approaches for schedul-
ing of chemical processes: A review. Computers and Chemical Engineering 28, 2109–2129.
Floudas, C. A., Visweswaran, V., 1993. A primal-relaxed dual global optimization ap-
proach. Journal of Optimization Theory and Applications 78 (2), 187–225.
Foulds, L. R., Haugland, D., Jörnsten, K., 1992. A bilinear approach to the pooling prob-
lem. Optimization 24 (1 & 2), 165–180.
Frangioni, A., 2005. About Lagrangian methods in integer optimization. Annals of Oper-
ations Research 139 (1), 163–193.
Furman, K. C., Jia, Z., Ierapetritou, M. G., 2007. A robust event-based continuous time
formulation for tank transfer scheduling. Industrial and Engineering Chemistry Research
46 (26), 9126–9136.
Garvin, W. W., Crandall, H. W., John, J. B., Spellman, R. A., 1957. Applications of linear
programming in the oil industry. Management Science 3 (4), 407–430.
Goffin, J. L., Haurie, A., Vial, J. P., 1992. Decomposition and nondifferentiable optimiza-
tion with the projective algorithm. Management Science 38 (2), 284–302.
Goltz, H.-J., Matzke, D., 1998. Practical Aspects of Declarative Languages. Vol. 1551
of Lecture Notes in Computer Science. Springer, Ch. University Timetabling Using Con-
straint Logic Programming, pp. 320–334.
Gomory, R. E., 1958. Outline of an algorithm for integer solutions to linear programs.
Bulletin of the American Mathematical Society 64 (5), 275–278.
Grossmann, I. E., 2002. Review of nonlinear mixed-integer and disjunctive programming

techniques. Optimization and Engineering 3 (3), 227–252.

Grossmann, I. E., 2005. Enterprise-wide optimization: A new frontier in process systems
engineering. AIChE Journal 51 (7), 1846–1857.
Gueddar, T., Dua, V., 2010. Novel model reduction techniques for refinery-wide energy
optimization. Applied Energy Journal.
Guignard, M., 2003. Lagrangean relaxation. Top 11 (2), 151–228.
Gupta, S., Karimi, I. A., 2003. An improved milp formulation for scheduling multiproduct
multistage batch plants. Industrial and Engineering Chemistry Research 42, 2365–2380.
Hart, W. D., 1978. L.P. behavior - recursion example comments. ACM SIGMAP Bulletin
25, 29–33.
Haverly, C. A., 1978. Studies of the behavior of recursion for the pooling problem. ACM
SIGMAP Bulletin 25, 19–28.
Haverly, C. A., 1980. Recursion model behavior: more studies. ACM SIGMAP Bulletin
28, 39–41.
Held, M., Karp, R. M., 1971. The traveling-salesman problem and minimum spanning
trees: Part ii. Mathematical Programming 1 (1), 6–25.
Held, M., Karp, R. M., 1974. Validation of subgradient optimization. Mathematical Pro-
gramming 6 (1), 62–88.
Hoda, S., van Hoeve, W.-J., Hooker, J. N., 2010. A systematic approach to mdd-based
constraint programming. In: 16th International Conference on Principles and Practice of
Constraint Programming. Lecture Notes in Computer Science. Springer.
Hooker, J. N., 2007. Integrated Methods for Optimization. Springer.
Hopcroft, J. E., Ullman, J. D., 1979. Introduction to Automata Theory, Languages and
Computation. Addison Wesley.
Hui, C., Gupta, A., van der Meulen, H. A. J., 2000. A novel milp formulation for short-term
scheduling of multi-stage multi-product batch plants with sequence-dependent constraints.
Computers and Chemical Engineering 24 (12), 2705–2717.
Ierapetritou, M. G., Floudas, C. A., 1998. Effective continuous-time formulation for short-
term scheduling. 1. Multipurpose batch processes. Industrial and Engineering Chemistry
Research 37 (11), 4341–4359.
ILOG Inc., 2007. ILOG CPLEX 11.0 User’s Manual.
Janak, S. L., Lin, X., Floudas, C. A., 2004. Enhance continuous-time unit-specific event-
based formulation for short-term scheduling of multipurpose batch processes: Resource
consraints and mixed storage policies. Industrial and Engineering Chemistry Research
43 (10), 2516–2533.

Jia, Z., Ierapetritou, M. G., Kelly, J. D., 2003. Refinery short-term scheduling using con-
tinuous time formulation: Crude-oil operations. Industrial and Engineering Chemistry
Research 42 (13), 3085–3097.
Johnson, E. L., Nemhauser, G. L., Savelsbergh, M. W. P., 2000. Progress in linear

programming-based algorithms for integer programming: An exposion. INFORMS Journal
on Computing 12 (1), 2–23.
Joly, M., Pinto, J. M., 2003. Mixed-integer programming techniques for the scheduling
of fuel oil and asphalt production. Chemical Engineering Research and Design 81 (4),
427–447.
Kallrath, J., 2002. Planning and scheduling in the process industry. OR Spectrum 24 (3),
219–250.
Karuppiah, R., Furman, K. C., Grossmann, I. E., 2008. Global optimization for scheduling
refinery crude oil operations. Computers and Chemical Engineering 32 (11), 2745–2766.
Kelley, J. E., 1960. The cutting-plane method for solving convex programs. Journal of the
Society for Industrial and Applied Mathematics 8 (4), 703–712.
Kelly, J. D., Mann, J. L., 2003. Crude oil blend scheduling optimization: an application
with multimillion dollar benefits. Hydrocarbon Processing 82 (6), 47–53.
Kesavan, P., Allgor, R. J., Gatzke, E. P., Barton, P. I., 2004. Outer approximation al-
gorithms for separable nonconvex mixed-integer nonlinear programs. Mathematical Pro-
gramming 100 (3), 517–535.
Kondili, E., Pantelides, C. C., Sargent, R. W. H., 1993. A general algorithm for short-term
scheduling of batch operations. I: MILP formulation. Computers and Chemical Engineering
17 (2), 211–227.
Ku, H., Karimi, I. A., 1988. Scheduling in serial multiproduct batch processes with finite
interstage storage: a mixed integer linear program formulation. Industrial and Engineering
Chemistry Research 27, 1840–1848.
Land, A. H., Doig, A. G., 1960. An automatic method of solving discrete programming
problems. Econometrica 28 (3), 497–520.
Lasdon, L., Joffe, B., 1990. The relationship between distributive recursion and succes-
sive linear programming in refining production planning models. In: NPRA Computer
Conference. Seattle, WA.
Lee, H., Pinto, J. M., Grossmann, I. E., Park, S., 1996. Mixed-integer linear programming
model for refinery short-term scheduling of crude oil unloading with inventory manage-
ment. Industrial and Engineering Chemistry Research 35 (5), 1630–1641.

Lemaréchal, C., 1974. An algorithm for minimizing convex functions. In: Proceedings
IFIP’74 Congress. pp. 552–556.
Leyffer, S., 2001. Integrating SQP and branch-and-bound for mixed integer nonlinear
programming. Computational Optimization and Applications 18 (3), 295–309.
Li, W., Hui, C., Li, A., 2005. Integrating cdu, fcc and product blending models into refinery
planning. Computers and Chemical Engineering 29 (9), 2010–2028.
Li, Z., Ierapetritou, M. G., 2010. Production planning and scheduling integration through
augmented Lagrangian optimization. Computers and Chemical Engineering 34 (6), 996–
1006.
Lin, Y., Schrage, L., 2009. The global solver in the lindo api. Optimization Methods and
Software 24 (4-5), 657 – 668.
Liu, Y., Karimi, I. A., 2007. Scheduling multistage, multiproduct batch plants with
nonidentical parallel unitsandunlimitedintermediatestorage. Chemical Engineering Science
62 (6), 1549–1566.
Manne, A. S., 1958. A linear programming model of the U. S. petroleum refining industry.
Econometrica 26 (1), 67–106.
Maravelias, C. T., Grossmann, I. E., 2003. New general continuous-time state-task net-
work formulation for short-term scheduling of multipurpose batch plants. Industrial and
Engineering Chemistry Research 42 (13), 3056–3074.
Maravelias, C. T., Grossmann, I. E., 2006. On the relation of continuous- and discrete-time
state-task network formulations. AIChE Journal 52 (2), 843–849.
Maravelias, C. T., Papalamprou, K., 2009. Polyhedral results for discrete-time production
planning mip formulations for continuous processes. Computers and Chemical Engineering
33 (11), 1890–1904.
Maravelias, C. T., Sung, C., 2009. Integration of production planning and scheduling:
Overview, challenges and opportunities. Computers and Chemical Engineering 33 (12),
1919–1930.
Marchetti, P. A., Cerdá, J., 2009. An approximate mathematical framework for resource-
constrained multistage batch scheduling. Chemical Engineering Science 64, 2733–2748.
Margot, F., 2008. Symmetry in integer linear programming. Tepper Working Paper 2008
E-37.
Marsten, R. E., Hogan, W. W., Blankenship, J. W., 1975. The boxstep method for large-
scale optimization. Operations Research 23 (3), 389–405.

McCormick, G. P., 1976. Computability of global solutions to factorable nonconvex pro-
grams: Part 1 - convex underestimating problems. Mathematical Programming 10, 147–
175.
Méndez, C. A., Cerdá, J., 2003. Dynamic scheduling in multiproduct batch plants. Com-
puters and Chemical Engineering 27 (8-9), 1247–1259.
Méndez, C. A., Cerdá, J., 2007. A precedence-based monolithic approach to lotsizing and
scheduling of multiproduct batch plants. In: 17th European Symposium on Computer
Aided Process Engineering - ESCAPE17. Vol. 24 of Computer Aided Chemical Engineer-
ing. pp. 679–684.
Méndez, C. A., Cerdá, J., Grossmann, I. E., Harjunkoski, I., Fahl, M., 2006a. State-
of-the-art review of optimization methods for short-term scheduling of batch processes.
Computers and Chemical Engineering 30, 913–946.
Méndez, C. A., Grossmann, I. E., Harjunkoski, I., Kaboré, P., 2006b. A simultaneous
optimization approach for off-line blending and scheduling of oil-refinery operations. Com-
puters and Chemical Engineering 30, 614–634.
Méndez, C. A., Henning, G. P., Cerdá, J., 2001. An milp continuous-time approach to
short-term scheduling of resource-constrained multistage flowshop batch facilities. Com-
puters and Chemical Engineering 25, 701–711.
Meyer, C. A., Floudas, C. A., 2006. Global optimization of a combinatorially complex

generalized pooling problem. AIChE Journal 52 (3), 1027–1037.
Michel, L., van Hentenryck, P., 2003. Comet in context. In: Principles of computing &
knowledge: Proceedings of the Paris C. Kanellakis Memorial Workshop. Vol. 41 of ACM
International Conference Proceeding Series. pp. 95–107.
Milano, M., Wallace, M., 2006. Integrating operations research in constraint programming.
4OR: A Quarterly Journal of Operations Research 4 (3), 175–219.
Misener, R., Floudas, C. A., 2009. Advances for the pooling problem: Modeling, global
optimization, and computational studies. Applied and Computational Mathematics 8 (1),
3–22.
Mitchell, M., 1998. An Introduction to Genetic Algorithms. The MIT Press.
Moro, L. F. L., Pinto, J. M., 2004. Mixed-integer programming approach for short-term
crude oil scheduling. Industrial and Engineering Chemistry Research 43, 85–94.
Mouret, S., Grossmann, I. E., Pestiaux, P., 2008. Multi-operations time-slots model for
crude-oil operations scheduling. In: 18th European Symposium on Computer Aided Pro-
cess Engineering - ESCAPE18. Vol. 25 of Computer Aided Chemical Engineering. pp.
593–598.

Mouret, S., Grossmann, I. E., Pestiaux, P., 2009a. A novel priority-slot based continuous-
time formulation for crude-oil scheduling problems. Industrial and Engineering Chemistry
Research 48 (18), 8515–8528.
Mouret, S., Grossmann, I. E., Pestiaux, P., 2009b. Tightening the linear relaxation of a
mixed integer nonlinear program using constraint programming. In: Integration of AI and
OR Techniques in Constraint Programming for Combinatorial Optimization Problems.
Vol. 5547 of Lecture Notes in Computer Science. Springer, pp. 208–222.
Mouret, S., Grossmann, I. E., Pestiaux, P., 2010a. Integration of refinery planning and
crude-oil scheduling using Lagrangian decomposition. To be submitted.
Mouret, S., Grossmann, I. E., Pestiaux, P., 2010b. Time representations and mathematical
models for process scheduling problems. Computers and Chemical Engineering.
Neiro, S. M. S., Pinto, J. M., 2006. Lagrangean decomposition applied to multiperiod plan-
ning of petroleum refineries under uncertainty. Latin American Applied Research 36 (4),
213–220.
Nemhauser, G. L., Wolsey, L. A., 1999. Integer and Combinatorial Optimization. Wiley-
Interscience.
Pantelides, C. C., 1994. Unified frameworks for optimal process planning and schedul-
ing. In: Rippin, D., Hale, J., Davis, J. (Eds.), Proceedings of the Second International
Conference on Foundations of Computer-Aided Process Operations. pp. 253–274.
Pesant, G., 2004. A regular language membership constraint for finite sequences of vari-
ables. In: 10th International Conference on Principles and Practice of Constraint Pro-
gramming. Vol. 3258 of Lecture Notes in Computer Science. Springer, pp. 482–495.
Pham, V., Laird, C. D., El-Halwagi, M., 2009. Convex hull discretization approach to the
global optimization of pooling problems. Industrial and Engineering Chemistry Research
48 (4), 1973–1979.
Pinto, J. M., Grossmann, I. E., 1995. A continuous time mixed integer linear programming
model for short term scheduling of multistage batch plants. Industrial and Engineering
Chemistry Research 34 (9), 3037–3051.
Pinto, J. M., Joly, M., Moro, L. F. L., 2000. Planning and scheduling models for refinery
operations. Computers and Chemical Engineering 24, 2259–2276.
Pinto, J. M., Moro, L. F. L., 2000. A planning model for petroleum refineries. Brazilian
Journal of Chemical Engineering 17 (4-7), 575–586.
Prasad, P., Maravelias, C. T., 2008. Batch selection, assignment and sequencing in multi-
stage multi-product processes. Computers and Chemical Engineering 32 (6), 1106–1119.

Quesada, I., Grossmann, I. E., 1992. An lp/nlp based branch and bound algorithm for
minlp optimization. Computers and Chemical Engineering 16, 937–947.
Quesada, I., Grossmann, I. E., 1995a. A global optimization algorithm for linear fractional
and bilinear programs. Journal of Global Optimization 6 (1), 39–76.
Quesada, I., Grossmann, I. E., 1995b. Global optimization of bilinear process networks
with multicomponent flows. Computers and Chemical Engineering 19 (12), 1219–1242.
Reddy, P. C. P., Karimi, I. A., Srinivasan, R., 2004. A new continuous-time formulation
for scheduling crude oil operations. Chemical Engineering Science 59 (6), 1325–1341.
Régin, J.-C., 2003. Global constraints and filtering algorithms. In: Constraints and Integer
Programming Combined. Kluwer, Ch. 1.
Robertson, G., Palazoglu, A., Romagnoli, J. A., 2010. Refinery scheduling of crude oil
unloading, storing, and processing considering production level cost. In: 20th European
Symposium on Computer Aided Process Engineering - ESCAPE20. Vol. 28 of Computer
Aided Chemical Engineering. pp. 1159–1164.
Rossi, F., van Beek, P., Walsh, T. (Eds.), 2006. Handbook of Constraint Programming.
Elsevier.
Sahinidis, N. V., 1996. Baron: A general purpose global optimization software package.
Journal of Global Optimization 8 (2), 201–205.
Sahinidis, N. V., 2003. Global Optimization and Constraint Satisfaction. Vol. 2861 of
Lecture Notes in Computer Science. Springer, Ch. The Branch-and-Reduce Approach, pp.
1–16.
Schilling, G., Pantelides, C. C., 1996. A simple continuous-time process scheduling formu-
lation and a novel solution algorithm. Computers and Chemical Engineering 20 (Suppl. 2),
S1221–S1226.
Shah, N., 1996. Mathematical programming techniques for crude oil scheduling. Computers
and Chemical Engineering 20, 1227–1232.
Shah, N., Pantelides, C. C., Sargent, R. W. H., 1993. A general algorithm for short-
term scheduling of batch operations. 2. Computational issues. Computers and Chemical
Engineering 17 (2), 229–244.
Shaw, P., 1998. Using constraint programming and local search methods to solve ve-
hicle routing problems. In: 4th International Conference on Principles and Practice of
Constraint Programming. Vol. 1520 of Lecture Notes in Computer Science. Springer, pp.
417–431.

Tawarmalani, M., Sahinidis, N. V., 2004. Global optimization of mixed-integer nonlinear
programs: A theoretical and computational study. Mathematical Programming 99 (3),
563–591.
van Hentenryck, P., 1989. Constraint satisfaction in logic programming. The MIT Press.
van Hoeve, W.-J., Pesant, G., Rousseau, L. M., Sabharwal, A., 2009. New filtering algo-
rithms for combinations of among constraints. Constraints 14 (2), 273–292.
Wassick, J. M., 2009. Enterprise-wide optimization in an integrated chemical complex.

Computers and Chemical Engineering 33 (12), 1950–1963.
Waterer, H., Johnson, E. L., Nobili, P., Savelsbergh, M. W. P., 2002. The relation of time
indexed formulations of single machine scheduling problems to the node packing problem.
Mathematical Programming 93 (3), 477–494.
Wenkay, L., Hui, C., Hua, B., Tong, Z., 2002. Scheduling crude oil unloading, storage, and
processing. Industrial and Engineering Chemistry Research 41 (26), 6723–6734.
Westerlund, T., Pettersson, F., 1995. An extended cutting plane method for solving convex
minlp problems. Computers and Chemical Engineering 19 (Suppl. 1), S131–S136.
Wicaksono, D. S., Karimi, I. A., 2008. Piecewise milp under- and overestimators for global
optimization of bilinear programs. AIChE Journal 54 (4), 991–1008.
Wikipedia, 2010. Oil refinery. http://en.wikipedia.org/wiki/Oil_refinery.
Yunes, T., Aron, I. D., Hooker, J. N., 2010. An integrated solver for optimization problems.
Operations Research 58 (2), 342–356.
Zhang, J., Kim, N., Lasdon, L., 1985. An improved successive linear programming algo-
rithm. Management Science 31 (10), 1312–1331.
Zhang, X., Sargent, R. W. H., 1996. The optimal operation of mixed production facilities:
A general formulation and some approaches for the solution. Computers and Chemical
Engineering 20 (6-7), 897–904.

Appendices
178
Appendix A
On Tightness of Strengthened Constraints
We present mathematical results on the tightness of strengthened constraints.
• Constraint (2.17) is at least as tight as constraint (2.4). Indeed, assume constraint

(2.17) is satisfied for W 0 such that v1 , v2 ∈ W 0 :
X
Ziv1 + Ziv2 ≤ Ziv ≤ 1
v∈W 0
• Constraint (2.6) is at least as tight as constraints (2.5). Indeed, assume constraint

(2.6) is satisfied for v1 , v2 ∈ W :
Ei1 v1 ≤ Ei1 v1 + Ei1 v2
≤ Si2 v1 + Si2 v2 + H · (1 − Zi2 v1 − Zi2 v2 )
≤ Si2 v2 + H · (1 − Zi2 v2 ) + Si2 v1 − H · Zi2 v1
≤ Si2 v2 + H · (1 − Zi2 v2 )

(2.18) is satisfied for W 0 such that v1 , v2 ∈ W 0 :
X
Ei1 v1 + Ei1 v2 ≤ Ei1 v
v∈W 0
!
X X
≤ Si2 v + H · 1− Z i2 v
v∈W 0 v∈W 0
X
≤ Si2 v1 + Si2 v2 + H · (1 − Zi2 v1 − Zi2 v2 ) + (Si2 v − H · Zi2 v )
v∈W 0 \{v 1 ,v2 }
≤ Si2 v1 + Si2 v2 + H · (1 − Zi2 v1 − Zi2 v2 )
Appendix A. On Tightness of Strengthened Constraints 179

(2.19) is satisfied for W 0 :
X X X X X X
Ei1 v ≤ Ei1 v + Div ≤ Si2 v + H · (1 − Zi2 v )
v∈W 0 v∈W 0 i∈T v∈W 0 v∈W 0 v∈W 0
i1 <i<i2
• Constraint (2.21a) is at least as tight as constraint (2.13a). Indeed, assume constraint

(2.21a) is satisfied for W 0 such that v 0 ∈ W 0 :
X
Siv ≤ Siv0 ≤ ti
v 0 ∈W 0
• Constraint (2.21b) is at least as tight as constraint (2.13b). Indeed, assume constraint

(2.21b) is satisfied for W 0 such that v ∈ W 0 :
X
Siv ≥ Siv + (Siv0 − H · Ziv0 )
v 0 ∈W 0 \{v}
X X
≥ Siv0 − H · Ziv0
v 0 ∈W 0 v 0 ∈W 0 \{v}
X X
≥ ti − H · (1 − Ziv0 ) − H · Ziv0
v 0 ∈W 0 v 0 ∈W 0 \{v}
≥ ti − H · (1 − Ziv )

(2.26) is satisfied for sets of operations W1 and W2 such that v1 ∈ W1 and v2 ∈ W2 :
X
Ei1 v1 ≤ Ei1 v
v∈W1
X X
≤ Si2 v + H · (1 − Zi 2 v )
v∈W2 v∈W2
X
≤ Si2 v2 + H · (1 − Zi2 v2 ) + (Si2 v − H · Zi2 v )
v∈W2 \{v2 }
≤ Si2 v2 + H · (1 − Zi2 v2 )
Appendix A. On Tightness of Strengthened Constraints 180

Appendix B
Crude-Oil Operations Scheduling
Examples

1 3
4 7
5 8
2 6
Figure B.1: Refinery crude-oil scheduling system for COSP1.
Table B.1: Data for COSP1.

Vessel 1 0 100% A 1,000
Vessel 2 4 100% B 1,000
Tank 1 [0, 1,000] 100% A 250
Tank 2 [0, 1,000] 100% B 750
Tank 1 (mix X) [0, 1,000] 100% C 500
Tank 2 (mix Y) [0, 1,000] 100% D 500
Crudes Property 1 (sulfur concentration) Gross margin ($/bbl)
Crude A 0.01 9
Crude B 0.06 4
Crude C 0.02 8
Crude D 0.05 5
Crude mixtures Property 1 (sulfur concentration) Demand (Mbbl)
Crude mix X [0.015, 0.025] [1,000, 1,000]
Crude mix Y [0.045, 0.055] [1,000, 1,000]
Appendix B. Crude-Oil Operations Scheduling Examples 181

1 4
5 11
6 12
2 7
8 13
9 14
3 10
Figure B.2: Refinery crude-oil scheduling system for COSP2 and COSP3.

Vessel 1 0 100% A 1,000
Vessel 2 3 100% B 1,000
Vessel 3 6 100% C 1,000
Tank 1 [0, 1,000] 100% A 200
Tank 2 [0, 1,000] 100% B 500
Tank 3 [0, 1,000] 100% C 700
Tank 1 (mix X) [0, 1,000] 100% D 300
Tank 2 (mix Y) [0, 1,000] 100% E 500
Tank 3 (mix Z) [0, 1,000] 100% F 300
Crudes Property 1 Property 2 Gross margin ($/bbl)
Crude A 0.01 0.04 1
Crude B 0.03 0.02 3
Crude C 0.05 0.01 5
Crude D 0.0167 0.333 1.67
Crude E 0.03 0.23 3
Crude F 0.0433 0.133 4.33
Crude mixtures Property 1 Property 2 Demand (Mbbl)
Crude mix X [0.01, 0.02] [0.03, 0.038] [1,000, 1,000]
Crude mix Y [0.025, 0.035] [0.018, 0.027] [1,000, 1,000]
Crude mix Z [0.04, 0.048] [0.01, 0.018] [1,000, 1,000]

Vessel 1 0 100% A 500
Vessel 2 4 100% B 500
Vessel 3 8 100% C 500
Tank 1 [0, 1,000] 100% D 200
Tank 2 [0, 1,000] 100% E 200
Tank 3 [0, 1,000] 100% F 200
Tank 1 (mix X) [0, 1,000] 100% G 300
Tank 2 (mix Y) [0, 1,000] 100% E 500
Tank 3 (mix Z) [0, 1,000] 100% F 300
Crudes Property 1 Gross margin ($/bbl)
Crude A 0.01 1
Crude B 0.085 6
Crude C 0.06 8.5
Crude D 0.02 2
Crude E 0.05 5
Crude F 0.08 8
Crude G 0.03 3
Crude mixtures Property 1 Demand (Mbbl)
Crude mix X [0.025, 0.035] [500, 500]
Crude mix Y [0.045, 0.065] [500, 500]
Crude mix Z [0.075, 0.085] [500, 500]
1 5
14
6
15
2 7
8 16
9 17
3 10
18
11
19
12
13
Figure B.3: Refinery crude-oil scheduling system for COSP4.

Vessel 1 0 100% A 600
Vessel 2 5 100% B 600
Vessel 3 10 100% C 600
Tank 1 [100, 900] 100% D 600
Tank 2 [100, 1,100] 100% A 100
Tank 3 [100, 1,100] 100% B 500
Tank 4 [100, 1,100] 100% C 400
Tank 5 [100, 900] 100% E 300
Tank 6 [100, 900] 100% E 600
Tank 1 (mix X) [0, 800] 100% F 50
Tank 2 (mix Y) [0, 800] 100% G 300
Tank 3 (mix Z) [0, 800] 100% H 300
Tank 4 (mix W) [0, 800] 100% E 300
Crudes Property 1 Gross margin ($/bbl)
Crude A 0.03 3
Crude B 0.05 5
Crude C 0.065 6.5
Crude D 0.031 3.1
Crude E 0.075 7.5
Crude F 0.0317 3.17
Crude G 0.0483 4.83
Crude H 0.0633 6.33
Crude mixtures Property 1 Demand (Mbbl)
Crude mix X [0.03, 0.035] [600, 600]
Crude mix Y [0.043, 0.05] [600, 600]
Crude mix Z [0.06, 0.065] [600, 600]
Crude mix W [0.071, 0.08] [600, 600]

Appendix C
Mathematical Models for Crude-Oil
Operations Scheduling Problems
C.1 MOS Model
X X XX
max Gc · Vivc
s.t. Variable bound and time constraints (3.1)
Cardinality constraints (3.2)
Precedence constraints (3.3)
Continuous distillation constraints (3.4)
Variable constraints (3.5)
Operation constraints (3.6)
Resource constraints (3.8)
Demand constraint (3.9)
Clique-based assignment constraint (3.10)
Clique-based non-overlapping constraint (3.11)

X
Ziv ≤ Z(i−1)v0 i ∈ T, i 6= 1, v ∈ W
0
v ∈W
N Ovv0 =1
X
Ziv ≥ 1 i∈T
v∈W
Siv , Div , Eiv , Vivt , Vivc , Ltir , Lirc ≥ 0 i ∈ T, v ∈ W, c ∈ C, r ∈ R
Ziv ∈ {0, 1} i ∈ T, v ∈ W
Appendix C. Mathematical Models for Crude-Oil Scheduling Problems 185

C.2 MOS-SST Model
C.2 MOS-SST Model
X X XX
max Gc · Vivc
ti−1 ≤ ti i ∈ T, i 6= 1
X
Siv ≤ ti i ∈ T, W 0 ∈ clique(GN O )
v∈W 0
X X
Siv ≥ ti − H · (1 − Ziv ) i ∈ T, W 0 ∈ clique(GN O )
v∈W 0 v∈W 0
X
Ziv ≥ 1 i∈T
v∈W
Ziv ∈ {0, 1} i ∈ T, v ∈ W
ti ∈ [0, H] i∈T

C.3 MOS-FST Model
C.3 MOS-FST Model
X X XX
max Gc · Vivc
Ziv ∈ {0, 1} i ∈ T, v ∈ W
i−1
ti = ·H i∈T
n

C.4 SOS Model
C.4 SOS Model
X X XX
max Gc · Vivc
Clique-based non-overlapping constraints (4.1)
Biclique-based non-overlapping constraints (4.2)
Symmetry-breaking constraints (4.3)

X
Ziv ≥ 1 i∈T
v∈W
Ziv ∈ {0, 1} i ∈ T, v ∈ W

Appendix D
Mathematical Model for the Refinery
Planning Problem
In this section, the ANN model developed in Gueddar and Dua (2010) for CDU simulations
is presented. It is based on a layered directed graph which represents the model calculations
(see Fig. D.1). Each node in the input/output layers correspond to one input/output
variable. Each node j = 1, . . . , Nn in the intermediate layer l = 1, . . . , Nl corresponds to an
activation variable alj and a transformed variable hlj . The activation variables are calculated
from the transformed variables of the previous layer using an affine expression while the
transformed variables are calculated by applying the hyperbolic tangent to their associated
activation variable. The ANN equations are expressed as follows.
Nx
X
a1j = 1
wji xi + b1j j = 1, . . . , Nn (D.1)
i=1
hlj = tanh alj l = 1, . . . , Nh , j = 1, . . . , Nn (D.2)

Nn
X
l l−1
alj = wji hi + blj l = 2, . . . , Nh , j = 1, . . . , Nn (D.3)
i=1
Nn
X
uk = Wki hN
i + Bk
h
k = 1, . . . , No (D.4)
i=1
The model uses the following parameters:
• Nx is the number of inputs

• No is the number of outputs
• Nh is the number of intermediate layers
Appendix D. Mathematical Model for the Refinery Planning Problem 189

x1
x2 u1
x3 u2
Output layer
x4 Intermediate layers
Input layer
Figure D.1: Layered artificial neural network.
• Nn is the number of nodes in each intermediate layer

• wjil , blj , Wki , Bk are parameters specific to the ANN
Dua (2010) demonstrates how to tune the ANN parameters by minimizing the total pre-
diction error as well as the ANN complexity. This tuning step is performed by solving
a training MINLP. We denote CDUANN(x, u) the set of constraints defining the relation
between the ANN inputs x and outputs u. In particular, the model inputs include crude
ip
properties qC and CDU cut points τ k (k ∈ {naphta, kerosene, diesel}), and the model out-
puts include cut yields αijk and crude cut properties q1ijkp . All crude property inputs are
fixed while CDU cut points are variable. The cut point between diesel and residue cuts can
take three discrete values defining the three CDU operating modes. All the outputs are
variable. The full refinery planning model is expressed as follows.

X
max pl xlS (sales revenue maximization)
l∈L
xij
X
i
s.t. 0≤ F ≤C i∈I (crude availability)
j∈J
τk ≤ τk ≤ τk k∈K (CDU cut point limits)

n o n o
ip k
CDUANN qC , τ , αijk , q1ijkp (CDU model)
X X ij
FR · H ≤ xF ≤ F R · H (CDU flowrate limitations)
i∈I j∈J
xijk
1 =α
ijk
· xij
F (i, j, k) ∈ I × J × K (CDU yield calculation)
X ijk X jkl
x1 = x2 (j, k) ∈ J × K (pool mass balance)
i∈I l∈L
(pool quality balance)

q1ijkp xijk jkp
xjkl
X X
1 = q2 2 (j, k, p) ∈ J × K × P
i∈I l∈L nonlinear
xjkl
XX
l
2 = xS l∈L (product mass balance)
j∈J k∈K
xlS ≤ Dl l∈L (maximum product demand)
(product quality requirement)

q2jkp xjkl
XX
lp l
2 ≤ Z xS (l, p) ∈ L × P
j∈J k∈K nonlinear
xij ijk jkl l

F , x1 , x2 , xS , α
ijk
≥0
q1ijkp , q2jkp , τ k ∈ R

Optimal Scheduling of Refinery Crude-Oil Operations

Uploaded by

Copyright:

Available Formats

Optimal Scheduling of Refinery Crude-Oil Operations

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Optimal Scheduling of Refinery Crude-Oil Operations

Uploaded by

Copyright:

Available Formats

Carnegie Mellon University

Optimal Scheduling of Refinery Crude-Oil

Follow this and additional works at: http://repository.cmu.edu/dissertations

in Partial Fulfillment of the Requirements

for the degree of

Carnegie Mellon University

List of Figures xii

2 Time Representations and Mathematical Models for Process Scheduling

3 Short-Term Scheduling of Crude-Oil Operations 67

4 Single-Operation Sequencing Model for Crude-Oil Operations Scheduling 92

5 Tightening the Linear Relaxation of a Crude-Oil Operations Scheduling

6 Integration of Refinery Planning and Crude-Oil Scheduling using La-

A On Tightness of Strengthened Constraints 179

B Crude-Oil Operations Scheduling Examples 181

C Mathematical Models for Crude-Oil Operations Scheduling Problems 185

D Mathematical Model for the Refinery Planning Problem 189

1.1 Optimization techniques used in different MINLP solvers. . . . . . . . . . . 12

2.1 Resource requirements for case study. . . . . . . . . . . . . . . . . . . . . . 22

3.1 Data for COSP1. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71

4.1 Maximal cliques and bicliques for COSP2 and COSP3. . . . . . . . . . . . . 93

5.1 Cost data for crude-oil operations scheduling problems. . . . . . . . . . . . 113

6.1 Crude-oil scheduling data for case study. . . . . . . . . . . . . . . . . . . . . 126

B.1 Data for COSP1. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 181

2.1 Four steps optimization approach. . . . . . . . . . . . . . . . . . . . . . . . 20

3.1 Example of tank schedule. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70

4.1 Symmetric sequences of operations for COSP1. . . . . . . . . . . . . . . . . 96

5.1 Branch & cut algorithm with McCormick cuts. . . . . . . . . . . . . . . . . 116

List of Figures xii

B.1 Refinery crude-oil scheduling system for COSP1. . . . . . . . . . . . . . . . 181

D.1 Layered artificial neural network. . . . . . . . . . . . . . . . . . . . . . . . . 190

List of Figures xiii

1. Develop a unified modeling approach for solving process scheduling problems

In this chapter, an overview of single-stage and multi-stage batch scheduling problems

1.1 Single-Stage and Multi-Stage Batch Scheduling

and often require additional big-M constraints.

1.2 Optimization of Oil Refineries

1. Refinery planning (see section 1.2.1)

- for LPG, see Pinto and Moro (2000)

1.2.1 Refinery Planning

1.2.2 Crude-Oil Operations Scheduling

1. Maximization of crude gross margins (chapters 3 and 4)

Shah (1996) proposed to use mathematical programming techniques to find crude-oil

Crude Vessels Storage Tanks Charging Tanks CDU

Figure 1.4: Crude-oil scheduling problem 1 from Lee et al. (1996).

1.3 Mixed-Integer Optimization Tools

1. mixed-integer linear programming (MILP)

1.3.1 Mixed-Integer Linear Programming

Table 1.1: Optimization techniques used in different MINLP solvers.

1.3.2 Mixed-Integer Nonlinear Programming

A number of industrial applications of mixed-integer optimization include nonlinearities

1. NLP-based branch-and-bound (Leyffer, 2001)

1. branch-and-reduce (Tawarmalani and Sahinidis, 2004)

1.3.3 Constraint Programming

Constraint programming is an alternative optimization approach to classical Operations

1.3.4 Lagrangian Relaxation

Lagrangian relaxation is a relaxation technique which aims at solving mathematical models

complicating constraints), this transformation can be described as follows:

1. subgradient method (Held and Karp, 1971; Fisher, 1981)

3. boxstep method (Marsten et al., 1975)

As explained in Frangioni (2005), problem (3) is equivalent to the following convexified