Accepted Manuscript: Computers & Industrial Engineering
Accepted Manuscript: Computers & Industrial Engineering
Accepted Manuscript: Computers & Industrial Engineering
PII: S0360-8352(17)30437-0
DOI: http://dx.doi.org/10.1016/j.cie.2017.09.020
Reference: CAIE 4906
Please cite this article as: Salehi, M., Jalalian, M., Siar, M.M.V., Green transportation scheduling with speed control:
trade-off between total transportation cost and carbon emission, Computers & Industrial Engineering (2017), doi:
http://dx.doi.org/10.1016/j.cie.2017.09.020
This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers
we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and
review of the resulting proof before it is published in its final form. Please note that during the production process
errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.
Green transportation scheduling with speed control: trade-off
between total transportation cost and carbon emission
1
Department of Industrial Engineering, Payame Noor University, Tehran, Iran
2
Department of Industrial Engineering, K. N. Toosi University of Technology, Tehran, Iran.
*
1
Corresponding author.
Email addresses: M_salehi61@yahoo.com (M. Salehi), Mahdi.Jalalian@email.kntu.ac.ir (M. Jalalian),
mvalisiar@email.kntu.ac.ir (M.M. Vali Siar)
1
Green transportation scheduling with speed control: trade-off
between total transportation cost and carbon emission
Abstract
Transportation has a significant portion in greenhouse gas emission. Efficient transportation
scheduling can decrease transportation costs and environmental damages by reducing fuel
consumption and greenhouse gas emission. These advantages have made green transportation
scheduling more attractive. This paper addresses the problem of integrated green truck
transportation scheduling and driver assignment. We propose a bi-objective mixed integer
nonlinear programming (MINLP) model to minimize total transportation-related costs (TTC)
and total carbon emission (TCE). We consider TTC and TCE as manufacturer measure and
environmental sustainability measure respectively. There is a contradiction between two
objectives. We consider the ability of speed control for trucks to create the trade-off between
TTC and TCE. A linearization technique is utilized to reduce the complexity of the proposed
model. The Augmented ε-constraint method is used to solve the model. Also, we develop a
constructive heuristic (CH) approach to get high-quality solutions in an efficient CPU time.
Some scenarios are applied to evaluate the model and heuristic approach. The results show
the model achieves high-quality solutions in reasonable time. Also, the model and heuristic
have a better performance compared to scenarios. Statistical tests are done to approve the
comparisons.
1. Introduction
Transportation has a great effect on economic growth and is a foundational infrastructure in
any country. Efficient transportation scheduling improves logistics performance and supply
chain and increases the competitive advantage of companies (Guo et al. 2016). People and
goods transportation is one of the main reasons for increasing greenhouse gas emission. It
leads to climate change and increase of global warming (Fuglestvedt et al., 2008). In 2010,
transportation sector emitted about 15% of global greenhouse gas which the road
transportation accounted for 10.5% of this amount (ECOFYS, 2010). It is forecasted that the
transportation will be responsible for 30-50% of CO2 emissions by 2050 (Nakicenovic and
Swart, 2000). This amount of emission pollutes the atmosphere and has great damaging
effects on the environment. Therefore, some policies are needed to turn mentioned
disadvantages to the opportunities. We propose a new approach to decrease the damaging
effect and get some advantages without any investment. This paper aims to address a multi-
period integrated truck transportation scheduling and driver assignment to minimize CO2
emissions and transportation costs.
There are a large variety of papers in the context of transportation scheduling. Topics include
truck transportation scheduling (Zhang et al., 2010; El Hachemi et al, 2013), train
transportation scheduling (Shafia et al., 2012; Wang and Yun 2013), block transportation
scheduling (Lee et al., 2009; Roh and Cha 2011, Joo and Kim, 2014), maritime transportation
scheduling (Huang and Karimi, 2006, Nishi and Izuno, 2014; Siddiqui and Verma, 2015;
2
Hennig et al., 2015) and Integrated transportation scheduling with production planning and
supply chain management (Xu et al. 2017 Karaoğlan and Kesen, 2017)
Negative effects of transportation on the environment have led researchers to focus on green
transportation scheduling in recent years. Bauer et al. (2010) considered minimizing
greenhouse gas emissions in an intermodal freight transportation problem by services
scheduling and freight routing. They proposed an integer programming formulation for the
problem. Li et al. (2013) studied green train scheduling problem. They proposed a mixed
integer programming model to minimize carbon and energy emission and total passengers’
time. They used fuzzy multi objective optimization algorithm for solving the model.
Kontovas (2014) conceptualized formulate for the green ship routing and scheduling
problem. They combined ship air emissions and the vehicle routing problem. Huang et al.
(2016) presented a bi-objective mathematical model to optimize the timetables of urban rail
transit systems and operational energy consumption. They used genetic algorithm for solving
the problem. Tanimizu and Amano (2016) proposed an integrated scheduling method for
production and transportation problem. They presented a bi-objective mathematical model to
minimize total tardiness and emitted carbon of vehicles. Guo et al. (2016) focused on green
transportation scheduling problem considering pickup time and transport mode selection.
They proposed a bi-objective integer nonlinear programming model and developed a
memetic optimization approach to solve the model. Guo et al. (2017) studied an integrated
production and transportation scheduling problem. They presented a mathematical model to
minimize total production costs and transportation costs. The objective function includes
carbon emission cost caused by production and transportation. They developed a harmony
search-based memetic algorithm for solving the problem.
We develop a bi-objective mixed integer nonlinear programming (MINLP) model for green
transportation scheduling problem which exists in some make-to-order manufacturing
companies. The first objective includes minimizing total transportation related-cost as a
measure of manufacturing company. TTC includes transport cost, earliness and tardiness cost
for delivery and holding cost in the manufacturing company. The second objective includes
minimizing total carbon emission as a measure of environmental sustainability.
Manufacturing companies have two main resources for transporting the goods, including
vehicles and drivers. The attention of drivers' assignment status besides trucks consideration
reveals a more comprehensive view on the problem and facilitates the company’s operations.
It is difficult for companies to assign drivers manually (Sargut et al., 2017). If number of
orders is high in the planning period, then more trucks and drivers are needed while the
number of drivers and trucks are limited. Also, there is eligibility requirment which is related
to the drivers’ license i.e. not all drivers can drive any truck. In this situation, the problem
will be more difficult. Therefore, we integrate the driver assignment with truck transportation
scheduling problem.
There is a conflict between the objectives. We propose new approach to create a balance
between them by the truck speed control. Speed control affects the transportation costs and
carbon emission, i.e. different speed levels are assumed for vehicles and each level leads to a
different TTC and TCE values. The augmented ε-constraint method is applied to solve the
proposed bi-objective model. Also, we develop an efficient constructive heuristic to solve the
large size problems in suitable times. Table 1 summarizes the research papers on green
transportation scheduling. To the best of our knowledge, the proposed bi-objective model
integrating the transportation scheduling and drivers' assignment considering truck speed
control has not been studied in the literature.
3
Table1. Summary of the literature on green transportation scheduling
The rest of this paper is organized as follows. Section 2 describes the considered problem and
explains the proposed mathematical model. Section 3 presents the constructive heuristic and
briefly explains the augmented ε-constraint method. Section 4 presents computational results.
Section 5 finishes the paper and brings up some offers for future works.
2. Problem description
Consider a transportation scheduling problem existing in a manufacturing company. When
the manufacturing of products is finished, then the company transports the requested quantity
of orders to a distribution center. Finally, the orders can be delivered to customers by other
companies. The quantity of transported orders is determined based on the customers’
demands.
Each order has a delivery date. It is better to deliver orders exactly on the delivery date. If
orders reach customers before the delivery date, then they may not require the orders at that
time. If orders reach the customers after the delivery date, then they will be dissatisfied and
4
customer loyalty may decrease. So, we use earliness and tardiness penalty costs for orders.
There is a holding cost for a finished product which is stored in the company. A certain
number of different trucks is available to the company. The trucks differ in terms of capacity
and their properties. Each truck has a specific weight and carbon emission factor depending
on its type. There are some possible speed levels for trucks. Each level has a certain amount
of fuel consumption, carbon emission and duration of transportation. Also, a certain number
of drivers is available to the company. Not all drivers can drive any truck. The pickup date of
all orders should be determined at the beginning of the planning horizon. An order can be
transported by two or more trucks. But all of its demand should be loaded and transported on
a single fixed day. In summary, after the company specified the pickup date for an order
based on the delivery date, then the trucks and corresponding drivers should be determined to
transport the orders.
An MINLP model for a multi-period trucks transportation scheduling problem is proposed to
minimize total transportation costs and total carbon emission. We use a linearization
technique to mitigate the difficulty of the problem. The MILP model is the basis of
computations. In the following, at first the MINLP programming has been explained and then
the linearization is done.
5
dem o Demand of order o
Decision Variables
Qoctd =1, if order o is picked up at planning day d by truck t with driver c and is 0 otherwise
v od =1, if all demands of order o is picked up at planning day d and is 0 otherwise
y otl =1, if order o is carried by truck t with speed level l and is 0 otherwise
z ooctd =0, if order o and are picked up at planning day d by truck t with driver c
simultaneously and is 1 otherwise
xooctd =2, if order o and are picked up at planning day d by truck t with driver c
simultaneously, and is 1 if just order o or is picked up else is 0 otherwise
6
t
min yotl (( wt t SPtl2 ) dis awt dis ) (2)
oO tT lL efc t
Subject to.
yotl 1 o O, t T (15)
lL
dis
DTot STot yotl o O, t T , l L (18)
SPtl
7
potd , awt , STot , DTot 0 o O, t T , d D (21)
Objective functions are presented in (1) and (2). Objective (1) represents the minimizing total
transportation related-costs and objective (2) represents total carbon emission. Objective (1)
contains four parts. It includes transport cost, holding cost in the companies and earliness and
tardiness cost for delivery time. The first objective tries to deliver the orders on time with
lowest carry cost. To do this, the planners can determine the appropriate speed level for
trucks. In case that the delivery schedule is before the delivery date, more tardiness and
holding cost can be avoided by choosing fast speed level. In another case, increasing the
speed levels causes the orders to be delivered earlier than the delivery date and thus the
earliness cost will be paid. Controlling speed levels as a decision variable enables the model
to prevent above costs. Objective (2) determines the trucks speed. Each speed level causes a
certain amount of fuel consumption which affects the CO2 emissions. Objective (1) and (2)
are based on the study of Guo et al. (2016). We change the objective functions of this study
which eventually leads to objectives (1) and (2). The changes include considering variable
speed for trucks and its efficiency that affects CO2 emission. The model is capable for
maximum usage of trucks with high efficiency and minimum usage of trucks with low
efficiency. Truck efficiency is dependent on the engine and structural efficiency which is
related to different factors such as the duration of its use.
Constraint (3) ensures that all orders should be carried in the planning period. Each order can
be loaded when all its demand is produced in manufacturing company. This is done by
constraint (4). Constraint (5) guarantees that each order is loaded by trucks just in one unique
day. The orders can be carried by one or more trucks by constraint (6). Constraint (7)
stipulates that each truck and its allocated driver should not be changed until the delivery of
the order. Also, each driver can drive some trucks due to his/her driving certificate.
Constraint (8) determines the orders carried by same truck simultaneously. constraint (9),
(10) and (11) are used to convert the to which is a binary variable. Constraint
(12) limits the amount of orders based on the trucks capacity. Constraint (13) ensures that all
demands must be responded. Constraint (14) computes the actual weight of the loads carried
by each truck. Constraint (15) ensures only one speed level is chosen for each truck to carry
the orders. Constraint (16) puts limitation on speed level. Constraint (17) determines the
pickup date of each order by each truck. Constraint (18) computes the delivery time of orders
based on the truck speed level. There is a limitation for orders that are not carried
simultaneously by same truck. So, constraint (19) suggests the relation between the loading
of next order and the delivery of previous order. Finally, constraint (20) and (21) are logical
constraints.
8
STot DTzooctd o, o(oo) O, c Ct , t T , d D (22)
The MINLP model is converted to a mixed integer linear programming (MILP) model
considering above constraints. The MILP model is as follows:
min efct y otl (( wt t SPtl2 ) dis awt dis ) (2)
oO tT lL t
Subject to.
(3)-(12), (15)-(18), (20)-(31)
There is a conflict between the objectives. We use truck speed control to create a balance
between the objectives. To have a better sight of this subject, we generated a problem with 5
orders, 2 drivers and 2 trucks with different speed levels. The planning horizon includes 3
days. The problem data is provided in Table 2.
9
Table2. Generated problem
orders trucks driver
number quantity production finish Due date number capacity number License for
time trucks
1 1120 Day 2 3
1 1100 1 1 and 2
2 800 Day 2 3
3 1000 Day 1 2
2 950 2 1 and 2
4 350 Day 1 2
5 250 Day 1 2
We solved the above problem via the augmented ε-constraint method based on the MILP model.
The model suggests different Pareto solutions which are shown in Figure1.
Each Pareto solution has its own planning with ceratin scheduling and truck-driver
assignment. In Pareto solution A, all trucks carry orders at maximum speed levels. In this
situation, fast transportation prevents the delays and inventory accumulation in the
manufacturing company. On the other hand, high speed level may cause earliness. Therefore,
considering the tardiness, earliness and holding cost put a limitation in highest speed level for
trucks. The best results of TTC is obtained in Pareto solution. Also, the energy consumption
and CO2 emissions are in their worst situation. Whatever get closer to Pareto solution B, the
lower level of TCE is reached. This is due to the use of trucks with lower speed level. In
Pareto solution B, all trucks carry orders with minimum speed levels. This leads to the lowest
CO2 emissions. In this case, the delay time and inventory levels are at their highest values.
Different factors affect the choice of a Pareto solution among all obtained Pareto solutions.
Company policies, management perspective, time limit for orders and demand volume are
some of these factors. For example, a management policy for a situation which a large
number of orders should be responded in a short planning period is different from a situation
which the same amount of orders can be responded in a longer planning period.
We considered Pareto solution C because both TTC and TCE are at an acceptable level in
normal situation. The obtained planning for Pareto solution C is shown in Table 3 and Figure
2.
Table3. Obtained planning for Pareto solution C
10
3. Solution method
Single-objective problems can reach optimal solutions. But there is no absolute optimal
solution which optimizes the whole objectives simultaneously in multi-objective problems.
So different techniques are used to get high quality solutions. Some of them such as
weighted-sum method and -constraint method have become more interesting for
researchers. -constraint method has some advantages in comparison with the weighted-sum
method. These advantages include generating non-extreme solutions in addition to extreme
one, achieving unique Pareto solution in each run, controlling the number of Pareto solutions
via grid points and no need to scale objectives. For more details, the readers are referred to
Mavrotas (2009) study. Besides these advantages, Mavrotas (2009) improved -constraint
method and provided an augmented -constraint method with ability to find the high-quality
solutions.
In this paper, augmented -constraint method has been used to solve different problems . We
briefly explain the implementation of this method for discussed problem.
1) The range of each objective should be calculated. So, it is necessary to find the best and
worst value of each objective. -constraint method suggests using payoff table. In this
table, each objective is optimized without considering the other objectives. So the best
value of each objective is calculated. Then the maximum (minimum in maximization
objectives) value for each objective is approximated as the worst (nadir) value. Mavrotas
(2009) has pointed out the obtained solutions are not guaranteed Pareto optimal solutions.
Therefore, he proposed using lexicographic optimization for objective functions. In this
method, at first, the objective (1) is optimized without considering the second objective.
So the min is obtained. Then the second objective is optimized considering the
value of as a constraint. It causes the second objective to be optimized ( )
while the optimal value of first the objective is kept. Afterward, the objective (2) is
optimized without considering the first objective. So min is obtained. Then the
first objective is optimized considering the value of as a constraint. It causes the
first objective to be optimized ( ) while the optimal value of the second objective
is kept. Finally, the best (ideal) and worst (nadir) value of objective (1) is obtained based
on the and . Also the best and worst value of objective (2) is obtained based on the
and .
2) The range of each objective is divided into equal intervals based on the number of grid
points.
3) One objective is considered as the main objective and the other is used in constraints.
4) Each time the problem paug is solved to reach unique Pareto optimal (non-dominated)
solution.
Problem Paug:
s
min f1 ( 2 )
r2
Subject to.
f 2 s2 e2 (32)
(3)-(12), (15)-(18), (20)-(31).
11
We consider objective (1) as the main objective. . Where, ub2 and r2 are
the upper bound and the range of objective (2) respectively. is a very small number
( to ). s2 and itr are non-negative surplus variable and the iteration number
respectively. We refer the interested readers to the study of Mavrotas (2009) for getting a
detailed explanation.
The complexity of transportation scheduling has led problems to be hard to be solved with
exact methods (Guo et al., 2016). For our proposed model, the complexity of the problem
with different sizes is shown in Figure 3. The trend of results indicates computational time
drastically increases by increasing the size of the problems. Therefore, a method which
obtains high-quality solutions in a reasonable time is needed to overcome the complexity.
The problem with O orders, T trucks and C drivers is denoted as O×T×C.
Constructive heuristic
We propose a constructive heuristic which named CH in this paper. The heuristic algorithm
includes three main steps:
1) A matrix containing different truck speed levels is defined. In this matrix one speed level
is assigned to each truck. So we know the speed of each truck while we do not know
which order carry by which truck. To find the trucks assignments and the related
schedules, the MILP model is transformed to the problem Pfix.
2) The mentioned matrix is considered as a parameter for MILP model. Unlike the presented
model in the previous section, Pfix is a single single-objective model which the speed
levels l ( ) are considered as parameter. So by defining the new parameter which
is , the fixed speed single MILP model (Pfix) is as bellow;
Pfix ;
min ( (r Q
oO cC t t T d D
t octd ) (ST
oO tT
ot f o ) ho max(0, due DT
oO t T
o ot ) eo
(1)
max(0, DT
oO t T
ot dueo ) to
Subject to.
(3)-(12), (15)-(18), (20)-(31)
3) As can be seen, the above model is a single-objective problem where the speeds of trucks
are fixed. Based on truck speed, the model assigns trucks to orders and defines
scheduling. Based on the assignment, total carbon emission is calculated via objective (2).
The pseudo code of the constructive heuristic algorithm is described in Figure 4.
Figure 4. Constructive heuristic algorithm
CH algorithm starts with the maximum speed of trucks. Based on the speed, an optimal
assignment and scheduling obtained by Pfix. So the value of objective (2) can be calculated. In
each iteration, the truck speeds are reduced and new assignment and scheduling are obtained.
Therefore, different Pareto solutions can be obtained which are different in scheduling,
12
assignment and trucks speeds. Some Pareto solutions may be dominated or weakly non-
dominated solutions. So we use the fast non-dominated sorting approach proposed by Deb et
al. (2002) in step 3 to find high quality non-dominated Pareto solutions. The procedure of the
approach is described in Figure 5. In this Figure, nb is the number of solutions that dominate
the solution b. Also S b is a set of solutions dominated by the solution b.
Both Cov (A,B) and Cov (B,A) should be calculated for comparing A and B.
Distance metrics (Czyzżak and Jaszkiewicz,1998): These performance measures represent
the quality of a solution set that is relevant to a reference set. These metrics can be computed
as follows;
1
Dist1 {min {d ( x A , xRe )}
card {Re} xReRe x AA
(34)
where A and Re are the solution set of algorithm and the reference set respectively. The
reference set is composed of all Pareto optimal solutions or the high-quality non-dominated
ones. and are the range of the - objective
j.
13
The first metric shows the average distance from to the nearest solution in A and the
second determines the worst case. The lower values show higher quality in solutions.
Spacing metric (Tan et al., 2006): This metric shows how evenly the obtained solutions are
distributed along the Pareto solution. This metric is calculated as follows;
1
1 card{ A}
2
SPC (d E d ) 2 /d (36)
card{ A} i 1
Where is the Euclidean distance (in the objective space) between solution i and its closest
solution. Smaller values of SPC mean that the solutions are distributed more evenly along the
relevant Pareto solution.
Cardinality of Pareto solutions (CRD): This metric includes the number of obtained Pareto
solutions
4.2 Comparison of augmented ε-constraint method, CH and scenarios (for small sizes)
We generated 8 instances to compare the augmented ε-constraint method and CH algorithm.
Figure 6 shows the comparison based on CPU time.
According to Figure 6, the number of orders, trucks and drivers have a substantial effect on
computational time. In detail, by increasing the size of the problems, the CPU time increases
because of the computational complexity growth. Increasing rate for CPU time in CH
algorithm is far less than the augmented ε-constraint method. The exponential growth for
CPU time is visible in the augmented ε-constraint method. Based on this measure, CH
algorithm outperforms the augmented ε-constraint method because it consumes much less
time for computations.
The comparison of the augmented ε-constraint method and CH algorithm based on the other
performance metrics is reported in Table 4 and Figure 7.
Table 4. Comparing aug.eps and CH algorithm based on performance metrics
Augmented ε-constraint CH
Prob. O×T×D COV COV
Dist1 Dist2 SPC CRD Dist1 Dist2 SPC CRD
(Aug. ε-CH) (CH-Aug. ε)
1 5×2×2 1.00 0.00 0.00 0.01 3 0.00 0.10 0.12 0.01 5
2 5×4×4 0.45 0.07 0.09 0.49 3 0.53 0.05 0.08 0.06 6
3 8×2×2 1.00 0.00 0.00 1.56 4 0.00 0.20 0.22 0.43 8
4 8×4×4 0.53 0.14 0.22 0.86 5 0.44 0.05 0.27 0.29 8
5 10×2×2 0.60 0.01 0.04 0.43 6 0.33 0.10 0.14 0.51 10
6 10×4×4 0.67 0.07 0.35 0.39 6 0.85 0.01 0.03 0.08 11
7 12×2×2 0.56 0.09 0.08 0.21 9 0.23 0.11 0.13 0.23 14
8 12×4×4 0.14 0.10 0.45 0.43 8 0.27 0.03 0.06 0.08 15
Average 0.62 0.06 0.15 0.55 5.5 0.33 0.08 0.13 0.21 9.63
14
Figure 7. Comparing aug.eps and CH based performance metrics
The results in Table 4 and Figure 7 present the comparison of the augmented ε-constraint
method and CH algorithm based on the COV, Dist1, Dist2, SPC and CRD. the augmented ε-
constraint method is a known method to reach Pareto optimal solutions. CH has better
performance based on CPU time, SPC and CRD metrics compared with the augmented ε-
constraint method. This means the proposed heuristic has obtained more solutions which are
distributed more evenly along the frontier in less time. More solutions can give managers
more flexibility to choose a solution. In this situation, managers can survey more solutions
with more precise consideration. Besides to CRD, the distribution of solutions is important.
The ideal situation happens when more solutions scatter in a wider scope of the feasible
space. It means more non-dominated solutions in a wide range can give managers more
confidence to choose one solution. On the other hand, if a significant number of solutions are
located in a restricted area then the flexibility to make decisions will be limited.
In some cases, CH has better performance based on COV and distance metrics compared
with the augmented ε-constraint method. In other cases, the augmented ε-constraint method
has obtained better solutions by consuming more CPU time in terms of coverage and
distances metrics.
For more survey, we compared the model and CH algorithm with defined scenarios.
augmented ε-constraint method and CH algorithm reach to different solutions while the
scenarios obtain only one solution. In order to make the comparison possible, we used
TOPSIS (Technique for Order of Preference by Similarity to Ideal Solution) procedure to
find an ideal Pareto solution from all obtained Pareto solutions. TOPSIS procedure proposed
by Hwang and Yoon (1981). This procedure is adapted to our proposed problem and is
represented in Figure 8.
Figure 8. TOPSIS procedure
We consider equal weight for objective (1) and (2). TOPSIS procedure suggests a solution
which has the lowest distance from the ideal solution (A+) and the greatest distance from the
nadir solution (A-) among all of the Pareto solutions. The Pareto solution proposed by
TOPSIS for the augmented ε-constraint method and CH algorithm is compared with obtained
result by each scenario. The comparison is reported in Table 5.
Table 5. Comparing the Euclidean distance from ideal and nadir solution in small size problems
Rk
Prob. O×T×D
k=minimum k=maximum k=average k=random k =Aug. ε k =CH
15
7 12×2×2 0.51 0.42 0.62 0.54 0.94 0.91
8 12×4×4 0.58 0.47 0.64 0.56 0.88 0.89
Average 0.59 0.52 0.74 0.65 0.92 0.91
Both augmented ε-constraint method and CH algorithm are superior to the scenarios based on
the Euclidean distances from the ideal and nadir solution. The model and heuristic have a
minimum distance from the ideal solution and maximum distance from the nadir solution.
Also, there is a little difference between augmented ε-constraint method and CH algorithm.
So, we compared the performance of the augmented ε-constraint and CH algorithm based on
performance metrics including COV, dist1, dist2, SPC, CRD and Rk to make the comparisons
statistically convincing.
The obtained results of metrics were normally distributed and the paired t-test was applied.
For each hypothesis test, the p-value result is reported in Table 6. The significance level was
assumed 0.05. It means if the result is smaller than 0.05 then the difference between the
augmented ε-constraint and CH is statistically significant.
Table 6. Paired-sample t-test for comparing aug.eps and CH based on performance metrics (p-value)
Performance metric aug.eps - CH
Dist1 0.57
Dist2 0.77
SPC 0.04
CRD 0.00
COV 0.13
Ck 0.35
Based on the above results, the difference between the augmented ε-constraint method and
CH algorithm is statistically significant based on the CRD and SPC metrics. CH has
performed better than augmented ε-constraint method based on these metrics. Also, there is
no significant difference between these two methods based on the Dist1, Dist2, COV and Ck.
16
The augmented ε-constraint method failed to reach solutions in large size problems.
Therefore, we studied the performance of CH algorithm via metrics including CRD and CPU
time. Figure 9 shows the trend of these measures on the large size problems.
Figure 9. Trend of CPU time and CRD for large size problems
Figure 9 shows as the size of the problem increases, more computational time is spent. Also,
this figure shows the size of problems affects the number of obtained solutions. Based on
these two metrics, the CH algorithm is able to reach numerous non-dominated solutions in an
effective time.
We compared the CH algorithm with scenarios to evaluated the heuristic in large size
problems. The results are given in Table 7.
Table 7. Comparing the Euclidean distance from ideal and nadir solution in large size problems
Ck
Prob. O×T×D
k=minimum k=maximum k=average k=random k =CH
The results show CH algorithm has a better performance compared to the scenarios. The
effect of different factors on the transportation costs and CO2 emissions are discussed in next
section.
Above results show the augmented ε-constraint method and CH algorithm improved the
objectives compared to the scenarios. The amount of changes in the first and second objective
values are investigated in Table 8 and 9. and denote the amount of
changes in method compared to the scenario for total transportation costs and total
carbon emission, respectively. The results are reported as a percentage and calculated as
follows;
TTCK TTCK
CTTCK , K 100 k , k K (42)
TTCK
17
TCE K TCE K
CTCE K , K 100 k , k K (43)
TCE K
Table 8. Percentage changes in TTC
Prob. O×T×D
=rand
=rand
=max
=max
=min
=min
=avg
=avg
1 5×2×2 450.48 2.96 113.40 229.93 446.00 2.17 111.65 227.24
2 5×4×4 529.80 1.165 139.76 150.81 529.34 1.14 139.60 150.63
3 8×2×2 106.90 1.05 15.73 16.15 89.14 -0.51 8.23 10.17
4 8×4×4 84..86 -4.54 14.32 8.73 78.80 -7.68 20.14 5.17
5 10×2×2 138.89 -13.86 14.64 2.24 132.40 -16.20 11.53 10.62
6 10×4×4 138.56 -13.98 14.50 2.11 134.37 -15.49 12.48 0.32
7 12×2×2 81.13 -13.02 12.40 7.37 71.91 -13.05 14.02 6.31
8 12×4×4 231.80 3.92 8.00 15.39 182.97 3.20 8.00 15.43
9 15×6×6 - - - - 180.91 -1.25 5.12 10.10
10 15×12×12 - - - - 239.41 1.05 14.61 11.37
11 25×6×6 - - - - 48.17 2.18 10.15 7.13
12 25×12×12 - - - - 280.90 0.20 29.48 33.29
13 35×6×6 - - - - 55.42 -6.65 8.39 7.83
14 35×12×12 - - - - 38.09 -6.23 5.12 8.19
15 45×6×6 - - - - 48.87 -1.09 10.90 13.72
16 45×12×12 - - - - 32.19 3.39 5.45 6.11
Average 209.70 -4.53 41.60 54.10 161.80 -3.43 25.93 32.73
Prob. O×T×D
=rand
=rand
=max
=max
=min
=min
=avg
=avg
18
4 8×4×4 2.19 250.52 54.01 178.14 0.84 116.74 45.06 95.70
5 10×2×2 -5.39 170.17 88.07 45.00 -8.10 142.16 56.98 32.62
t 10×4×4 -5.39 88.12 6.69 27.87 -8.98 140.71 53.97 17.06
7 12×2×2 1.80 344.89 50.46 73.10 4.10 518.89 10.18 52.61
8 12×4×4 -0.40 91.68 25.35 48.25 1.54 211.88 36.66 28.17
9 15×6×6 - - - - 2.30 180.53 92.20 55.26
10 15×12×12 - - - - -0.17 210.11 107.82 110.64
11 25×6×6 - - - - -1.06 190.42 59.68 50.37
12 25×12×12 - - - - 2.34 128.46 70.01 78.05
13 35×6×6 - - - - 1.76 114.52 53.91 61.29
14 35×12×12 - - - - -3.45 168.16 85.27 77.26
15 45×6×6 - - - - -2.56 98.60 43.28 46.80
16 45×12×12 - - - - 3.42 79.47 40.71 52.64
Average 2.29 281.30 113.00 94.47 0.55 199.18 85.89 58.22
Minimum transportation costs occur when the orders are delivered to the customers at the
nearest time to the delivery date. This situation is done by scenario maximum which all
trucks move with highest speed level. In this situation, the minimum holding cost and
tardiness penalty cost are created. In some cases, the scenario maximum performs better than
the augmented ε-constraint method and CH algorithm. However, the proposed model and the
algorithm can improve the total transportation costs obtained by scenario maximum in some
cases. This improvement is done by efficient scheduling. In this situation, selecting the proper
speed level leads to lower earliness penalty cost beside the tardiness and the hodling cost
compared with scenario maximum.
The minimum value of total carbon emission is obtained when trucks move with minimum
speed level. This situation is created by the scenario minimum. The model and CH algorithm
are able to improve the total carbon emission compared with scenario minimum. This
improvement is due to the maximum use of trucks with high efficiency.
5. Conclusion
Transportation is an important infrastructure in every country. It has a significant effect on
economic growth. Transportation has environmental damaging effects besides all its
advantages. This sector emits a considerable amount of greenhouse gas. Efficient
transportation scheduling can decrease damaging environmental effects and improve supply
chain operations.
In this paper, we studied the integrated green truck transportation scheduling and driver
assignment. We focused on total transportation costs as an index of manufacturing company
and total carbon emission as an index of environmental sustainability. The objectives conflict
with each other. The trade-off between two mentioned objectives was investigated.
19
We proposed a new multi-objective nonlinear mixed-integer programming model. The model
was linearized via a linearization technique. Innovations of this paper includes the assignment
of drivers and considering different speed levels for trucks. The computational results show
the proposed model is efficient. The results suggets that the number of orders has a
significant impact on CPU time. We proposed an efficient constructive heuristic to overcome
the complexity of the problem which is the other innovation of our work. Several instances
were generated and the comparisons were done between proposed constructive heuristic,
augmented ε-constraint method and four introduced scenarios.
The results show that CH has a good performance and obtains high-quality non-dominated
solutions in reasonable CPU time. The results were compared based on the statistical tests to
reach further insights.
In terms of managerial insights of this work, manufacturers and transportation companies can
use the proposed model and heuristic algorithm to make a trade-off between total
transportation costs and total carbon emission. They can utilize the obtained Pareto solutions
and choose a solution considering their policies. Following and implementing the proposed
methods have a great impact on decreasing detrimental environmental effects along with cost
reduction.
There are some ideas for future research. Developing efficient meta-heuristic and heuristic
methods and comparing with proposed CH is a good idea for further research. Considering
multiple distribution centers and multiple paths and extending the proposed mathematical
model to an integrated scheduling and routing problem are other interesting areas for future
research. There may be some uncertainties in real world situations such as uncertainty in
vehicle availability or demands. Working on this subject and applying techniques for
handling uncertainty in the problem discussed seems to be an appealing research topic.
References
Bauer, J., Bektaş, T., & Crainic, T. G. (2010). Minimizing greenhouse gas emissions in
intermodal freight transport: an application to rail service design. Journal of the Operational Research
Society, 61(3), 530-542.
Czyzżak, P., & Jaszkiewicz, A. (1998). Pareto simulated annealing—a metaheuristic technique
for multiple‐ objective combinatorial optimization. Journal of Multi‐ Criteria Decision Analysis,
7(1), 34-47.
Deb, K., Pratap, A., Agarwal, S., & Meyarivan, T. A. M. T. (2002). A fast and elitist
multiobjective genetic algorithm: NSGA-II. IEEE transactions on evolutionary computation, 6(2),
182-197.
ECOFYS, 2010. World GHG Emissions Flow Chart 2010, Kanaalweg 15-G, 3526 KL Utrecht,
The Netherlands.
El Hachemi, N., Gendreau, M., & Rousseau, L. M. (2013). A heuristic to solve the synchronized
log-truck scheduling problem. Computers & Operations Research, 40(3), 666-673.
Fuglestvedt, J., Berntsen, T., Myhre, G., Rypdal, K., & Skeie, R. B. (2008). Climate forcing from
the transport sectors. Proceedings of the National Academy of Sciences, 105(2), 454-458.
Guo, Z., Zhang, D., Liu, H., He, Z., & Shi, L. (2016). Green transportation scheduling with
pickup time and transport mode selections using a novel multi-objective memetic optimization
approach. Transportation Research Part D: Transport and Environment.
20
Guo, Z., Shi, L., Chen, L., & Liang, Y. (2017). A harmony search-based memetic optimization
model for integrated production and transportation scheduling in MTO manufacturing. Omega, 66,
327-343.
Hennig, F., Nygreen, B., Furman, K. C., & Song, J. (2015). Alternative approaches to the crude
oil tanker routing and scheduling problem with split pickup and split delivery. European Journal of
Operational Research, 243(1), 41-51.
Huang, Y., Yang, L., Tang, T., Cao, F., & Gao, Z. (2016). Saving energy and improving service
quality: bicriteria train scheduling in urban rail transit systems. IEEE Transactions on Intelligent
Transportation Systems, 17(12), 3364-3379.
Huang, C., & Karimi, I. A. (2006). Scheduling trans-shipment operations in maritime chemical
transportation. Industrial & engineering chemistry research, 45(6), 1955-1973.
Joo, C. M., & Kim, B. S. (2014). Block transportation scheduling under delivery restriction in
shipyard using meta-heuristic algorithms. Expert Systems with Applications, 41(6), 2851-2858.
Kontovas, C. A. (2014). The green ship routing and scheduling problem (GSRSP): a conceptual
approach. Transportation Research Part D: Transport and Environment, 31, 61-69.
Karaoğlan, İ., & Kesen, S. E. (2017). The coordinated production and transportation scheduling
problem with a time-sensitive product: a branch-and-cut algorithm. International Journal of
Production Research, 55(2), 536-557.
Lee, W. S., Lim, W. I., & Koo, P. H. (2009, July). Transporter scheduling based on a network
flow model under a dynamic block transportation environment1. In Computers & Industrial
Engineering, 2009. CIE 2009. International Conference on (pp. 311-316). IEEE.
Li, X., Wang, D., Li, K., & Gao, Z. (2013). A green train scheduling model and fuzzy multi-
objective optimization algorithm. Applied Mathematical Modelling, 37(4), 2063-2073.
Mavrotas, G. (2009). Effective implementation of the ε-constraint method in multi-objective
mathematical programming problems. Applied mathematics and computation, 213(2), 455-465.
Nakicenovic, N. and Swart, R., 2000. Special report on emissions scenarios. Special Report on
Emissions Scenarios, Edited by Nebojsa Nakicenovic and Robert Swart, pp. 612. ISBN 0521804930.
Cambridge, UK: Cambridge University Press, July 2000., 1.
Nishi, T., & Izuno, T. (2014). Column generation heuristics for ship routing and scheduling
problems in crude oil transportation with split deliveries. Computers & Chemical Engineering, 60,
329-338.
Roh, M. I., & Cha, J. H. (2011). A block transportation scheduling system considering a
minimisation of travel distance without loading of and interference between multiple transporters.
International journal of production research, 49(11), 3231-3250.
Sargut, F. Z., Altuntaş, C., & Tulazoğlu, D. C. (2017). Multi-objective integrated acyclic crew
rostering and vehicle assignment problem in public bus transportation. OR Spectrum, 1-26.
Shafia, M. A., Sadjadi, S. J., Jamili, A., Tavakkoli-Moghaddam, R., & Pourseyed-Aghaee, M.
(2012). The periodicity and robustness in a single-track train scheduling problem. Applied Soft
Computing, 12(1), 440-452.
Siddiqui, A. W., & Verma, M. (2015). A bi-objective approach to routing and scheduling
maritime transportation of crude oil. Transportation Research Part D: Transport and Environment,
37, 65-78.
Tan, K. C., Goh, C. K., Yang, Y. J., & Lee, T. H. (2006). Evolving better population distribution
and exploration in evolutionary multi-objective optimization. European Journal of Operational
Research, 171(2), 463-495.
21
Tanimizu, Y., & Amano, K. (2016). Integrated Production and Transportation Scheduling for
Multi-objective Green Supply Chain Network Design. Procedia CIRP, 57, 152-157.
Wang, W. F., & Yun, W. Y. (2013). Scheduling for inland container truck and train
transportation. International journal of production economics, 143(2), 349-356.
Xu, S., Liu, Y., & Chen, M. (2017). Optimisation of partial collaborative transportation
scheduling in supply chain management with 3PL using ACO. Expert Systems with Applications, 71,
173-191.
Zitzler, E. (1999). Evolutionary algorithms for multiobjective optimization: Methods and
applications, Ph. D. thesis, Swiss Federal Institute of Technology (ETH) Zurich, Switzerland. TIK-
Schriftenreihe Nr. 30, Diss ETH No. 13398, Shaker Verlag, Aachen, Germany.
Zhang, R., Yun, W. Y., & Kopfer, H. (2010). Heuristic-based truck scheduling for inland
container transportation. OR spectrum, 32(3), 787-808.
22
2.54E+15
A
2.49E+15
2.44E+15
2.39E+15
2.34E+15
TCE
2.29E+15 C
2.24E+15
2.19E+15 B
2.14E+15
2.09E+15
15300 17300 19300 21300 23300 25300 27300
TTC
23
Figure 2. Obtained planning for Pareto solution C based on each truck
24
Figure 3. Complexity of different sizes of problems based on CPU-time
25
Constructive heuristic
Input: set of orders, trucks, drivers and days / truck speed levels / capacity of trucks / demand of orders/ due
date of orders/ finish production time of orders/ distances/ related costs to transportation and CO2 emissions
Output: set of non-dominated Pareto solutions NDSPS.
begin
Step 0 (initialization)
Set counter Z= 0;
Set SPS= ø; let SPS represent the set of Pareto solutions; Let represent the speed level of truck in
counter Z;
Set the speed matrix at fast speed ;
Step 1.
Apply model Pfix to schedule orders using speed matrix ( ) and new parameter (Find objective
(1))
Based on the assignment, total carbon emission is calculated (Find objective (2))
let AZ represent the feasible solution for order schedule with truck and drive assignment in step Z;
SPS =A0 SPS;
Let spn _ slow represent the set of orders that their truck speed levels are not slow;
Step 2. (looking for energy efficient solutions)
while spn _ slow do
26
Fast non-dominated sort approach
Input: set of solutions (SPS)
Output: set of non-dominated solutions (NDSPS)
for each b B
Sb
nb 0
for each q B
if (b q) (i.e if b dominates q) then
Sb Sb q
elseif (q b) then
nb nb 1
end
If n b 0 then
NDSPS=NDSSPS {b}
end
end
report the set of non-dominated solutions NDSPS.
Figure 5. Fast non-dominated sort approach
27
Constructive heuristic Augmented ε-constraint
20000
18000
16000
14000
12000
10000
8000
6000
4000
2000
0
05*2*2 25*4*4 48*2*2 6 8*4*48 10*2*2
1 0 10*4*4
12 14
12*2*2 112*4*4
6 18
Figure 6. Comparing aug.eps and CH based on CPU time
28
Figure 7. Comparing aug.eps and CH based performance metrics
29
TOPSIS
Input: Pareto solutions, Original weight for each objective.
Output: Best Pareto solution.
A matrix is formed; m denotes the Pareto solution number and n denotes the
objective function number. Each cell of matrix denotes the value of objective function n for
Pareto solution m;
All values of the matrix are normalized; Let fab denote the value of objective b ( in
Pareto solution a ( ). Let nab denotes the normalize value of objective b in Pareto
solution a;
f ab
nab
m b n (37)
f
a 1
2
ab
Calculate normalized weight for all objectives; Let wb denote the common weight for
objective b. Let denote normalized weight for objective b in Pareto solution a,
respectively;
ab wb nab a m, b n (38)
+ -
Find ideal solution (A ) and nadir solution (A ); Let and denote the best and the worst
value of objective b;
A
min (ab | b {1,2} (t1 , t 2 )
a (39)
A
max (ab | b {1,2}) (t1 , t 2 )
a
Calculate Euclidean distance of each Pareto solution from ideal solution and nadir solution;
and denotes the Euclidean distance form ideal and nadir solution, respectively;
n
d a 2 ( ab tn ) 2
b 1 a m (40)
n
d a 2 (
b 1
ab tn ) 2
Calculate the relative closeness of each Pareto solution; Let Ri denote the relative closeness
of Pareto a ( Ri ;
d a
Ra a m (41)
d a d a
Sort Ra in descending order;
Choose a Pareto solution which its relative closeness value is closer to 1;
Report Best Pareto solution
Figure 8. TOPSIS procedure
30
5000 20
4200 18
CPU time
3400 16
2600 14
CRD
1800 12
1000 10
15*12*12
25*12*12
35*12*12
45*12*12
15*12*12
25*12*12
35*12*12
45*12*12
15*6*6
25*6*6
35*6*6
45*6*6
15*6*6
25*6*6
35*6*6
45*6*6
Problems
Problems
Figure 9. Trend of CPU time and CRD for large size problems
31
Highlights
32