Applied Energy 189 (2017) 221–232

Contents lists available at ScienceDirect

Applied Energy
journal homepage: www.elsevier.com/locate/apenergy

Steady-state optimisation of a multiple cryogenic air separation unit and

compressor plant
Richard Adamson a,⇑, Martin Hobbs b, Andy Silcock b, Mark J. Willis a
a
School of Chemical Engineering and Advanced Materials, Newcastle University, Newcastle upon Tyne NE1 7RU, United Kingdom
b
BOC Gases Ltd., Bawtry Road, Brinsworth, Rotherham S60 5NT, United Kingdom

h i g h l i g h t s

Material and power consumption modelling of a multi-unit gas separation network.

Piece-wise linear data-based modelling of compressor and ASU power consumption.

Development and implementation of a MILP approach to network optimisation.

Demonstrates achievable optimal demand side load management scheduling benefits.

a r t i c l e i n f o a b s t r a c t

Article history: The development and on-line application of a steady-state optimisation strategy for a multiple cryogenic
Received 9 September 2016 air separation unit and compressor plant is discussed. Implemented using mixed integer linear program-
Received in revised form 1 December 2016 ming (MILP), it is demonstrated that the optimiser improves site efficiency at steady state by reduction of
Accepted 11 December 2016
power consumption by up to 5% (a significant saving for such an energy intensive process) while meeting
Available online 23 December 2016
customer demand specifications. This is achieved through determination of the production distribution of
the air separation units and optimal load distribution of the compression network, while simultaneously
Keywords:
ensuring network material balance and network component operating constraints are met. In addition,
Air separation
Compressors
the work demonstrates achievable benefits of demand side load management during peak power pricing
MILP periods, using liquid oxygen as an effective energy storage device. A key constituent of the optimisation
Optimisation strategy is linear modelling to predict individual unit power consumption. Piece-wise linear data-based
Piece-wise linear modelling models of compressor and air separation unit power are shown to provide accurate models which
improve existing on-site power prediction by up to 80% for compressors and 60% for the air separation
units.
Ó 2016 Elsevier Ltd. All rights reserved.

1. Introduction electricity is at its most expensive) e.g. see Manenti and Rovaglio
[2]. The purpose of this paper is to report the development and
Cryogenic air separation plants are energy intensive processes on-line application of a steady-state optimisation strategy that
consuming significant amounts of power (electricity) as a result aims to minimise power consumption of a network of air separa-
of air separation into oxygen, nitrogen etc. and the subsequent tion and gas compression units. This is achieved by optimal pro-
compression or liquefaction of the gas products. Key peculiarities duction distribution of the air separation units and load
of industrial air separation are well documented; air is the only distribution of the compression network.
raw material (it is free and supply is unlimited), the cost of power Related to this work therefore, is the optimisation of industrial
(electricity) is the primary operational cost, >90% [1], which varies gas supply chains that comprise of networks of pipework, com-
throughout the day and the networks simultaneously produce gas pressors and gas production units which are operated to meet cus-
and liquid that may be used to satisfy customer demand (the tomer demand and optimised for economic and environmental
stored liquid being used to satisfy demand at times when cost reductions, e.g. see Üster and Dilaveroğlu [3], Azadeh et al.
[4] and Gao and You [5], who use a mixed integer nonlinear pro-
gramming (MINLP) algorithm to optimise economic and environ-
⇑ Corresponding author.
mental objectives. In Cortinovis et al. [6], a MINLP approach to
E-mail addresses: richard.adamson1@newcastle.ac.uk (R. Adamson),
mark.willis@newcastle.ac.uk (M.J. Willis). the optimisation of a natural gas transmission system comprising

http://dx.doi.org/10.1016/j.apenergy.2016.12.061
0306-2619/Ó 2016 Elsevier Ltd. All rights reserved.
222 R. Adamson et al. / Applied Energy 189 (2017) 221–232

Nomenclature
F flow rate (HCMS/h)
Abbreviations PD discharge pressure (bar)
ASU air separation unit y flow/Boolean auxiliary variable
HP high pressure CkW spot power cost (£/MWh)
LN liquid nitrogen Cs liquid use cost (£/HCM)
LP low pressure e machine efficiency (kW/HCM)
MP medium pressure J cost function (£/hr)
GO gaseous oxygen W power consumption (kW)
IC internally compressed
LO liquid oxygen Subscripts
ME model mean error (kW) c compressor
TLO total liquid oxygen k pseudo-machine number
s liquid back-up supply
Parameters v valve
^
b model coefficient j compressor or ASU number
CLO liquid make cost (£/HCM) m model number
d Boolean coefficient u unit (ASU)

of parallel compressors power consumption is also reported. This horizons of hours, not days. The implementation of an MINLP
allows the monitoring of power regression curves against real time approach using nonlinear empirical models, reported in Adamson
data to track compressor performance and identifying when main- et al. [17] generally yielded excessive solution times rendering
tenance is required. The optimal load sharing of a network of com- the on-line application impractical. In this work the steady-state
pressors has also been considered by Han et al. [7] and Abbaspour optimisation problem is formulated as an MILP in order to ensure
et al. [8] who developed off-line optimisation approaches while robust and efficient on-line optimisation.
Paparella et al. [9] and Xenos et al. [10] considered the on-line opti- To develop a MILP optimisation model, the power consumption
misation of a network of air compression units. In Paparella et al. of network components (ASU and compressors) are determined
[9] the optimal load distribution of a number of parallel compres- using a piece-wise linear modelling approach. Lin et al. [18] pre-
sors in a natural gas pipe-line were considered. While Xenos et al. sent a review of the use of piecewise linearisation techniques, find-
[10] optimised an industrial compressor station that served com- ing methods can be used to efficiently discover approximated
pressed air to air separation and chemical plants. Where compres- globally optimal results. There are many recent examples in liter-
sor load sharing performance can be robustly modelled, Øvervåg ature regarding the use of piece-wise linearisation techniques to
[11] uses a MINLP model predictive controller to optimally load partially or fully formulate a MILP approach to gas network optimi-
share using efficiency curve data. sation. For example, Martin et al. [19] formulates a large MILP
Development of predictive models of both compressor and air problem to solve the optimisation of natural gas networks using
separation unit power consumption is a particular focus of this piece-wise linear approximations of nonlinear constraints at
work. The operating principles of compressors (e.g. their character- steady state. While Kolb et al. [20] describe how gas network opti-
istic performance and operation limits) are normally described in misation can be better achieved by developing piece-wise linear
terms of compressor efficiency and detailed hybrid models of approximations of network components. In Camponogara et al.
power consumption in industrial multistage compressors have [21] and Aguiar et al. [22] gas-lifted oil field production costs are
been developed, e.g. see Han and Han [12]. Similarly, mechanistic optimised by combining convex nonlinear regression curves with
models of air separation units exist within the literature which piece-wise linear approximations of these curves at given pressure
may be used as the basis for prediction of power consumption and routing constraints. While Domschke et al. [23] solve a com-
e.g. see Huang et al. [13]. However, the development of a set of plex natural gas network optimisation problem by integrating
robust and reliable models from fundamental principles would piece-wise linear approximations to partially linearise a nonlinear
be an onerous task for a complex mix of industrial air separation cost function. Furthermore, in Correa-Posadaa and Sánchez-Martın
units and compressors. Therefore, an empirical data-based mod- [24] the piecewise linearisation of general gas flow equations and
elling approach is used in this work. Related work includes Puranik the development of linear models of other gas network compo-
et al. [14] who used nonlinear regression models as the basis of a nents, such as machine power are discussed. Adopting a similar
MINLP approach to the optimisation of an oxygen and nitrogen approach, in this work, it is demonstrated that piecewise linear
customer network. Similarly, in Cao et al. [15], an MINLP approach models can accurately predict the power consumption of both air
was used to determine the optimal production rates required to separation units and compressors with prediction accuracies com-
meet customer nitrogen demands and assess dynamic compressor parable to the best nonlinear alternative.
performance. In addition, Kopanos et al. [16] and Xenos et al. [10] Where power consumption can be accurately predicted at a
also describe power consumption modelling using nonlinear given time of use, the operational network cost can be determined
empirical models to capture the relationship between power con- ahead of use, subject to the predicted price of power. Flexible
sumption, flows, temperatures and pressures for use within an power loads can be selectively purchased on spot power markets
optimiser to improve the operational costs of a parallel network at variable prices to reduce overall commodity costs in conjunction
of air compressors and to optimally schedule maintenance. with demand side load management, Merkert et al. [25]. We refer
In all previous work on air separation unit and compressor plant to the examples in Karagiannopoulos et al. [26] in which flexible
optimisation, site optimisation for gas network demand is devel- loads are manipulated into an overall less energy efficient
oped using a MINLP approach with dynamic optimisation using operation strategy whilst still delivering operational cost savings
demands predicted for the days and weeks ahead. A key by reducing load at peak power price. Similarly, Zhang et al. [27]
operational aspect of the process considered in this work is that use a MILP scheduling approach to liquefy air off-peak and
customer demands are unknown and unpredictable, with time generate power directly via turbine at peak on underutilised air
 R. Adamson et al. / Applied Energy 189 (2017) 221–232 223

FGO,v1

FTLO FGO,s
FGO,u1
u1 FGO,c4
FLO,u1

FGO,c1 FGO,c5

FGO,c2 FGO,c6
FGO,u2
u2
FLO,u2 FGO,c3 FGO,MP

FGO,v2

FGO,HP
FGO,u3i
FGO,u3ii
u3
FLO,u3
FLN,u3

Fig. 1. Margam supply network and compression configuration for oxygen production (Table 1 provides a description of each network flow). The network comprises of three
air separation units (ASUs) (u1, u2 and u3) fed by four air compressors, three centrifugal oxygen compressors (c1, c2 and c3), three reciprocating oxygen compressors (c4, c5
and c6), one spill valve (v1), one cross over valve (v2) and flows to liquid oxygen (FLO) and nitrogen storage (FLN) tanks.

separation plants. However, expansion processes deliver a lower 1.1. The multiple cryogenic air separation unit (ASU) and compressor
round-trip efficiency than load shifting the power requirement of plant
usable product production. In this work, we consider a similar
approach to Xenos et al. [28] in which liquid oxygen is used The multiple cryogenic air separation unit (ASU) and compres-
directly as an energy storage device for load management benefits sor plant considered in this work is an oxygen gas supply network
by consuming liquid and temporarily stopping product compres- operated to meet customer demands of oxygen gas at two pres-
sion, load shifting additional power consumption to off-peak sures. A schematic of the network (which is the Margam site in
periods for liquid production. ASUs are considered fixed loads the UK) is shown in Fig. 1. It comprises a number of centrifugal
and external product compression is flexibly load managed as
non-dispatchable for power price peak-shaving activities.
The key contribution of this paper is the combined minimisa- Table 1
Margam oxygen gas supply and compression network components and descriptions.
tion of the production distribution of a network of air separation
units and external compressors using a MILP technique which Flow Description
incorporates accurate piece-wise linear network component power F GO;u1 Production flow of gaseous LP oxygen from ASU u1
models. In particular, we develop an optimisation strategy that F GO;u2 Production flow of gaseous LP oxygen from ASU u2
must operate using unpredictable and frequent changes in cus- F GO;u3i Production flow of gaseous MP oxygen from ASU u3
tomer demands (not day ahead predictions). Our on-line applica- F GO;u3ii Production flow of gaseous HP oxygen from ASU u3
F GO;c1 LP to MP oxygen flow through centrifugal compressor c1
tion results demonstrate that the optimiser improves site
F GO;c2 LP to MP oxygen flow through centrifugal compressor c2
efficiency at steady state by reduction of power consumption by F GO;c3 LP to MP oxygen flow through centrifugal compressor c3
up to 5% (a significant saving for such an energy intensive process1) F GO;c4 LP to HP oxygen flow through reciprocating compressor c4
while meeting customer demand specifications using fast, efficient F GO;c5 MP to HP oxygen flow through reciprocating compressor c5
and readily available commercial (and open source) software. Fur- F GO;c6 MP to HP oxygen flow through reciprocating compressor c6
thermore, we demonstrate the optimiser can be used as an advanced F GO;v 1 Flow of oxygen spilled from the LP line through valve v1
F GO;v 2 Flow of oxygen expanded from HP to MP through valve v2
scheduling tool to more than recover the costs of demand side load
F GO;s Flow of oxygen vaporised from liquid oxygen storage
management actions taken to reduce the network’s exposure to vari- F GO;MP Flow of MP oxygen to the customer
ation in spot market power pricing. F GO;HP Flow of HP oxygen to the customer
F LO;u1 Production flow of liquid oxygen from ASU u1
F LO;u2 Production flow of liquid oxygen from ASU u2
F LO;u3 Production flow of liquid oxygen from ASU u3
1
e.g. the Air Liquide Group has more than 400 air separation units worldwide. The F TLO Total production flow of liquid oxygen from ASUs to storage
group’s total electricity consumption in 2010 corresponded to more than one F LN;u3 Production flow of liquid nitrogen from ASU u3
thousandth of the world’s total electricity consumption [29].
224 R. Adamson et al. / Applied Energy 189 (2017) 221–232

and reciprocating compressors with various combinations of series the oxygen gas throughout the network in order to minimise total
and parallel operation. Three air separation units (ASUs) provide operating costs. The outputs of the optimiser are the optimal ASU
production flows of gaseous and liquid oxygen. Gaseous oxygen production rate set-points and the running requirement of each
(GO) is compressed and sent directly to the customer while liquid compressor within the network. For significant demand changes,
oxygen (LO) is stored in vessels on site for gas network back up this will also allow the prediction of the optimal network configura-
during compressor trips, to meet customer demands at times of tion, i.e. whether a particular ASU or compressor should be turned on
high electricity prices and to meet local bulk liquid customer or off. To implement the optimisation strategy a mathematical rep-
demands by road tanker. Two ASUs produce low pressure (LP) resentation of the oxygen gas network must be constructed. The
and the third is an internally compressed (IC) ASU capable of pro- three components of this optimisation model are (a) steady state
ducing medium pressure (MP) and high pressure (HP) GO directly, material balance relationships which are used to ensure an optimal
along with liquid nitrogen (LN). Four separate air compressors feed configuration meets customer demand (b) a description of the
the three ASUs with high pressure compressed air. power consumption of each of the ASUs and compressors within
Three centrifugal compressors raise LP GO to MP GO and three the network which relates operating cost to the individual flows
reciprocating compressors raise LP GO or MP GO to HP GO. A pres- throughout the network (c) a description of the total operating cost
sure control loop allows cross over between the HP GO to MP GO of the gas supply network. The mathematical model is then used in
stream and the network is supported, when required, by an oxygen conjunction with process operational data (operating pressures,
tank pump vaporiser system converting LO to GO when the HP flow-rates etc.) to optimise the plant configuration.
pipeline pressure falls due to under-production or over-
consumption of HP gas. When there is overproduction of LP gas 2.1. Steady-state material balances
for the current compression configuration, a pressure controlled
LP GO gas spill valve will open. However, in normal operation In order to define the material balance relationships across the
the spill valve, v1, remains shut as ASUs should preferentially be site it is assumed that there are no dynamic gas losses due to spill
ramped down to prevent spill as production of oxygen consumes from machines or fouling. Furthermore, as the network is relatively
power. In other words, an open spill valve often implies an incor- local and pipelines between machines short, pipeline friction losses
rect process configuration. are not modelled. The material balance relationships can therefore
All ASUs produce gaseous nitrogen streams and a secondary be defined based upon the steady-state temperature and pressure
nitrogen network consisting of four compressors is operated in corrected2 flowrates through the various components of the
parallel to the oxygen network. As the site is operated in oxygen network.
lead, nitrogen gas is almost always produced in excess and spilled
to atmosphere after heat exchange with the compressed air feed. 2.1.1. Low pressure oxygen
For this paper, the production and supply of nitrogen gas is there- Referring to Fig. 1, the flow, F, in hundred cubic meters per hour,
fore not considered as part of the plant-wide steady-state optimi- HCMs/h, of gaseous oxygen, GO, production from the air separation
sation strategy. units u1 and u2 is equal to the flow of gaseous oxygen through
compressors c1, c2, c3 and c4 plus the gaseous oxygen flow leaving
1.2. Current operating policy the network via the low pressure spill valve, v1,
F GO;u1 þ F GO;u2 ¼ F GO;c1 þ F GO;c2 þ F GO;c3 þ F GO;c4 þ F GO;v 1 ð1Þ
The network’s manager, BOC Gases, receives communication of
independent demands of MP and HP GO from the customer. ASU
production rate set points, managed by a linear model predictive 2.1.2. Medium pressure gas flow
controller and the compression network configuration and loads The customer demand of medium pressure, MP, gaseous oxygen
are then manipulated from the current configuration to meet the is met by the production of MP from ASU3, u3i, compression of LP
customer order. To do this, the operators attempt to minimise gas from c1, c2 and c3, the cross over flow through the let-down
overall power usage of the plant which is a difficult task as all net- valve from the HP line, v2 minus the suction inlet of MP required
work components vary in specifications, including capacities and for the HP compressors, c5 and c6,
efficiencies. Furthermore, the characteristics of all network compo-
F GO;MP ¼ F GO;u3i þ F GO;c1 þ F GO;c2 þ F GO;c3  F GO;c5  F GO;c6 þ F GO;v 2
nents change over time and vary dependant on current operating
conditions leading to inconsistency in decisions resulting in sub- ð2Þ
optimal re-configuration of the plant.
Current re-configuration strategies include ramping units indi- 2.1.3. High pressure gas flow
vidually in order to meet changes in customer demand instead of The customer demand of high pressure, HP, gaseous oxygen is
load sharing between units and selecting compressor loads and met by the production of HP from ASU3, u3ii, compression of MP
the production distribution of the ASUs using knowledge of gas from c5 and c6, and the liquid back up supply flow, F GO;s , minus
machine efficiencies. As key performance indicators of machine the cross over flow through the let-down valve from the HP to the
wear and fouling, machine efficiencies are calculated monthly as MP line, v2,
the average power consumption divided by the average flow
through the machine. This method fails to accurately capture the F GO;HP ¼ F GO;u3ii þ F GO;s þ F GO;c4 þ F GO;c5 þ F GO;c6  F GO;v 2 ð3Þ
relationship between the operating efficiency of a unit and the pro-
cess variables that affect it (temperatures, pressures, flows etc.) at 2.1.4. Liquid oxygen production
given throughputs over time. Liquid oxygen is produced by air separation units u1, u2 and u3,
and the total site production, F TLO , is given by,
2. Plant-wide modelling for steady-state optimisation
F TLO ¼ F LO;u1 þ F LO;u2 þ F LO;u3 ð4Þ
The concept of real-time optimisation scheme (e.g. see [30]) can
be tailored to the operation of the integrated plant of ASU’s and com- 2
Site flow meters are calibrated regularly and twinned with pressure and
pressors. In this application, the aim of the real-time (or steady- temperature meter readings to provide a calculated standardised flow which assumes
state) optimisation scheme is to determine the optimal flows of ideal gas behaviour.
 R. Adamson et al. / Applied Energy 189 (2017) 221–232 225

Each ASU requires a minimum LO production flow at all times feed temperature was observed as roughly constant due to temper-
to purge hydrocarbons from the column liquid sump. When addi- ature control of the product streams exiting the ASU which negate
tional LO production is required to build liquid stocks, a production the majority of temperature variation. In addition, compressed air
target can be met by optimisation to select the most efficient avail- streams into ASUs are cooled by direct coolers and chillers to
able ASUs. cooler-than-ambient conditions. In this work, the effect of temper-
ature on power consumption on the models is therefore smaller
2.2. Power consumption of each unit within the network than the influence of discharge pressure and gas flow due to tem-
perature control management. Auxiliaries such as cooling water
The efficiency of a compressor is also known as its specific and water chiller duty have not been included in the models of net-
power. This represents the power, W (kW), required to compress work power consumption as these are not a manipulated variable
1 HCM of gaseous oxygen. Specific power, e, (kW/HCM), is nor- of the optimisation technique but a systematic cost of operating
mally used to calculate the power required by the machine to com- the site.
press the swept volume through the machine. Kurz et al. [31]
describe the differences between the power consumption (and 2.3. Total operating cost of the network
hence efficiency) of centrifugal and reciprocating compressors at
a constant discharge pressure and temperature. As gas flow The total network operating cost, J (£/h) is a function of the sum
increases through a centrifugal compressor, the power consump- of compressor and air separation unit power multiplied by the cost
tion increases proportionally presenting a constant efficiency, of power, C kW (£/kWh) plus the cost of consuming liquid oxygen, C S
however, machines exhibit nonlinear behaviour approaching the (£/HCM), when a backup flow, F S , is required plus the cost of pro-
lower and upper limits of operation. Reciprocating compressors ducing LO, C LO (£/HCM),
have higher fixed mechanical power requirements, simply moving !
X
6 X
3
the piston, resulting in an efficiency relationship which favours the J ¼ C kW  W cj þ W uj þ C s  F s þ C LO  F TLO ð8Þ
maximum design gas flowrate. In this work it is assumed that j¼1 j¼1
specific power may be mathematically modelled as a function of
flow through the machine and the discharge pressure, therefore The cost of power, C kw , is the current spot market price plus
the power consumption ðW cj Þ of compressor cj, is given by, non-commodity costs such as charges and taxes, the price of liquid
back up consumption is assumed to be the average cost of its gen-
W cj ¼ F GO;cj  ecj ðF GO;cj ; PDj Þ ð5Þ eration and the cost of making LO is the average cost of production,
which involves consuming liquid nitrogen.
where the efficiency profile (functional relationship) depends on
the type of product compressor, either centrifugal or reciprocating.
The power consumption of the two low pressure oxygen pro- 3. Data-based modelling of machine efficiency/power
ducing air separation units is defined using a modified version of
Eq. (5). Air separation unit power may be calculated using the dif- As discussed in Section 2.2, machine efficiency (and hence
ference between the power of the air compressor/s feeding the power) is a nonlinear function of flow through a machine and
separation column and the power generated from the turbine the discharge pressure. With years of historical data available in
(which is used to balance ASU column temperatures). As ASU the data historian, it is possible to produce efficiency curves as well
power is directly related to total oxygen production (gaseous oxy- as discover the flow limits for each machine (discussed in Section 4
gen plus liquid oxygen), the power consumption of units u1 and of the paper). 15 min averaged data of oxygen gas flow, machine
u2, W ujðj¼1;2Þ , is given by, discharge pressure, power usage and other key normal operation
indicators, such as recycle valve opening position etc., were com-
W ujðj¼1;2Þ ¼ ðF GO;uj þ F LO;uj Þ  euj ððF GO;uj þ F LO;uj Þ; PDj Þ ð6Þ piled over 6 months for each machine in the compression arrange-
where the machine efficiency is taken to be a function of total oxy- ment. Data was pre-screened to remove missing data and data not
gen production rate (gaseous oxygen plus liquid oxygen production recorded during normal operation, e.g. at machine start up. Both
rates) and the air compressor discharge pressure, which for an ASU
is the column pressure. Column pressure is known to be a function
of total oxygen production rate as it increases with increasing com-
pressed air supply flow rate. As ASU u2 has two modes of operation,
1400
being supplied by the main air compressor only, or both the main
and booster air compressors in parallel feeding the unit, power 1300
models of each air compressor feed operation were developed.
Power (kW)

The third ASU on the plant (u3) generates MP, HP and LO to 1200
storage simultaneously therefore the power consumption of oxy-
1100
gen production is,

W u3 ¼ ðF GO;u3i þ F GO;u3ii þ F LO;u3 Þ  eu3 ððF GO;u3i þ F GO;u3ii þ F LO;u3 Þ; PDj Þ 1000

 F LN;u3  eLN ð7Þ 900

45
However, ASU u3 also produces LN directly through internal 40 90
compression. As the network is optimised to minimise the cost of 35 80
70
gaseous oxygen production, the LN product flow, F LN;u3 , is multi- 30 60
Pressure (bar)
plied by a fixed LN production efficiency term, eLN , to approximate 25 50 Flow (HCMs/hr)
the power used to produce LN. This is then subtracted from the
overall ASU power consumption of the unit. Fig. 2. Power consumption as a function of flow (HCMs/h) and discharge pressure
(bar). The process data demonstrates a multivariate nonlinear relationship. The data
Kurz et al. [31] describe how gas density, a function of temper- shown is for reciprocating compressor (c4) with the centrifugal compressors also
ature, is known to have an effect on the work done by a compres- demonstrating nonlinear relationships between power, flow and discharge
sor. Whilst studying the empirical data sets, product compressor pressure.
226 R. Adamson et al. / Applied Energy 189 (2017) 221–232

univariate and multivariate plots of the process data were ‘eye-

balled’ in order to ascertain interdependencies between the vari-
ables before data-based modelling was undertaken. As an example,
the nonlinear relationship between the power consumption (and
hence efficiency) of a reciprocating compressor and the flow and
discharge pressure may be observed in Fig. 2. Where the compres-

Power (kW)
sor recycles due to anti-surge control at low flow rates and is over-
loaded at high flow rates, a reduction in machine efficiency is
observed. As a result, machines often run more efficiently per unit
of gas when fully loaded and the resulting non-linear relationships
are convex. Linear regression would not capture these
complexities.
In Adamson et al. [17] and Xenos et al. [10] polynomial nonlin-
ear regression models of compressor or ASU power were used.
Referring to Eqs. (5)–(7), efficiency is multiplied by flow to calcu-
late power, where efficiency is a function of flow and pressure. This
introduces combination and modelling non-linearity and can be
simplified by building multivariate regression models of power Flow (HCMs/hr)
directly. A typical non-linear model being of the form (where the
^0;j ; . . . ; b
b ^4;j are model coefficients), Fig. 3. Plot of power with respect to flow for compressor c4 between the limits
shown in Fig. 2. The separate colours denote flow data within the constraint
boundary of each piece-wise linear model. To assist visualisation of the model
^ þb
^ cj ¼ b ^ F ^ ^ 2 ^ 2
1;j GO;cj þ b2;j P Dj þ b3;j F GO;cj þ b4;j P Dj ð9Þ
W predictions the discharge pressure is not shown, rather, the black lines present the
0;j
prediction obtained using each piece-wise linear model at the discharge pressure
However, a nonlinear model cannot be embedded within a MILP midpoint of each region. Additional, straight-line relationships would be obtained
given any discharge pressure within a particular constraint boundary.
approach to network optimisation which requires all components
of the optimisation model to be linear.
3.1.1. Multivariate piece-wise linear modelling of the compressors
Each piece-wise segment of the compressor data are regressed
3.1. Piecewise linear modelling to the following model,
^ þb
^ cj;k ¼ b
W ^ F GO;cj þ b
^ PDj ð11Þ
0;k 1;k 2;k
Piece-wise linear models are the combination of many linear
models to provide a pseudo-nonlinear overall model. The typical For the compressors, the discharge pressure3 is used to define
structure of a univariate piecewise linear model is (where u1 is a the model breakpoints, i.e. denote the points at which one model
single input and y is a process output and N are the number of data begins at and the previous one ends. To develop the models, plant
records), data was pre-screened and erroneous data points indicating non-
running, non-steady state or operation outside normal operating
8 9
> b ^11 u
^01 þ b u10 6 u1j < u11 > range were removed. Fig. 3 shows a typical set of data for compres-
>
> 1j >
>
>
> >
> sor c4. Here, four piece-wise linear models were found to provide the
<b ^02 þ b
^12 u1j u11 6 u1j < u12 =
yjðj¼1;...NÞ ¼ best fit with the lowest modelling error to generate estimated power
> ...
>
>
..
.
>
>
> consumption from the measured flow and pressure data.
>
> >
>
:^ ^ u1j ; The model equations and the optimal breakpoints are,
b0k þ b 1k u1k1 6 u1j < u1k 8 9
> 320:75 þ 22:01PD4 þ 10:26F GO;c4 26:33 6 PD4 < 30:12 >
The constraints represent the regions in which each model is >
> >
>
< 285:47 þ 18:88PD4 þ 11:06F GO;c4 30:12 6 P < 34:68 =
valid (between the minimum, u10 and maximum, u1k value of u1). ^ c4 ¼
W
D4

A number of approaches have been developed to create linear >

> 244:07 þ 15:49PD4 þ 12:23F GO;c4 34:68 6 PD4 < 40:21 >
>
>
: >
;
piecewise models from large data sets e.g. see Yang et al. [32]. 182:19 þ 12:18P D4 þ 13:39F GO;c4 40:21 6 PD4 < 43:14
The primary difficulty being to optimally determine the total num- ð12Þ
ber of models, their parameters (b ^ Þ and the breakpoints
^01 ; . . . ; b
1k
In other words, the empirical model defines ‘pseudo-machine’
(u11 ; . . . ; u1k ) that define the regions of validity for each model. In
models, each with a lower and upper pressure bound, i.e. for the
this work a pragmatic approach to piece-wise linear modelling
purpose of building the steady-state optimiser the model of com-
was adopted, simplifying the model development task by assuming
pressor c4 may be considered as four separate (linear) machines
a maximum of five linear models, kmax ¼ 5, are required for accu-
with the appropriate pseudo-machine being determined by the
rate relationship capture. This gives the following cost function,
current discharge pressure.

X
kmax
Jm ¼ ^ b
kyj  b ^ u k 8y ; u 2 ðu 3.1.2. Univariate piece-wise linear modelling of the ASUs
0k 1k 1j 1 j 1j 1k1 6 u1j < u1k Þ ð10Þ
k¼1 Ideally, a piece-wise linear model of an ASU would use identical
methodology to compressors. However, as column pressure is
A nonlinear optimiser, similar to the method employed by Szucs
directly related to the total oxygen production rate, unlike pipeline
et al. [33], is then used to minimise (10) in order to determine each
and discharge pressure of compressors, the current column pres-
of the model parameters b ^ ;b
^ ðk ¼ 1; . . . ; kmax Þ and iteratively
0k 1k sure cannot be fed into the model to estimate the ASU power con-
determine the position of the model break-points u1k ; . . . ; u1kmax 1 . sumption at the optimised flow. Therefore, ASU piece-wise linear
In an outer-loop, the value of kmax is increased from an initial value
of two to the maximum value of five. The resulting models are val- 3
Models could have been developed using either flow or discharge pressure
idated on a second data-set and the piecewise linear model that breakpoints, however the use of pressure breakpoints were found to produce models
minimises the validation error chosen as the optimal model. that gave more accurate prediction of machine power.
 R. Adamson et al. / Applied Energy 189 (2017) 221–232 227

Pressure (bar)
Power (kW)

Flow (HCMs/hr) Flow (HCMs/hr)

Fig. 4. Plot of power with respect to flow for ASU u2. The separate colours denote Fig. 5. Plot of flow and discharge pressure of compressor c4 at steady state between
flow data within the constraint boundary of each piece-wise linear model. Black the limits shown in Fig. 2. Nonlinear regression models (16) were regressed to
lines present the prediction obtained using each piece-wise linear model within the selected points (in red) to determine the lower and upper operational bounds at a
particular regions. Six piece-wise regions are shown, three with one air compressor given pressure. The black lines are the predicted flow constraints (15) obtained
feed and three with two air compressor feeds. Two piece-wise linear models, one from model (16). Data points outside the regressed lower and upper flow bounds
from each air supply configuration, overlap with respect to total oxygen production were outside of ideal operating range, ie running with gas recycle or overloaded.
flow rate. (For interpretation of the references to colour in this figure legend, the reader is
referred to the web version of this article.)

power models must be univariate, only considering production

flow from the ASU, with flow breakpoints and are given by, Therefore, two polynomial regression fits using data on the edge
of the operating regions are used to capture the relationships
^
^ uj;k ¼ b ^
W 0;k;uj þ b1;k;uj ðF GO;uj þ F LO;uj Þ ð13Þ between flow and discharge pressure. The regression model can
ASU u2 can be fed by either the main air compressor only or then be used to determine the lower and upper flow constraint,
both the main and booster air compressors simultaneously. Each Eq. (15). As the steady state optimiser does not alter pressure,
compressed air supply configuration has a separate piece-wise lin- the bounds are fixed prior to solving and nonlinearity is not intro-
ear modelling strategy. Using, kmax ¼ 5 as a starting point, the duced into the cost function.
piece-wise linear optimiser determined that the optimal model Fig. 5 shows the upper and lower flow bounds of compressor c4,
comprised three piece-wise segments for each compressor feed with respect to discharge pressure. The lower (and upper) bound
operation. An overlap between models demonstrates network con- was obtained using the following regression model to the data
figurations where the operator should have considered turning off points indicated in Fig. 5.
one air compressor to reduce the overall ASU power consumption.
Fig. 4 shows the six piece-wise linear models with production flow F^min ^ ^ ^ 2
GO;cj ðP Dj Þ ¼ a0;cj þ a1;cj P Dj þ a2;cj P Dj ð16Þ
breakpoints for ASU u2 (the actual model coefficients are omitted
due to industrial confidentiality).
Analysis of the difference in operation of ASUs in summer and
winter shows a reduction in the maximum production rate range 5. Total operating cost model linearisation
of operation of ASU u2. This may be due to temperature differences
in the air increasing the difficulty of compressing warmer, less The total operating cost of the network, Eq. (8) may be rewritten
dense air in the summer. To represent this, the pseudo machine to include the piece-wise power models,
with the highest production flow range is removed from the opti- !
miser during the summer months. X
6 X
mj X
3 X
mj
J ¼ C kW  ^ cj;k  dcj;k þ
W ^ uj;k  duj;k
W þ C s  F s þ C LO  F TLO
j¼1 k¼1 j¼1 k¼1
4. Machine processing capacity
ð17Þ
Realistic modelling of machines requires the determination of The network cost is therefore a sum of each machine power,
the minimum and maximum flow production or processing ability which is a sum of all the pseudo-machine model powers
of the ASUs and compressors. For example, the lower and upper ðk ¼ 1; . . . ; mÞ, multiplied by the cost of power and the additional
flow limits of a compressor could be assumed constant and given liquid costs and mj are the number of piecewise linear models
by, obtained for each unit after model validation. In addition, Boolean
coefficients, dcj;k 2 f0; 1g and du;j;k 2 f0; 1g are introduced into the
F min max
GO;cj 6 F GO;cj 6 F GO;cj ð14Þ
cost function in order to remove machines from the network when
However, process operation limits are known to be a function of not required as well as ensuring that machines are operating
machine size, discharge pressure and weather conditions. For the between the minimum and maximum flow constraints of each
compressors, operational data shows the lower and upper limits pseudo-machine region. A zero indicates the machine is off (there-
of flow change with the discharge pressure of the machine, i.e. fore removing the respective power term from the operating cost)
and one multiplied by the power indicates it is on. To prevent the
F min max
GO;cj ðP Dj Þ 6 F GO;cj 6 F GO;cj ðP Dj Þ ð15Þ optimiser selecting multiple pseudo-machines simultaneously,
228 R. Adamson et al. / Applied Energy 189 (2017) 221–232

mutually exclusive constraints are then added for the sum of Boo- flows F s and F TLO . The minimisation is performed with respect to
lean coefficients for each machine. For example, for the ‘mj’ Boo- the material balance constraints, Eqs. (1)–(4), and the flow range
lean coefficients of models for compressor ‘j’, limits, Eq. (15), and the constraints that define the state of the Boo-
lean variable, Eq. (21).
X
mj
dcj;k 6 1 ð18Þ
k¼1 6. Implementation aspects

Prior minimisation of Eq. (17), the current plant discharge pres- A supply network mimic was produced in Microsoft Excel as an
sures are imported. With no further updates, the discharge pres- optimiser interface (see Fig. 6). The user inputs the current cus-
sure of each unit remains constant during optimisation therefore tomer demand, which is linked to the mass balance constraint
the only decision variables in the piece-wise models are flow and cells, to optimise the compression arrangement. The reasons for
the associated Boolean coefficients. As the multiplication of the using Excel were that (a) the steady-state optimiser had to be com-
Boolean and flow variables introduces combinational nonlinearity, patible with all computers within the company allowing the opti-
in order to use a MILP solver they must be removed. The combina- misation results to be integrated with exiting data access methods
tion of the estimated power model and the Boolean gives, using data historian Excel plug ins and implemented by the process
^ þb
^ cj;k ¼ ðb ^ P Þd þb ^ F operators, (b) it had to be cheap to run and easily updateable once
W 0;k 1;k Dj cj;k 2;k GO;cj;k  dcj;k ð19Þ
developed as machine specifications change, (c) the optimiser
^ F interface had to be easy for process operators to understand.
The second term in Eq. (19), b 2;k GO;cj;k  dcj;k , causes nonlinearity
therefore an auxiliary variable yc;j;k is introduced where, The optimised flows through the machine, the upper and lower
flow limits, and the estimated machine power are presented in a
ycj;k ¼ F GO;cj;k  dcj;k ð20Þ machine representation of the network.
An automated spreadsheet running scheduler positioned on a
When directly optimising for ycj;k , the variable’s lower and
24 h server accommodates the macro run spreadsheet solving
upper limits are indirectly subject to changes in the Boolean vari- method. On opening, the spreadsheet downloads the current
able value. For example for a pseudo-machine of a compressor, demand, pressure, flow, temperature, power price and machine
availability tag data required for power estimation and the mate-
F^min ^max
GO;cj ðP Dj Þ  dcj;k 6 F GO;cj 6 F GO;cj ðP Dj Þ  dcj;k ð21Þ
rial balance constraints. A visual basic program then automatically
Therefore, F GO;cj ¼ 0 if dcj;k ¼ 0 and F^min ^max
GO;cj ðP Dj Þ 6 F GO;cj 6 F GO;cj ðP Dj Þ
runs the MILP (solved using Excel’s built in solver routine) to min-
when dcj;k ¼ 1. imise overall steady state network cost, it then updates the
machine requirement running tags (0 = off, 1 = on) and ASU pro-
Therefore, the cost function used to minimise total network cost
duction flow set points and exports this information to the data
is,
historian. The scheduler runs the optimisation spreadsheet every
X
6 X
mj 15 min and as the results are only relevant at steady state and
J ¼ C kW  ^
ðb ^ ^
0;k;cj þ b1;k;cj P Dj Þ  dcj;k þ b2;k;cj  ycj;k the network must be reconfigured from the current position to
j¼1 k¼1
! the optimal configuration over time by the process operator.
X3 X
mj
þ ^ ^
b 0;k;uj  duj;k þ b1;k;uj  yujk þ C s  F s þ C LO  F TLO ð22Þ
6.1. Machine availability
j¼1 k¼1

This is minimised with respect to the binary variables Machines were often unavailable for periods of time during
dcj;k 2 f0; 1g, du;j;k 2 f0; 1g, the auxiliary variables, ycj;k , yujk and the maintenance or after a machine trip. A selection of dropdown text

Fig. 6. Spreadsheet-based optimiser operator interface to present the optimal production flow rate set points and compression configuration by network mimic (all
numerical values are false). The user inputs the customer demand combination in the top left corner ready for solving. The optimisation result is displayed within the relevant
boxes representing the machines in the network. Boxes prior to the ASUs display the air compressors required to meet the oxygen production rate.
 R. Adamson et al. / Applied Energy 189 (2017) 221–232 229

boxes were positioned on the operator interface and if the machine

was unavailable, the machine’s Boolean coefficient was forced to
be zero (using an equality constraint, e.g. dcj;k = 0). Adding this flex-
ibility allowed the optimiser to cater for any network configura-
tion. Forcing the Boolean coefficient to zero for liquid back up
supply also allowed liquid back up to be prevented for most situa-
tions as favoured by liquid schedulers who wish to preserve stocks.

Power (kW)
7. On-line application results

This section assesses the performance of the piece-wise linear

models, running estimations of machine power in parallel with
actual power consumption. It also demonstrates the application
of the steady state optimiser to the process plant, optimising
steady state network production and load distributions.

7.1. Prediction of power consumption

Time (mins)
Validation of piece-wise linear models was conducted on-line Fig. 8. Models of power consumption of ASU u2 over 85 h, where the black trend is
by accessing the actual unit power consumption and comparing the actual power, the red trend is the estimate obtained from the current site
this with each of the models prediction of power consumption specific power model, the green trend is the estimate obtained using the
multivariate nonlinear model, Eq. (9), and the blue trend shows the univariate
(using measured flow and pressure data) to determine the predic-
piece-wise linear model estimation of power. (For interpretation of the references
tion accuracy of the models. The piece-wise linear models are com- to colour in this figure legend, the reader is referred to the web version of this
pared with (a) the current site models, which use a fixed value article.)
specific power (which is the current average efficiency of the com-
pressor or ASU) multiplied by the flow through the machine, and
(b) the best-fit nonlinear multivariate models using Eq. (9). As a estimation of site power consumption when compared to the
representative example of the relative accuracies of the models existing method used (and a comparable accuracy between piece-
of the compressors, Fig. 7 shows the power consumption predic- wise linear and the nonlinear multivariate models). The ME of the
tions for compressor c4 in the network. piece-wise linear compressor models produced is typically equiva-
The results shown cover a period of 85 h of operation where the lent to around ±2% of total compressor power.
three models are used in parallel. Model accuracy was determined Fig. 8 shows the power consumption predictions for ASU u2,
by calculation of mean error (ME) between the actual and esti- with operation across the full range of piece-wise linear models.
mated power consumption at 15 min intervals. Over this period, During this period, the second air compressor feeding u2 was
the ME for the current on-site constant efficiency specific power turned off, resulting in a temporary power increase (observed on
model is 99.81 kW, for the nonlinear multivariate power model the fourth vertical grid line) this is because the first machine is
the ME is 20.57 kW and for the piece-wise linear power model ramped up prior to stopping of the second machine.
the ME is 20.10 kW. Similar results are obtained for the other The results cover the same period of 85 h of operation and the
compressors in the network with around an 80% improvement in model accuracy was again determined by calculation of the ME
at 15 min intervals. Over this period, the ME for the current
Power (kW)

Demand (HCMs)

Time (mins)

Fig. 7. Models of power consumption of compressor c4 over 85 h, where the black Time (hours)
trend is the actual power, the red trend is the estimate obtained from the current
site specific power model, the green trend line is the estimate obtained from the Fig. 9. Customer demand for MP (blue) and HP (red) oxygen gas and metered gas
multivariate nonlinear model, Eq. (9), and the blue trend is the multivariate piece- delivered over a period of 24 h. Three numbers indicate the selected steady state
wise linear model estimation of power, Eq. (12). (For interpretation of the points and real network gas flows are ramped up or down towards the steady state
references to colour in this figure legend, the reader is referred to the web version points over a period of time. (For interpretation of the references to colour in this
of this article.) figure legend, the reader is referred to the web version of this article.)
230 R. Adamson et al. / Applied Energy 189 (2017) 221–232

on-site specific power model is 832.95 kW and for the piece-wise Table 3
linear power model the ME is 322.77 kW. The ME of the piece- Optimised oxygen supply network outputs for steady state point 2 at three different
power price values.
wise linear ASU models produced is typically equivalent to around
±3% of total ASU power. Similar results were obtained for the two Cost function Optimal Power price Off-peak liquid
other ASUs demonstrating that the piece-wise linear models of component configuration event make

power consumption are around 60% more accurate when com- C kW £55/MWh £250/MWh £45/MWh
pared to power consumption estimation using the current on-site F GO;u1 0% 0% 0%
F GO;u2 85.3% 85.3% 61.5%
method. The use of the nonlinear multivariate model would
F GO;u3i 43.9% 43.9% 67.7%
increase the accuracy by a further 10% however; the optimiser F GO;u3ii 70.4% 70.4% 70.4%
should not be fed with pressure data at the current steady state dc1 0 0 0
production flows as pressure is a function of the total oxygen pro- dc2 0 0 0
duction rate. dc3 1 0 1
dc4 1 0 0
dc5 1 0 1
7.2. Power cost optimisation savings dc6 0 0 0
F GO;v 1 0% 100% 0%
F GO;v 2 84% 100% 41.5%
To demonstrate the operating cost savings that are achieved
F GO;s 0% 65.5% 0%
using the steady-state optimiser, steady state demands were opti- F LO;u1 0% 0% 0%
mised using the MILP solver and compared actual network config- F LO;u2 17% 17% 97.7%
urations and flows using process data. Fig. 9 shows the customer F LO;u3 0% 0% 0%
demand of MP and HP over a period of one day and the actual cus- F LN;u3 72.7% 72.7% 72.7%
tomer flows through the site flow meters. Large, temporary
increases in HP supply demonstrate that liquid back up was
required to flow through the meters and boost supply pipeline 7.3. Liquid oxygen as an energy storage device
pressure. Three steady state positions were used to estimate the
potential network power savings. In addition to the cost savings generated by reducing the overall
Table 2 shows a comparison between the actual network con- gas network power consumption for a given customer demand
figuration and ASU production flows and the output of the combination by optimisation, the potential benefits of planned
optimiser at steady state points 1–3, assuming a power price of demand side load management were simulated using the steady
£55/MWh. Binary values represent whether the machine was on state optimiser, subject to variation in power pricing. Optimal
or off and the flow from the network component is presented as steady state point 2 is considered at three power price values, (a)
a percentage of the upper limit. The estimated power consumption the average power price considered for typical steady state optimi-
of the actual network configuration and optimised solution is cal- sation at £55/MWh, (b) the price of power during a typical low
culated and the percentage difference between the two displayed. surplus peak at £250/MWh and (c) off-peak overnight pricing at
Table 2 demonstrates that compression changes are not always £45/MWh.
required to yield power reduction benefits as only steady state For steady state point 2 under average power price conditions,
point 2 involves a compression change over. This is understandable the optimal compression configuration is of three product com-
as the power consumption of the ASUs is significantly larger when pressors running, compressors c3, c4 and c5, with the combined
compared to individual product compressors. For the tested steady power consumption being the flexible load. Running the optimiser
state demands, both the process operator and the optimiser using the peak power pricing value automatically suggests turning
correctly decided ASU u1 was not required. Overall, the steady- off all running product compressors and consuming LO to meet the
state optimiser demonstrates a percentage reduction in power created shortfall in customer demand. In this case, the network is
consumption and ultimately network costs of up to 5% at steady reconfigured to open the cross over valve, v2, for HP to MP GO
state. expansion and the spill valve, v1, to expel LP GO to atmosphere.

Table 2
Recorded and optimised oxygen supply network outputs for steady state points 1–3.

Cost function component Steady state point 1 Steady state point 2 Steady state point 3
Actual Optimiser Actual Optimiser Actual Optimiser
F GO;u1 0% 0% 0% 0% 0% 0%
F GO;u2 56.7% 58.6% 64.3% 85.3% 80.2% 88.0%
F GO;u3i 57.3% 43.9% 85.5% 43.9% 87.8% 79.4%
F GO;u3ii 11% 32.6% 33.3% 70.4% 37.0% 68.9%
dc1 0 0 0 0 0 0
dc2 0 0 0 0 0 0
dc3 1 1 1 1 1 1
dc4 0 0 0 1 1 1
dc5 1 1 1 1 1 1
dc6 0 0 0 0 0 0
F GO;v 1 0% 0% 0% 0% 0% 0%
F GO;v 2 3% 50% 25% 84% 27% 0%
F GO;s 0% 0% 0% 0% 0% 0%
F LO;u1 0% 0% 0% 0% 0% 0%
F LO;u2 17% 17% 17% 17% 17% 17%
F LO;u3 36% 30% 0% 0% 0% 0%
F LN;u3 57.4% 53.2% 92.1% 72.7% 95.8% 100%
Power consumption differential (%) 4.89% 4.87% 3.25%
 R. Adamson et al. / Applied Energy 189 (2017) 221–232 231

ASU production rates and power consumptions are assumed fixed transitions, minimisation of liquid back up consumption, reduction
as temporarily ramping ASUs during peak is not considered safe of LP GO gas spill and minimisation of machine wear due to start-
practise. ing and stopping compressors. Furthermore, it would ensure the
In order to offset the liquid consumed in one hour during the customer order is adhered to during ramps and that all ASU ramp-
peak pricing simulation and determine the benefits of such a pro- ing constraints are adhered to for safety reasons. Large ramps can
cedure, steady state point 2 is optimised to recover the liquid con- take place over several hours with operators required to alter pro-
sumed over three hours at the off-peak power price. The total duction set point targets periodically to meet customer demands at
predicted LO consumed during peak is divided by three and used all times.
in addition to the required minimum liquid production target, i.e.
above what was required for LO purge. Optimisation results in an
Acknowledgments
alternative product compression configuration with an increased
ASU u3 MP GO production flow and a higher ASU u2 LO production
The authors gratefully acknowledge the financial support of
flow to replace the liquid consumed during peak. The cost incurred
EPSRC grant EP/G037620/1 and BOC Gases through the Biopharma-
generating product off peak in additional power consumption and
ceutical Bioprocessing Technology Centre at Newcastle University,
LN injection is calculated to be around 43% of the cost reduction
UK.
achieved during peak avoidance, a significant benefit. The opti-
miser output at each of the various power prices for the customer
demand combination is presented in Table 3. References

[1] Yan L, Yu Y, Li Y, Zhang Z. Energy saving opportunities in an air separation

8. Discussion and conclusions process. In: 12th international symposium on process systems engineering and
25th European symposium on computer aided process engineering; 2010.
It has been shown that piece-wise linear models of network [2] Manenti F, Rovaglio M. Market-driven operational optimization of industrial
gas supply chains. Comput Chem Eng 2013;56:128–41.
component power allow for improved estimation of power con- [3] Üster H, Dilaveroğlu Sß . Optimization for design and operation of natural gas
sumption (when compared to the existing site models) and achieve transmission networks. Appl Energy 2014;133:56–69.
a similar accuracy to nonlinear regression models. The benefit of [4] Azadeh A, Anaraki AZ, et al. Optimisation of multi–stage supply chain systems
by integrated simulation–variable neighbourhood search algorithm. Int J
generating piece-wise linear models is that they can be incorpo- Services Oper Manage 2015;21(1):1–26.
rated into a steady-state optimisation strategy solved using a MILP [5] Gao J, You F. Shale gas supply chain design and operations toward better
approach, a strategy not yet attempted by other authors. It has economic and life cycle environmental performance: MINLP model and global
optimization algorithm. ACS Sustain. Chem. Eng. 2015;3(7):1282–91.
been demonstrated using plant data that the developed optimisa-
[6] Cortinovis A, Mercangöz M, Zovadelli M, Pareschi D, De Marco A, Bittanti S.
tion strategy can reduce site power consumption costs by up to 5% Online performance tracking and load sharing optimization for parallel
during steady state conditions. operation of gas compressors. Comput Chem Eng 2016;88:145–56.
The fidelity of the optimisation scheme is dependent on the [7] Han IS, Han C, Chung CB. Optimization of the air- and gas-supply network of a
chemical plant. Chem Eng Res Des 2004;82(10):1337–43.
accuracy of the power consumption models. Generally, real-time [8] Abbaspour M, Chapman KS, Krishnaswami P. Nonisothermal compressor
optimisation strategies up-date model parameters when new station optimisation. J Energy Resour Technol, Trans ASME 2005;127:131–41.
steady-state conditions are detected. This is not the strategy [9] Paparella F, Domínguez L, Cortinovis A, Mercangöz M, Pareschi D, Bittanti S.
Load sharing optimization of parallel compressors. In: European control
adopted here, rather as with the existing on-site model, the aim conference (ECC), July 17–19, 2013, Zürich, Switzerland; 2013.
is to monitor model accuracy periodically updating the models, [10] Xenos DP, Cicciotti M, Kopanos GM, Bouaswaig AEF, Kahrs O, Martinez-Botas
e.g. on a monthly basis. It is suspected that greater improvements R, et al. Optimization of a network of compressors in parallel: Real Time
Optimization (RTO) of compressors in chemical plants – an industrial case
in model accuracy can be achieved through integration of further study. Appl Energy 2015;114:51–63.
variables known to affect efficiency into the multivariate power [11] Øvervåg TF. Centrifugal compressor load sharing with the use of MPC, master
models, such as ambient conditions [34]. of science in engineering cybernetics. Norwegian University of Science and
Technology; 2013.
In addition to providing energy savings, other benefits of the [12] Han IS, Han C. Modeling of multistage air-compression systems in chemical
steady-state optimiser are that the automated system releases processes. Ind Eng Chem Res 2003;42(10):2209–18.
operator time for other tasks, such as improving overall site control [13] Huang R, Zavala VM, Biegler LT. Advanced step nonlinear model predictive
control for air separation units. J Process Control 2009;19(4):678–85.
and aiding network configuration manipulation from the current
[14] Puranik Y, Kilinç M, Sahinidis NV, Li T, Gopalakrishnan A, Besancon B, et al.
network position to the optimal steady state end point. Further- Global optimization of an industrial gas network operation. AIChE J 2016;62
more, it could be used to provide decisions as to whether it is (9):3215–24.
cheaper to consume liquid and stop compressors during high [15] Cao Y, Swartz CLE, Baldea M. Design for dynamic performance: application to
an air separation unit. In: Proceedings of the 2011 American control
power price periods. Including forecasts of spot market power pric- conference; 2011.
ing, a schedule for liquid oxygen production can be developed to [16] Kopanos GM, Xenos DP, Cicciotti M, Pistikopoulos EN, Thornhill NF.
allow release of liquid inventory during periods of high power or Optimization of a network of compressors in parallel: operational and
maintenance planning – the air separation plant case. Appl Energy
non-commodity costs and production during low pricing periods. 2015;146:453–70.
In some cases, spilling gas or consuming liquid can be used to deli- [17] Adamson R, Hobbs M, Silcock A, Montague G. Real time optimisation of
ver an overall network cost benefit by avoiding consuming power industrial gas supply networks. IFAC-PapersOnLine 2015;48(8):355–60.
[18] Lin M-H, Carlsson JG, Ge D, Shi J, Tsai J-F. A review of piecewise linearization
at peak pricing (non-dispatchable activities) or by generating rev- methods. Math Prob Eng 2013;2013.
enues from frequency or demand side response activities (dis- [19] Martin A, Möller M, Moritz S. Mixed integer models for the stationary case of
patchable activities) as suggested by Merkert et al. [25]. This may gas network optimization. Math Program 2006;105(2–3):563–82.
[20] Kolb O, Lang J, Bales P. Adaptive linearized models for optimization of gas
also be useful for oxygen requests where a larger, under-loaded networks. Proc Appl Math Mech 2007;7:1061301–02.
and less efficient compressor is required instead of another. Steady [21] Camponogara E, de Castro MP, Plucenio A, Pagano DJ. Compressor scheduling
state optimisation cannot, however, automatically and optimally in oil fields. Optim Eng 2011;12(1):153–74.
[22] Aguiar MAS, Camponogara E, Silva TL. A Mixed-Integer convex formulation for
schedule over periods of lower power pricing to recover those
production optimization of gas-lifted oil fields with routing and pressure
stocks by producing liquid at other times but this can be achieved constraints. Braz J Chem Eng 2014;31(2):439–55.
either by introducing a liquid production target schedule or by [23] Domschke P, Geißler B, Oliver K, Jens L, Alexander M, Antonio M. Combination
developing a dynamic optimisation method [2]. of nonlinear and linear optimization of transient gas networks. INFORMS J
Comput 2011;23(4):605–17.
Dynamic optimisation could deliver considerable further bene- [24] Correa-Posadaa CM, Sánchez-Martın P. Gas network optimization: a
fits including reduction of power consumption during network comparison of piecewise linear models. Optimization Online; 2014.
232 R. Adamson et al. / Applied Energy 189 (2017) 221–232

[25] Merkert L, Harjunkoski I, Isaksson A, Säynevirta S, Saarela A, Sand G. [29] Li T, Roba T, Bastid M, Prabhu A. Real time optimization of air separation
Scheduling and energy – industrial challenges and opportunities. Comput plants. Proceedings of ISA Automation Week; 2011.
Chem Eng 2014;72:183–98. [30] Love J. Process automation handbook: a guide to theory and practice. Springer
[26] Karagiannopoulos S, Vrettos E, Andersson G, Zima M. Scheduling and real-time Sci & Bus Media 2007:849–99.
control of flexible loads and storage in electricity markets under uncertainty. [31] Kurz R, Winkelmann B, Mokhatab S. Efficiency and operating characteristics of
In: 11th international conference on the European energy market, Krakow; centrifugal and reciprocating compressors. Pipeline Gas J 2010;237(10).
2014. p. 1–5. [32] Yang L, Liua S, Tsokab S, Papageorgiou LG. Mathematical programming for
[27] Zhang Q, Grossmann IE, Heuberger CF, Sundaramoorthy A, Pinto JM. Air piecewise linear regression analysis. Expert Syst Appl 2016;44:156–67.
separation with cryogenic energy storage: Optimal scheduling considering [33] Szucs A, Kvasnica M, Fikar M. Optimal piecewise affine approximations of
electric energy and reserve markets. Process Syst Eng 2015;61(5):1547–58. nonlinear functions obtained from measurements. IFAC Proceedings Volumes
[28] Xenos DP, Noor IM, Matloubi M, Cicciotti M, Haugen T, Thornhill NF. Demand- (IFAC-PapersOnline); 2012.
side management and optimal operation of industrial electricity consumers: [34] Fu C, Gundersen T. Using exergy analysis to reduce power consumption in air
an example of an energy-intensive chemical plant. Appl Energy separation units for oxy-combustion processes. Energy 2012;44:60–8.
2016;182:418–33.