You are on page 1of 24

Scheduling of Crude Oil Operations Under Demand

Uncertainty: A Robust Optimization Framework Coupled


with Global Optimization
Jie Li, Ruth Misener and Christodoulos A. Floudas
Dept. of Chemical and Biological Engineering, Princeton University, Princeton, NJ 08544

DOI 10.1002/aic.12772
Published online October 28, 2011 in Wiley Online Library (wileyonlinelibrary.com).

Scheduling of crude oil operations is an important component of overall refinery operations, because crude oil costs
account for about 80% of the refinery turnover. The mathematical modeling of blending different crudes in storage tanks
results in many bilinear terms, which transform the problem into a challenging, nonconvex, mixed-integer nonlinear
programming (MINLP) optimization model. In practice, uncertainties are unavoidable and include demand fluctuations,
ship arrival delays, equipment malfunction, and tank unavailability. In the presence of these uncertainties, an optimal
schedule generated using nominal parameter values may often be suboptimal or even become infeasible. In this article, the
robust optimization framework proposed by Lin et al. and Janak et al. is extended to develop a deterministic robust
counterpart optimization model for demand uncertainty. The recently proposed branch and bound global optimization
algorithm with piecewise-linear underestimation of bilinear terms by Li et al. is also extended to solve the nonconvex
MINLP deterministic robust counterpart optimization model and generate robust schedules. Two examples are used to
illustrate the capability of the proposed robust optimization approach, and the extended branch and bound global
optimization algorithm for demand uncertainty. The computational results demonstrate that the obtained schedules are
robust in the presence of demand uncertainty. V
C 2011 American Institute of Chemical Engineers AIChE J, 58: 2373–2396,

2012
Keywords: refinery, crude oil scheduling, mixed-integer nonlinear programming, nonconvex, global optimality, robust
optimization

Introduction minimizing crude changeovers, avoiding ship demurrage,


In recent years, refineries have to exploit all potential and managing crude inventories. The presence of crude
cost-saving alternatives because of their intense competition blending gives rise to bilinear terms in the mathematical for-
arising from fluctuating product demands, ever-changing mulation for scheduling, while discrete scheduling decisions
crude prices, and strict environmental regulations. Schedul- such as selecting a tank to unload or feed and the often com-
ing of crude oil operations is a critical component of the plex nonlinear nature of crude properties and qualities make
overall refinery operations,1–3 because crude oil costs can such a model challenging, nonlinear, nonconvex mixed-inte-
account for about 80% of the refinery turnover.4 Most refin- ger nonlinear problem (MINLP).
eries blend premium crudes with low-quality crudes over The crude oil scheduling problem has received consider-
time to exploit the higher profit margins of low-quality able attention with researchers developing different models
crudes. However, the low-cost crudes can lead to processing based on discrete- and continuous-time representations.3 Li
problems in crude distillation units (or CDUs) and down- et al.3 recently developed a novel unit-specific event-based
stream units, because they usually contain some less-than-de- continuous-time5–21 MINLP formulation for this problem.
sirable properties with high composition. Therefore, a key They incorporated many realistic operational features such
issue is to exploit blends of low-cost crudes and premium as single buoy mooring (SBM), multiple jetties, multiparcel
crudes to maximize profit margins and minimize the opera- vessels, single-parcel vessels, crude blending, brine settling,
tional problems at the same time. Optimal scheduling of crude segregation, and multiple tanks feeding one CDU at
crude oil operations using advanced mathematical optimiza- one time and vice versa. In addition, 15 important volume-
tion techniques such as mixed-integer linear programming based or weight-based crude property indices were consid-
(MILP) can increase profit margins using cheaper crudes, ered. To address this nonconvex MINLP problem, they
exploited recent advances in piecewise-linear underestima-
Additional Supporting Information may be found in the online version of this
article. tion of bilinear terms22–32 within a branch and bound algo-
Correspondence concerning this article should be addressed to C. A. Floudas at rithm for global optimization. The proposed model signifi-
floudas@titan.princeton.edu.
cantly reduced the number of bilinear terms and problem
V
C 2011 American Institute of Chemical Engineers size compared to the discrete-time formulation of Reddy

AIChE Journal August 2012 Vol. 58, No. 8 2373


et al.33 and Li et al.34 The computational results showed that separately. In their model, uncertain product demands were
the proposed branch and bound global optimization algo- represented by chance-constrained programming and fuzzy
rithm with piecewise-linear underestimation22–32 of the bilin- programming, and ship arrival delay was represented by a
ear terms was effective to address all tested examples and scenario approach. Although their model can cope with a
resulted in better integer feasible solutions. More impor- wide variety of uncertainties, the generated schedule from
tantly, these integer feasible solutions were guaranteed to be their model also results in composition discrepancy.
within 2% of global optimality. The above approaches cannot ensure the generated solu-
All of the aforementioned models assume that the parame- tion to be feasible for the nominal parameters. To overcome
ters used in the models are deterministic in nature. However, this disadvantage, Ben-Tal, Nemirovski, and coworkers42–45
frequent uncertainties in practice are unavoidable such as developed the robust optimization framework to handle
demand fluctuations, ship arrival delays, crude quality speci- uncertain parameters within linear and quadratic program-
ming problems. A similar robust optimization framework
fication variations, uncertainty on crude profit margin, de-
was also independently proposed by Ghaoui and co-
murrage cost, inventory cost, changeover cost and safety
workers.46,47 In their framework, various forms of parameter
stock penalty, and equipment malfunction, and tank unavail- uncertainties are explicitly addressed and the robust solution
ability. In the presence of these uncertainties, an optimal is guaranteed to be feasible for the nominal parameters. Lin
crude schedule obtained using nominal parameter values et al.,48 Janak et al.,49 Verderame and Floudas,50–52 and Li
may often be suboptimal or even become infeasible. Differ- et al.53 extended and developed the theory of the robust opti-
ent methodologies can be used to address this problem. In mization framework for general MILP problems with
general, there are two approaches to address uncertainties: bounded, bounded and symmetric, and several known proba-
reactive scheduling and preventive scheduling.35 Reactive bility distributions. Also, Bertsimas and coworkers54–56
scheduling is a process to revise the generated schedule from extended and applied a robust optimization framework for
nominal parameters when a disruption has occurred during linear and discrete programming. Recently, the robust opti-
the actual execution of the schedule. Equipment malfunction mization framework has been successfully extended and
and tank unavailability are events which are best modeled applied to addressing demand due date and demand amount
after realization by means of reactive scheduling techniques. uncertainty50 in the problem of operational planning of
Because of the ‘‘on-line’’ nature of reactive scheduling, an large-scale industrial bath plants,57 demand and processing
time uncertainty51 in the problem of integration of opera-
updated schedule from the nominal schedule must be gener-
tional planning and medium-term scheduling for large-scale
ated in a timely manner and hence heuristic approaches are
industrial batch plants,57 and demand and transportation
often utilized.36,37 Preventive scheduling seeks to accommo- uncertainty52 in multisite planning problem.56 In addition,
date future uncertainty at the scheduling stage. The uncer- they applied an alternative framework based on conditional
tainty related to demand, ship arrival time, crude specifica- value-at-risk theory to address demand due date and demand
tion, crude profit margin, demurrage cost, inventory cost, amount uncertainty58 in the operational planning of batch
changeover cost, and safety stock penalty can be explicitly processes,44 and demand and transportation uncertainty in
taken into account through preventive approaches such as the operational planning of multisite batch plants.59
two-stage stochastic programming, parametric programming, To the best of our knowledge, the robust optimization frame-
fuzzy programming, chance constraint programming, robust work has not yet been extended and applied for crude oil sched-
optimization techniques, and risk mitigation techniques.35 uling operations under uncertainty. In this article, we address
Among these approaches, uncertain parameters are often rep- the crude oil scheduling problem described by Li et al.3 for a
resented by scenarios or nonscenarios. For detailed reviews typical marine-access refinery under demand uncertainty. The
on planning and scheduling under uncertainty, the reader is unit-specific event-based continuous-time formulation devel-
directed to Li and Ierapetritou,38 and Verderame et al.35 oped by Li et al.3 is used as the basis. The theory of robust opti-
Robust optimization focuses on developing preventive mod- mization framework is used to develop the robust counterpart
els to minimize the effects of uncertainties on the perform- optimization model where a new approach is proposed to con-
ance measure such as profit and operating cost. Its main vert demand equality constraints to inequalities. Then, the
objective is to ensure that the generated solutions are robust, branch and bound global optimization algorithm from Li et al.3
while maintaining a high level of solution quality. Li et al.39 is extended to solve the proposed deterministic robust counter-
developed scenario-based models for demand and ship arrival part optimization model and generate robust schedules. The
uncertainties separately and obtained more robust schedules computational results show that the schedule obtained from the
compared to the nominal schedules. However, the number of proposed deterministic robust counterpart is robust in the pres-
scenarios exponentially increases with the number of uncer- ence of demand uncertainty.
tain parameters and hence makes their model intractable for
practical problems with large number of uncertain parameters. Problem Statement
Cao et al.40 proposed an optimization model based on chance- Consider Figure 1, which shows a schematic of crude oil
constrained programming to generate robust schedules under unloading, storage and processing in a typical marine-access
demand uncertainty during scheduling of crude oil operations. refinery. It involves offshore facilities for crude unloading
Their approaches avoided the drawback of enumerating sce- such as a SBM station, onshore facilities for crude unloading
narios. However, their approach cannot be used to deal with such as B jetties, I (i ¼ 1, 2, 3,…,I) crude storage tanks, and
uncertain parameters following a discrete probability distribu- U (u ¼ 1, 2, 3,…,U) CDUs. The pipeline connecting the
tion.41 More importantly, their approach results in composi- SBM station with crude tanks is called the SBM line, and it
tion discrepancy. Recently, Wang and Rong41 developed a normally has a substantial holdup. In this study, we assume
two-stage robust optimization model for crude oil scheduling that the refinery has no separate charging tanks, and hence
problem to address demand and ship arrival uncertainty crude storage tanks also act as charging tanks. Very large

2374 DOI 10.1002/aic Published on behalf of the AIChE August 2012 Vol. 58, No. 8 AIChE Journal
2. All jetties are identical.
3. Holdup of the jetty pipeline is negligible.
4. Crude mixing is perfect in each storage tank.
5. Crude changeover times are negligible.
6. During operation, CDUs never shut down.
The objective is to maximize the gross profit, which is the
revenue computed in terms of crude margins minus the oper-
ating costs such as demurrage and safety stock penalties.
During the above crude oil scheduling operations, frequent
uncertainties are unavoidable such as demand fluctuations
and ship arrival delays. The most common uncertainty arises
from: (1) demand, (2) ship arrival, (3) crude quality specifi-
cations, and (4) some economic coefficients. These uncertain
parameters can be described using discrete or continuous dis-
tributions. In some cases, only limited knowledge about the
distribution is available, for example, the uncertainty is
bounded, or the uncertainty is symmetrically distributed in a
Figure 1. Schematic of crude oil unloading, blending,
certain range. In the best situation, the distribution function
and processing.
for the uncertain parameter is given, for instance, as a nor-
[Color figure can be viewed in the online issue, which is
mal distribution with known mean and standard deviation. In
available at wileyonlinelibrary.com.]
this article, we focus on demand uncertainty which is consid-
ered as: (a) bounded, (b) symmetrical and bounded, (c) fol-
crude carriers (VLCCs) carrying multiparcels can dock at lowing a known distribution such as normal distribution, or
the SBM station and unload crudes into storage tanks. Sin- (d) following an unknown probability distribution. It should
gle-parcel vessels carrying single crude for each can berth at be noted that the random demand parameters considered in
jetties and unload crudes into storage tanks. Different types this article are assumed to be fully independent.
of crudes C (c ¼ 1, 2, 3,…,C) can be allowed to blend in
these crude storage tanks. After blending, they are fed into Deterministic Mathematical Formulation for
CDUs for processing. Scheduling of Crude Oil Operations and Branch
Given: and Bound Global Optimization Algorithm
1. V ships, their expected arrival times, their crude par- The model of Li et al.3 was developed based on unit-spe-
cels, and parcel sizes; cific event-based continuous-time representation,5–21 which
2. B Jetties, jetty-tank and SBM-tank connections, crude is different from other variants such as discrete-time,5,9 pro-
unloading transfer rates, and SBM pipeline holdup volume cess slots,2 and unit slots.60–62 The differences among those
and its resident crude; time representations are discussed by Floudas and Lin5,6 and
3. I storage tanks, their capacities, their initial crude vol- Li et al.63 In the model of Li et al.,3 they defined jetties,
umes and compositions, and crude quality specifications or storage tanks (i), and CDUs (u) as units (m), and treated all
limits; identical jetties as one single resource. For each unit m, the
4. U CDUs, their processing rates, and crude quality scheduling horizon [0, H] is divided into N (n ¼ 1, 2,…,N)
specifications or limits; event points (Figure 2). They also defined two binary varia-
5. Scheduling horizon H and product demands; bles to denote parcel-to-tank and tank-to-CDU connections,
6. Economic data: crude margins, demurrage, crude respectively.
changeover costs, and safety stock penalties.
8
Determine: <1 if parcel p is unloaded to
1. Unloading schedule for each ship including the tim- Xðp; i; nÞ ¼ tank i during event point n 8ðp; iÞ 2 SP;I
ings, rates, and tanks for all parcel transfers; :
0 otherwise
2. Inventory and crude concentration profiles of all stor-
age tanks;
3. Charging schedule for each CDU including the feed
tanks, feed rates, and timings.
Subject to the operating practices:
1. Only one VLCC can dock at the SBM station at a time.
2. The unloading sequence of VLCC parcels is known a
priori.
3. A parcel can unload to at most one storage tank at any
time, but may unload to multiple tanks over time.
4. Each tank needs 8 h to settle and remove brine after
each crude receipt.
5. A storage tank cannot receive and feed simultaneously.
6. Multiple tanks can feed a CDU simultaneously, and
vice versa.
Assumptions:
1. Only one crude resides in the SBM line at the end of
each parcel transfer. Crude flow is plug flow in the SBM. Figure 2. Event points definition for each unit.

AIChE Journal August 2012 Vol. 58, No. 8 Published on behalf of the AIChE DOI 10.1002/aic 2375
Table 1. Data for Example 1
Tanker Arrival Time Parcel No: (Crude, Parcel Size, kbbl)
VLCC-1 0 1: (C2, 10) 2: (C1, 300) 3: (C4, 300) 4: (C3, 340)
Initial Crude
Composition (kbbl) Crude Concentration Range in Tanks
Tank Initial Inventory (kbbl) Capacity (kbbl) Heel (kbbl) C1 or C3 C2 or C4 C1 or C3 Min-Max C2 or C4 Min-Max
T1 300 700 50 200 100 0–1 0–1
T2 300 700 50 100 200 0–1 0–1
T3 200 700 50 50 150 0–1 0–1
T4 300 700 50 130 170 0–1 0–1
T5 80 700 50 50 30 0–1 0–1

Range of Crude Concentration for CDUs


C1 C2 C3 C4 Key Comp.Range Range of Demand/8 h Demand
CDU Min–Max Min–Max Min–Max Min–Max Min–Max Min–Max (kbbl)
CDU1 0–1 0–1 0–0 0–0 0.0045–0.006 50–100 600
CDU2 0–0 0–0 0–1 0–1 0.014–0.0153 50–100 600

Flow Rate Limit (kbbl/8 h)


Parcel-Tank Tank-CDU Demurrage Changeover Safe Inventory Concentration of Margin
Min–Max Min–Max Cost (k$/8 h) Loss (k$/instance) Penalty ($/bbl/8 h) Crude Key Composition ($/bbl)
10–400 0–100 100 5 0.2 C1 0.005 3
Tanks 1, 4 store crude 1–2 (Class 1); 2–3, 5 store crude 3–4 (Class 2) C2 0.006 4.5
CDU 1 processes crudes 1–2; 2 processes crudes 3–4 C3 0.0165 5
The desire safety stock is 1200 kbbl C4 0.0145 6

8 tightening LB and UB, and so on. The details about these


<1 if tank i feeds CDU u
Yði; u; nÞ ¼ during event point n 8ði; uÞ 2 SI;U strategies can also be found in Li et al.3
:
0 otherwise Motivating example
Consider Example 1 involving one SBM pipeline, five
where, SP,I ¼ {(p, i)| parcel p that can be unloaded to tank i} storage tanks (T1–T5), and two CDUs (CDU101–CDU102).
and SI,U ¼ {(i, u)| tank i that can feed CDU u}. One VLCC carrying three crude parcels (300 kbbl C1, 300
The complete nonconvex MINLP model is presented in kbbl C4, and 350 kbbl C3, unloaded in that sequence)
Appendix A. For a detailed discussion of the model, the arrives at time zero. At time zero, the SBM pipeline is
reader is referred to Li et al.3 holding 10 kbbl C2 from the last parcel. The scheduling
horizon is about 72 h (i.e., 3 days). The nominal demands
Basic Components of Branch and Bound Global for both CDUs are 600 kbbl. Table 1 presents the complete
Optimization Algorithm data.
At each node in the branch and bound tree, a piecewise-lin- We solve this example with the model of Li et al.3 and
ear relaxation of the node is minimized, and the node is the proposed branch and bound global optimization algo-
branched to create two child nodes. After solving the piece- rithm using GAMS 22.6/CPLEX 11.0.0 on Dell OPTIPLEX
TM
wise-linear relaxation, a pool of feasible solutions (denoted as 960 of IntelVR Xeon CPU 3.0 GHz with 2 GB RAM run-
Pool-1) including the final solve (which is the best or optimal ning Linux. The computational performance is given in
integer solution for the relaxation) is obtained. The lower Table 2. The best solution of $ 5631.707K was obtained
bound (LB) is updated with this final solve, if the final solve within 15.3 CPU seconds, and it is guaranteed to be within
is greater than the current LB. Each solution from the Pool-1
is used to fix the current values of the binary variables, initial- Table 2. Model and Solution Statistics for Example 1
ize the continuous variables, and locally minimize the result-
Robust Solution Robust Solution
ing NLP. All generated feasible local optimal solutions from Nominal for Bounded for Normal
another pool denoted as Pool-2. If the smallest objective value Solution Uncertainty Distribution
in the Pool-2 is less than the current upper bound (UB), then Event points 3 3 3
UB is updated with this value. At each step, the nodes with GR 7 7 7
relaxations within a predetermined tolerance (denoted as e) of Binary variables 45 45 45
the current UB are eliminated. The algorithm terminates with Continuous variables 753 1007 1007
e-convergence. Comprehensive coverage of the branch and Constraints 1159 1825 1825
Bilinear terms 70 210 210
bound algorithms can be found in the textbooks of Flou- Obj (UB, K$) 5631.707 5614.974 5614.974
das.64,65 Appendix B presents an outline of the strategies that LB (K$) 5687.152 5671.139 5667.320
are used in the branch and bound global optimization algo- Gap (%) 0.97 0.99 0.92
rithm, which include piecewise-linear underestimators, branch- CPU time (s) 15.3 72.7 72.7
ing strategy, solution improvement strategy, optimality-based Y ¼ 0.05, e ¼ 0.01.

2376 DOI 10.1002/aic Published on behalf of the AIChE August 2012 Vol. 58, No. 8 AIChE Journal
Figure 3. Nominal schedule for Example 1.

1% of global optimality. The operational schedule (i.e., nom- inventory level of T3 at t ¼ 72 h reaches its minimum
inal schedule) is illustrated in Figure 3. Figures 4 and 5 capacity (50 kbbl) from Figure 4. Thus, T3 cannot feed
illustrate inventory profiles of tanks T1–T5 and feed rates to more crudes to CDU102. T2 can feed (12.56.25)  24
CDUs CDU101–102 from T1–T5. kbbl ¼ 150 kbbl additional crudes to CDU102. On the
Now let us consider an increase in the demands for both other hand, the inventory level of T2 at t ¼ 40.2 h is 150
CDUs from the nominal value of 600 kbbl to 705 kbbl. We kbbl, which indicates that T2 can feed maximum 100 kbbl
freeze the unloading schedule, tank-to-CDU connections, additional crudes to CDU102 during [16.2, 40.2] h. Hence,
start and end times of tank feeding CDU operations from the maximum demand of CDU102 can reach 700 kbbl. In
Figure 3 and solve the model again with the new demands a brief, we cannot simply increase the feed rates to
(i.e., 705 kbbs for both CDUs). However, we cannot gener- CDU102 from tanks T2, and T3 to meet its new demand
ate a feasible schedule. This reason is analyzed as follows. of 705 kbbl.
From Figure 3, T2 and T3 feed CDU102, and T5 does not. If we do not freeze the unloading schedule, tank to CDU
Although the feed rates to CDU102 do not reach its maxi- connections and those start and end times for tank feeding
mum feed rate (12.5 kbbl/s) at any time from Figure 5, the CDU operations from Figure 3, and the demands of

Figure 4. Inventory profiles of T1–T5 for Example 1 from the nominal schedule in Figure 3.
[Color figure can be viewed in the online issue, which is available at wileyonlinelibrary.com.]

AIChE Journal August 2012 Vol. 58, No. 8 Published on behalf of the AIChE DOI 10.1002/aic 2377
Min=Max cT x þ d T y
x;y

s:t: Ex þ Fy ¼ e
Ax þ By  p (1)
x  x  x
y ¼ 0; 1
Assume that the left-hand-side coefficients A and B and
the right-hand-side parameters p of the inequality constraints
are uncertain parameters. The true realization of an uncertain
parameter is represented by

a~ ¼ ð1 þ e  nÞ  a (2)
where a is an uncertain parameter with a~ being its true
realization, e represents a give (relative) uncertain level, and n
stands for a random variable.
In the robust optimization framework, a solution (x, y) is
called robust if (1) (x, y) is feasible for the nominal problem,
(2) whatever are the true values of the coefficients and right-
hand-side parameters, (x, y) must satisfy the lth inequality
constraint with an error of at most d max[1, |pl|], where d is
Figure 5. Feed rates of T1–T5 to CDU101 and CDU102 a given infeasibility tolerance. Then, any inequality con-
for Example 1 from the nominal schedule in straint l in Eq. 1 becomes
Figure 3.
X X X
[Color figure can be viewed in the online issue, which is alm xm þ a~lm xm þ blk yk
available at wileyonlinelibrary.com.]
m62Ml m2Ml k62Kl
X
þ b~lk yk  p~l þ d max½1; jpl j 8l ð3Þ
k2Kl
CDU101 and CDU102 increase to 705 kbbl, then another
where Ml and Kl are the subsets that encompass those
new schedule is generated, which is illustrated in Figure 6.
uncertain parameters for the given constraint l.
The new schedule in Figure 6 is much different from that in
Figure 3. The differences may confuse the refiners and dis- Bounded uncertainty
rupt plant operation. In practice, it is desirable to keep the Assume that the uncertain parameters vary in a bounded
plant operation as close as possible to the nominal schedule interval. They are represented by
(Figure 3). This example highlights the importance of gener-
ating reliable, efficient, and robust schedules and hence moti- ja~lm  alm j  ejalm j
vates us to develop systematic and effective techniques that  
b~lk  blk   ejblk j (4)
generate robust schedules which may accommodate demand
uncertainty. jp~l  pl j  ejpl j
Overview of the Robust Optimization Framework
The deterministic robust counterpart for Eq. 3 is derived
The robust optimization framework developed by Lin
as follows
et al.,48 Janak et al.,49 Verderame and Floudas,50–52 and Li
et al.53 is used to address the various forms of uncertainty in !
X X X X
this article. In the following, we present in brief the general alm xm þ blk yk þ e jalm jum þ jblk jyk
robust optimization approach, which explicitly takes into m k m2Ml k2Kl (5)
account the various forms of parameter uncertainty within
 pl  ejpl j þ d max½1; jpl j 8l
constraints and/or objective function.
Consider a generic deterministic MILP problem where um  xm  um.

Figure 6. Operational schedule for Example 1 with demands of 705 kbbl.

2378 DOI 10.1002/aic Published on behalf of the AIChE August 2012 Vol. 58, No. 8 AIChE Journal
The deterministic robust counterpart of the original uncer- where X is a positive parameter with j ¼ exp (X2/2).
tain MILP problem can be derived as follows
Min=Max cT x þ d T y Known probability distribution
x;y;u
The true values of the uncertain parameters are obtained
s:t: Ex þ Fy ¼ e
from their nominal values by random variables
Ax þ By  p
!
X X X X a~lm ¼ ð1 þ enlm Þalm
alm xm þ blk yk þ e jalm jum þ jblk jyk (6)
m k m2Ml k2Kl b~lk ¼ ð1 þ enlk Þblk (9)
 pl  ejpl j þ d max½1; jpl j 8l p~l ¼ ð1 þ enl Þpl
 u m  xm  u m 8m 2 Ml
x  x  x
Assume that the probability distributions of the random
y ¼ 0; 1 8k variables nlm, nlk, and nl are known. Then, Eq. 3 becomes
Bounded and symmetric uncertainty
Assume that the uncertain parameters are distributed X X X
around their nominal values randomly and symmetrically alm xm þ alm ð1 þ enlm Þxm þ blk yk
represented as follows m62Ml m2Ml k62Kl
X
þ blk ð1 þ enlk Þyk  pl ð1 þ enl Þ þ d max½1; jpl j 8l ð10Þ
a~lm ¼ ð1 þ enlm Þalm ; b~lk ¼ ð1 þ enlk Þblk ;
p~l ¼ ð1 þ enl Þpl k2Kl
(7)
where nlm, nlk, and nl are random variables distributed
Eq. 10 is assumed to be violated by the user-specified pa-
symmetrically in the interval [1, 1].
rameter j (0  j  1). In other words
The deterministic robust counterpart for Eq. 3 is derived
as follows 2 (
X X X X X X X
alm xm þ blk yk þ e4 jalm jum þ jblk jwlk Pr alm xm þ blk yk  pl þ e alm nlm xm
m k m2Ml k2Kl m k m2Ml
3 ! )
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
ffi X
X X þ blk nlk yk  pl nl > d max½1; jpl j j 8l ð11Þ
þjpl ju0l þ X a2lm z2lm þ b2lk v2lk þ p2l z02l 5
k2Kl
m2Ml k2Kl

 pl þ d max½1; jpl j 8l ð8Þ


where Note that if the tolerance for the constraint violation is
low, then the value of j should be similarly small.
 ulm  xm  zlm  ulm 8 m 2 Ml We aggregate the respective random variables into one
 wlk  yk  vlk  wlk 8 k 2 Kl random variable.
 u0t  1 þ z0t  u0t 8l X X
nl ¼ alm nlm xm þ blk nlk yk  pl nl 8l (12)
The deterministic robust counterpart of the original uncer- m2Ml k2Kl
tain MILP problem can be derived as follows

Min=Max cT x þ d T y We define Fnl as the cumulative distribution for the aggre-


x;y;u;z;v;w;z0;u0 gate random variable nl, and F1
nl to be inverse cumulative
s:t: Ex þ Fy ¼ e distribution.
Ax þ By  p Fnl ðkÞ ¼ Prfnl  kg ¼ 1  j 8l (13)
2
X X X X
alm xm þ blk yk þ e4 jalm julm þ jblk jwlk F1
nl ð1  jÞ ¼ f ðk; jalm jxm ; jblk jyk ; jpl jÞ 8l (14)
m k m2Ml k2Kl
3
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
X X ffi Combing Eqs. 12–14, the generic deterministic robust
þjpl ju0l þ X a2lm z2lm þ b2lk v2lk þ p2l z02l 5 counterpart of the probabilistic constraint can be formu-
m2Ml k2Kl lated as follows for random variables following any
distribution.
 pl þ d max½1; jpl j 8l
X X
ulm  xm  zlm  um 8m 2 Ml alm xm þ blk yk  pl þ e  f ðk; jalm jxm ; jblk jyk ; pl Þ
wlk  yk  vlk  wlk 8k 2 Kl m k (15)
u0l  1 þ z0l  u0l 8l  d max½1; jpl j 8l
x  x  x
The deterministic robust optimization model for the origi-
yk ¼ 0; 1 8k nal uncertain MILP problem can be derived as follows

AIChE Journal August 2012 Vol. 58, No. 8 Published on behalf of the AIChE DOI 10.1002/aic 2379
 
Min=Max cT x þ d T y P P
 max 0;2j pli  pli
s:t: Ex þ Fy ¼ e P P P  j, then the deter-
i2Il i2Il
If let  alm xm  blk yk þ pli þd max½1;jpl j
Ax þ By  p
X X i62Il
m k

alm xm þ blk yk  pl þ e  f ðk; jalm jxm ; jblk jyk ; pl Þ ministic robust counterpart for Eq. 16 is represented by
m k
 d max½1; jpl j 8l
x  x  x X X X  
2j  1
yk ¼ 0; 1 8k alm xm þ blk yk  pli þ max 0; 
m k i62Il
j
X (20)
Unknown probability distribution pli þ d max½1; jpl j 8l
In this case, the right-hand side parameters in Eq. (1) are i2Il

uncertain and follow an unknown probability distribution, i.e.


X X X X The deterministic robust counterpart optimization model
alm xm þ blk yk  p~li þ pli þ d max½1; jpl j 8l for the original uncertain MILP problem can be derived as
m k i2Il i62Il
follows
P (16)
where pl ¼ i pli :
Its probabilistic form with constraint violation of the user- Min=Max cT x þ d T y
specified parameter j is represented by s:t: Ex þ Fy ¼ e
( Ax þ By  p
X X X  
Pr alm xm þ blk yk > p~li X X X 2j  1
m k i2Il alm xm þ blk yk  pli þ max 0;
) m k i62Il
j
X X
þ pli þ d max½1; jpl j j 8l  pli þ d max½1; jpl j 8l
i62Il i2Il
(
X X X x  x  x
) Pr  p~li >  alm xm  blk yk
m
yk ¼ 0; 1 8k
i2Il k
)
X
þ pli þ d max½1; jpl j j 8l ð17Þ
i62Il Up to this point, the robust optimization framework has
been reviewed for uncertain parameters following a bounded,
From Markov Inequality,50 the left-hand side of Eq. 17 is bounded and symmetric distribution, known probability dis-
( tribution, and unknown probability distribution. For simplic-
X X X
Pr  p~li >  alm xm  blk yk ity, universal values for e, j, and d are adopted. However,
i2Il m k Janak et al.49 highlighted that the robust optimization frame-
) work can easily be extended for cases where e can vary
X
þ pli þ d max½1; jpl j from parameter to parameter and j, and d can vary from
i62Il constraint to constraint. Next, we extend and apply the ro-
" # bust optimization framework to address the various forms of
P
E p~li demand uncertainty present during crude oil scheduling
i2Il operations.
 P P P ð18Þ
 alm xm  blk yk þ pli þ d max½1; jpl j
m k i62Il Robust Counterpart for Demand Uncertainty
during Scheduling of Crude Oil Operations
If the unknown distribution is approximately symmetric In the deterministic model of Li et al.,3 the deterministic
and the extremes are equidistant from the mean, for a given demand constraint is presented as follows
value of j
" # " #
X X X X
E p~li  max 0; 2j  pli  pli (19) VU ðu; nÞ ¼ DðuÞ 8u (21)
i2Il i2Il i2Il n

Combine Eqs. 18 and 19


( ) where VU(u,n) is total amount of crudes fed to CDU u during
X X X X
Pr  p~li >  alm xm  blk yk þ pli þ d max½1; jpl j event point n and D(u) is the nominal demand of CDU u. Note
i2Il m k i62Il that Eq. 21 features equalities. To apply the robust optimiza-
tion framework, Eq. 21 must be converted to inequality.
" # We notice that Eq. 21 is equivalent to the following two
P P constraints:
 max 0; 2j  pli  pli
 P P
i2Il
P
i2Il X
 alm xm  blk yk þ pli þ d max½1; jpl j VU ðu; nÞ  DðuÞ 8u (22)
m k i62Il n

2380 DOI 10.1002/aic Published on behalf of the AIChE August 2012 Vol. 58, No. 8 AIChE Journal
X
VU ðu; nÞ  DðuÞ 8u (23) Note that X is a positive parameter with j ¼ exp (X2/2)
n and the violation probability is 2j.

If we apply the robust optimization techniques to Eqs. 22 Known probability distribution


and 23 simultaneously, then it is not difficult to show that The generic robust deterministic counterpart of the proba-
we cannot generate any feasible solution to meet those ro- bilistic constraint can be formulated as follows for random
bust counterparts and Eq. 21 at nominal value simultane- variables following any distribution.
ously. To avoid this, we define a new set SC (sc ¼ 1 and 2) X
and new auxiliary variables rVU(u,n,sc). rVU ðu; n; scÞ þ e  f ½k; DðuÞ  DðuÞ þ d max½1; DðuÞ
X n
~
rVU ðu; n; scÞ  DðuÞ 8u; sc ¼ 1 (24a) 8u; sc ¼ 1 ð27aÞ
Xn
X
~
rVU ðu; n; scÞ  DðuÞ 8u; sc ¼ 2 (24b) rVU ðu; n; scÞ  DðuÞ þ e  f ½k; DðuÞ  d max½1; DðuÞ
n n

~
where DðuÞ is the true realization of the uncertain demand 8u; sc ¼ 2 ð27bÞ
parameter.
Note that the violation probability is 2j.
The robust optimization techniques can be easily applied
to Eq. 24a,b simultaneously. Uniform continuous distribution
Bounded uncertainty If the demand parameters are uncertain and follow a uni-
form continuous distribution, then f ½k; DðuÞ ¼
If the demand parameter is uncertain and varies in a
ð1  2jÞ  DðuÞ. The deterministic robust counterpart is
bounded interval [DL (u), DU (u)], then the robust counter-
represented by
parts of Eq. 24a,b are given by
X
X rVU ðu; n; scÞ þ e  ð1  2jÞDðuÞ  DðuÞ þ d max½1; DðuÞ
rVU ðu; n; scÞ  DL ðuÞ þ d maxð1; jDðuÞjÞ 8u; sc ¼ 1 n
n
8u; sc ¼ 1 ð28aÞ
(25a)
X X
rVU ðu; n; scÞ  DU ðuÞ  d maxð1; jDðuÞjÞ 8u; sc ¼ 2 rVU ðu; n; scÞ  DðuÞ þ e  ð1  2jÞDðuÞ  d max½1; DðuÞ
n n
(25b) 8u; sc ¼ 2 ð28bÞ
Normal distribution
Bounded and symmetric uncertainty
If the demand parameters follow a normal distribution,
If the demand parameters are uncertain and distributed then f ½k; DðuÞ ¼ F1 1
n ð1  jÞ  DðuÞ where Fn ð1  jÞ being
around the nominal values randomly and symmetrically as the inverse standard normal cumulative distribution function.
follows The robust counterpart is represented by
X
~
DðuÞ ¼ ½1 þ e  nðuÞ  DðuÞ rVU ðu; n; scÞ þ e  F1
n ð1  jÞ  DðuÞ
n
where n(u) are random variables distributed symmetrically in  DðuÞ þ d max½1; DðuÞ 8u; sc ¼ 1 ð29aÞ
the interval [1, 1]. X
The robust counterparts for Eq. 24a,b are derived as follows rVU ðu; n; scÞ  DðuÞ þ e  F1
n ð1  jÞ  DðuÞ
n
X
rVU ðu; n; scÞ þ e  DðuÞ  ½ruðu; scÞ þ X  rzzðu; scÞ  d max½1; DðuÞ 8u; sc ¼ 2 (29b)
n
 DðuÞ þ d max½1; DðuÞ 8u; sc ¼ 1 ð26aÞ The deterministic robust counterparts for other known
probability distributions are also presented in Supporting
where Information, Appendix S1.
ruðu; scÞ  1 þ rzðu; scÞ  ruðu; scÞ 8u; sc ¼ 1 Unknown probability distribution
rzzðuÞ  rzðuÞ  rzzðuÞ 8u; sc ¼ 1
rzðuÞ  0; rzzðuÞ  0 8u; sc ¼ 1 When the demand parameters are uncertain and follow an
unknown probability distribution, then the robust counter-
parts for Eq. 24a,b are given by
X
VU ðu; n; scÞ  DðuÞ þ e  DðuÞ  ½ruðu; scÞ X 1
n (26b) rVU ðu; n; scÞ  DðuÞ  d max½1; DðuÞ 8u; sc ¼ 1
þ X  rzzðu; scÞ  d max½1; DðuÞ 8u; sc ¼ 2 n
j

where (30a)
X  
2j  1
ruðu; scÞ  1 þ rzðu; scÞ  ruðu; scÞ 8u; sc ¼ 2 rVU ðu; n; scÞ  max 0;  DðuÞ þ d max½1; DðuÞ
j
rzzðuÞ  rzðuÞ  rzzðuÞ 8u; sc ¼ 2 n

rzðuÞ  0; rzzðuÞ  0 8u; sc ¼ 2 8u; sc ¼ 2 ð30bÞ

AIChE Journal August 2012 Vol. 58, No. 8 Published on behalf of the AIChE DOI 10.1002/aic 2381
Note that the violation probability is 2j. The amount of crude c fed from tank i to CDU u during
Besides the above robust counterparts, we also define some event point n for each sc [VI,U,C(i,u,c,n)] is computed by,
other auxiliary variables including rVI(i,n,sc), rVI,C(i,c,n,sc), X
rVI,U(i,u,n,sc), rVI,U,C(i,u,c,n,sc), and rEI,C(i,c,n,sc), and rVI;U ði; u; n; scÞ ¼ rVI;U;C ði; u; c; n; scÞ
rSSP(sc). These variables are connected with each other and c:ði;cÞ2SI;C
rVU(u,n,sc) using the following additional constraints:
8ði; uÞ 2 SI;U ; sc ð38Þ
We first relate rVI(i,n,sc) with the variable rVI,C(i,c,n,sc).
For each sc, the total crude volume in each tank i at the end
of event point n is equivalent to summation of the volume For each sc, the total amount of crudes [VU(u,n)] fed to
of each crude c in this tank i. each CDU u during event point n is given by,
X
X rVI;U ði; u; n; scÞ ¼ rVU ðu; n; scÞ 8u; n; sc (39)
rVI;C ði; c; n; scÞ ¼ rVI ði; n; scÞ 8i; n; sc (31) i:ði;uÞ2SI;U
c:ði;cÞ2SI;C
For each sc, the total amount of crudes fed to each CDU
For each sc, at any time, the crude c in tank i must meet u during each event point n must meet its minimum
U ðuÞ] and maximum [DU ðuÞ] processing rates.
[Dmin max
I ði; cÞ] and upper [EI ði; cÞ] fractions in this
its lower [Emin max

tank.  end 
U ðuÞ TU ðu; nÞ  TU ðu; nÞ  rVU ðu; n; scÞ
Dmin 8u; n; sc
start

rVI ði; n; scÞ  Emin


I ði; cÞ  rVI;C ði; c; n; scÞ 8ði; cÞ 2 SI;C ; n; sc
(40a)
(32a)
 end 
rVI;C ði; c; n; scÞ  rVI ði; n; scÞ  Emax
I ði; cÞ 8ði; cÞ 2 SI;C ; n; sc
rVU ðu; n; scÞ  Dmax
U ðuÞ TU ðu; nÞ  TU ðu; nÞ
start
8u; n; sc
(32b) (40b)

The crude fraction in the feed to any CDU u under each


Each tank i has several possible event points during which
the concentration of this tank i is the same as its initial com-
sc must also meet its minimum [EminU ðu; cÞ] and maximum
[Emax
U ðu; cÞ] fractions.
position. During the possible event points
X
rVU ðu; n; scÞ  Emin
U ðu; cÞ  rVI;U;C ði; u; c; n; scÞ
rVI;U;C ði; u; c; n; scÞ ¼ Einit
I ði; cÞ  rVI;U ði; u; n; scÞ i:ði;uÞ2SI;U

8ði; uÞ 2 SI;U ; ði; cÞ 2 SI;C ; ði; nÞ 2 SF;I ; n; sc ð33Þ 8ðu; cÞ 2 SU;C ; sc ð41aÞ
X
where SF,I denotes the possible event points during which the rVI;U;C ði; u; c; n; scÞ  rVU ðu; n; scÞ  Emax
U ðu; cÞ
concentration of this tank i is the same as its initial i:ði;uÞ2SI;U
composition.
8ðu; cÞ 2 SU;C ; sc ð41bÞ
At other event point on each tank i [i.e., (i, n) 62 SF,I]

rVI;U;C ði; u; c; n; scÞ ¼ rEI;C ði; c; n  1; scÞ  rVI;U ði; u; n; scÞ For every sc, the desired crude qualities feeding to CDUs
8ði; uÞ 2 SI;U ; ði; cÞ 2 SI;C ; ði; nÞ 62 SF;I ; n; sc ð34Þ
must be ensured within the minimum [emin U ðu; kÞ] and maxi-
mum [emaxU ðu; kÞ] acceptable limits on properties that in the
feed to CDU u.
At the end of each event point n, X
U ðu; kÞ 
emin rVI;U ði; u; n; scÞ
rVI;C ði; c; n; scÞ ¼ rEI;C ði; c; n; scÞ  rVI ði; n; scÞ X
i:ði;uÞ2SI;U
X
8ði; cÞ 2 SI;C ; n; sc ð35Þ  eC ðc; kÞ  rVI;U;C ði; u; c; n; scÞ
i:ði;uÞ2SI;U c:ði;cÞ2SI;C
For each sc, the total amount of crudes from storage tank 8ði; uÞ 2 SI;U ; sc ð42aÞ
i to CDU u during each event point n must meet its mini-
mum [Fmin I;U ði; uÞ] and maximum [FI;U ði; uÞ] feed rates.
max X X
eC ðc; kÞ  rVI;U;C ði; u; c; n; scÞ
h i
rVI;U ði; u; n; scÞ  Fmax
I;U ði; uÞ  T end
I;U ði; u; nÞ  T start
I;U ði; u; nÞ i:ði;uÞ2SI;U c:ði;cÞ2SI;C
X
 emax
U ðu; kÞ  rVI;U ði; u; n; scÞ
8ði; uÞ 2 SI;U ; sc ð36aÞ i:ði;uÞ2SI;U
h i 8ði; uÞ 2 SI;U ; sc (42b)
I;U ði; uÞ  TI;U ði; u; nÞ  TI;U ði; u; nÞ  rVI;U ði; u; n; scÞ
Fmin end start
0 1
8ði; uÞ 2 SI;U ; sc ð36bÞ X X
@ qc  rVI;U;C ði; u; c; n; scÞA
U ðu; kÞ 
emin
For each sc, if tank i does not feed CDU u at event point i:ði;uÞ2SI;U c:ði;cÞ2SI;C
n, then the total amount [VI,U(i,u,n)] charged should be zero. X X
 eC ðc; kÞ  qc  rVI;U;C ði; u; c; n; scÞ
rVI;U ði; u; n; scÞ  VI;U
max
ði; u; nÞ  Yði; u; nÞ 8ði; uÞ 2 SI;U ; sc i:ði;uÞ2SI;U c:ði;cÞ2SI;C

(37) 8ði; uÞ 2 SI;U ; sc ð43aÞ

2382 DOI 10.1002/aic Published on behalf of the AIChE August 2012 Vol. 58, No. 8 AIChE Journal
X X The average safety stock for each sc [SSP(sc)] is calcu-
eC ðc; kÞ  qc  rVI;U;C ði; u; c; n; scÞ
i:ði;uÞ2SI;U c:ði;cÞ2SI;C
lated as,
0 1
X X
 emax @ qc  rVI;U;C ði; u; c; n; scÞA  
U ðu; kÞ  P P
i:ði;uÞ2SI;U c:ði;cÞ2SI;C rVI ði; n; scÞ þ VIinit ðiÞ
i n
8ði; uÞ 2 SI;U ; sc (43b) rSSPðscÞ  SS  8 sc
Nþ1
where, eC(c,k) denotes the known blending index for a (45)
property k of crude c, and qc is the density of crude c.
The inventory balance for each storage tank i at the end where, SS is the desired safety stock of crude.
of event point n for each sc can be expressed as follows, The hard bounds for those auxiliary variables are given
X below:
rVI;C ði; c; n; scÞ ¼ rVI;C ði; c; n  1; scÞ þ VP;I ðp; i; nÞ
p:ðp;cÞ2SP;C
X
 EP ðp; cÞ  rVI;U;C ði; u; c; n; scÞ rSSPðscÞ  SS 8 sc (46)
u:ði;uÞ2SI;U

8ði; cÞ 2 SI;C ; n > 1; sc ð44aÞ

X 0  rEI;C ði; c; n; scÞ  1 8ði; cÞ 2 SI;C ; sc (47a; b)


rVI;C ði; c; n; scÞ ¼ VI;C
init
ði; cÞ þ VP;I ðp; i; nÞ  EP ðp; cÞ
p:ðp;cÞ2SP;C
X
 rVI;U;C ði; u; c; n; scÞ 8ði; cÞ 2 SI;C ; n ¼ 1; sc
u:ði;uÞ2SI;U VImin ðiÞ  rVI ði; n; scÞ  VImax ðiÞ 8ði; cÞ 2 SI;C ; sc
(44b) (48a; b)

8 n o
>
> min Fmax ði; uÞ  H; VIinit  VImin ði; nÞ; VImax ðiÞ  VImin ðiÞ if n ¼ 1
>
< I;U
( )
rVI;U ði; u; n; scÞ  8ði; uÞ 2 SI;U ; sc ð49Þ
> min FI;U ði; uÞ  H; VI ði; n  1Þ  VI ði; nÞ
max max min
>
> if n > 1
: VImax ðiÞ  VImin ðiÞ

The deterministic robust counterpart optimization model is rVI;C ði; c; n; scÞ ¼ rEI;C ði; c; n; scÞ  rVI ði; n; scÞ
denoted as ROM presented below. 8ði; cÞ 2 SI;C ; n; sc ð350 Þ

ðROMÞ Min PROFIT


s:t: Eqs: A1A43 in Appendix A
Eqs: 25a; b; 26a; b;27a; b; or 30a; b Remarks: Cao et al.40 and Wang and Rong41 used only
for corresponding distribution Eq. 23 to convert Eq. 21 to inequality. The corresponding
Eqs: 3145; and 4649 deterministic robust counterparts with different distributions
Eqs: A45A54 and Eqs: A55A61 from the robust framework are presented in Supporting In-
in Appendix A formation, Appendix S2. The deterministic robust counter-
part optimization model of Cao et al.40 and Wang and
The resulting mathematical model ROM is a nonconvex Rong41 (ROM-CWR) always maximizes the total amount
MINLP, and the sources of nonconvexities are the distinct fed to each CDU for the objective of profit maximization
bilinear terms (i.e., Eqs. A15, A16, 34 and 35) regardless of specific demands of CDUs, as illustrated
later.
VI;U;C ði; u; c; nÞ ¼ EI;C ði; c; n  1Þ  VI;U ði; u; nÞ
8ði; uÞ 2 SI;U ; ði; cÞ 2 SI;C ; ði; nÞ 62 SF;I ; n ðA150 Þ Global Optimization Approach
In the model ROM, additional bilinear terms (i.e.,
VI;C ði; c; nÞ ¼ EI;C ði; c; nÞ  VI ði; nÞ 8ði; cÞ 2 SI;C ; n Eqs. 34,35 or 340 ,350 ) are introduced. These bilinear
(A160 ) terms are relaxed using only the McCormick convex
and concave envelopes. The McCormick convex and
rVI;U;C ði; u; c; n; scÞ ¼ rEI;C ði; c; n  1; scÞ  rVI;U ði; u; n; scÞ concave envelopes for the bilinear terms (i.e., Eqs. 34,
35 or 340 , 350 ) in the model ROM are derived as
8ði; uÞ 2 SI;U ; ði; cÞ 2 SI;C ; ði; nÞ 62 SF;I ; n; sc ð340 Þ follows

AIChE Journal August 2012 Vol. 58, No. 8 Published on behalf of the AIChE DOI 10.1002/aic 2383
rVI;U;C ði; u; c; n; scÞ
8
>
>  rEI;C ði; c; n  1; scÞ  rVI;U
min
ði; u; n; scÞ þ rEmin
I;C ði; c; n  1; scÞ  rVI;U ði; u; n; scÞ  rEI;C ði; c; n  1; scÞ  rVI;U ði; u; n; scÞ
min min
>
>
<  rEI;C ði; c; n  1; scÞ  rV min ði; u; n; scÞ þ rEmax ði; c; n  1; scÞ  rVI;U ði; u; n; scÞ  rEmax ði; c; n  1; scÞ  rV min ði; u; n; scÞ
I;U I;C I;C I;U
>  rEI;C ði; c; n  1; scÞ  rVI;U
> max
ði; u; n; scÞ þ rE min
ði; c; n  1; scÞ  rV ði; u; n; scÞ  rE min
ði; c; n  1; scÞ  rVI;U ði; u; n; scÞ
max
>
> I;C I;U I;C
:
 rEI;C ði; c; n  1; scÞ  rVI;U ði; u; n; scÞ þ rEI;C ði; c; n  1; scÞ  rVI;U ði; u; n; scÞ  rEI;C ði; c; n  1; scÞ  rVI;U
max max max max
ði; u; n; scÞ
8ði; uÞ 2 SI;U ; ði; cÞ 2 SI;C ; n ð50Þ
8
>
>  rEI;C ði; c; n; scÞ  rVImin ði; n; scÞ þ rEmin
I;C ði; c; n; scÞ  rVI ði; n; scÞ  rEI;C ði; c; n; scÞ  rVI ði; n; scÞ
min min
>
>
<  rEI;C ði; c; n; scÞ  rV min ði; n; scÞ þ rEmax ði; c; n; scÞ  rVI ði; n; scÞ  rEmax ði; c; n; scÞ  rV min ði; n; scÞ
I I;C I;C I
rVI;C ði; c; n; scÞ
>
>  rEI;C ði; c; n; scÞ  rVI ði; n; scÞ þ rEI;C ði; c; n; scÞ  rVI ði; n; scÞ  rEI;C ði; c; n; scÞ  rVI ði; n; scÞ
max min min max
>
>
:
 rEI;C ði; c; n; scÞ  rVImax ði; n; scÞ þ rEmax
I;C ði; c; n; scÞ  rVI ði; n; scÞ  rEI;C ði; c; n; scÞ  rVI
max max
ði; n; scÞ
8ði; cÞ 2 SI;C ; n ð51Þ

The piecewise-linear relaxation of the model ROM


denoted as RROM is defined as follows

ðRROMÞ Min PROFIT


s:t: Eqs: A1A14; A17A43 in Appendix A
Eqs: B1B2 in Appendix B
Eqs: 25a; b; 26a; b; 27a; b; or 30a; b
for corresponding distribution
Eqs: 3133; 3645; 46 49; and 5051
Eqs: A45A54 and Eqs: A55A61
in Appendix A

Note that we do not use piecewise-linear relaxation for


those variables rEI,C(i,c,n,sc), rVI(i,n,sc), and rVI,U(i,u,n,sc)
in the model RROM.
A pool of feasible solutions including the final solve from
the model RROM is used to fix the current values of the
binary variables, initialize the continuous variables using
their current values, and locally minimize the resulting NLP.
The objective function for NLP changes to the following
XXXX
PROFIT ¼ CPROF ðcÞ  VI;U;C ði; u; c; nÞ
i u c n
X
 TCW ðvÞ  PEN  SSP  H
v
XX
 CSET  zðu; nÞ
u n
XXXXX
þ CPROF ðcÞ  rVI;U;C ði; u; c; n; scÞ
i u c n sc
X
 PEN  rSSPðscÞ  H ð52Þ
sc
The LP minimization and maximization problems for
tightening the lower and UBs for variables EI,C(i,c,n),
VI(i,n), and VI,U(i,u,n) can be stated as follows

ðROMLÞ Min z obbt


s:t: Eqs: A1A14; A17A43 in Appendix A
Eqs: B3B4 in Appendix B
Eqs: 25a; b; 26a; b; 27a; b; or 30a; b
for corresponding distribution
Eqs: 3133; 3645; 4649; and 5051
Eqs: A45A54 and Eqs: A55A61
in Appendix A Figure 7. Flow chart of the extended branch and bound
0  Xðp; i; nÞ; Yði; u; nÞ  1 global optimization algorithm from Li et al.3

2384 DOI 10.1002/aic Published on behalf of the AIChE August 2012 Vol. 58, No. 8 AIChE Journal
Table 3. Model and Solution Statistics for Example 2 ðROMUÞ Max z obbt
Robust Robust s:t: Eqs: A1A14; A17A43 in Appendix A
Solution for Solution for Eqs: B3B4 in Appendix B
Nominal Bounded Normal Eqs: 25a; b; 26a; b; 27a; b; or 30a; b
Solution Uncertainty Distribution
for corresponding distribution
Event points 3 3 3 Eqs: 3133; 3645; 4649; and 5051
GR 2 4 3
Binary variables 160 160 160 Eqs: A45A54 and Eqs: A55A61
Continuous variables 1189 2865 2577 in Appendix A
Constraints 5258 9076 8884 0  Xðp; i; nÞ; Yði; u; nÞ  1
Bilinear terms 192 576 576
Obj (UB, K$) 4795.037 4780.727 4789.280
LB (K$) 4884.673 4868.617 4882.344 where z_obbt becomes VI(i,n), EI,C(i,c,n), and VI,U(i,u,n)
Gap (%) 1.84 1.81 1.92 respectively to update VImin ði; nÞ, VImax ði; nÞ, Emin I;C ði; c; nÞ,
CPU time (s) 4063 51600 5630
Emax
I;C ði; c; nÞ, and V max
I;U ði; u; nÞ accordingly. The entire proce-
Y ¼ 0.50, e ¼ 0.02. dure of the branch and bound global optimization algorithm
for ROM is illustrated in Figure 7.

Figure 8. Robust schedule for Example 1 from model ROM with demand uncertainty following bounded
uncertainty.

Figure 9. Inventory profiles of T1–T5 for Example 1 from the robust schedule in Figure 8.
[Color figure can be viewed in the online issue, which is available at wileyonlinelibrary.com.]

AIChE Journal August 2012 Vol. 58, No. 8 Published on behalf of the AIChE DOI 10.1002/aic 2385
We use model ROM to generate the robust schedule
(Figure 8) within 72.7 CPU seconds and with the objective
of $ 5614.974K, which deviates only by 0.298% from that
of the nominal schedule. Note that the objective for the
robust schedule is very close to that of nominal schedule.
From the robust schedule (Figure 8), CDU101 processes 600
kbbl, and CDU102 processes 600 kbbl. Figures 9 and 10
illustrate the inventory profiles of T1–T5 and feed rates to
CDU101 and CDU102, respectively. Note that T1 and T4
can feed CDU101, and T2, T3, and T5 can feed CDU102.
From Figure 9, we observe that T1 has 210 kbbl additional
crudes that can be charged to CDU101, although T4 reaches
its minimum capacity at the end of scheduling horizon. Let
us examine Figure 10, the feed rates of T1 and T4 to
CDU101 are 3.97 kbbl/s and 3.68 kbbl/s, respectively, in
[29.2, 72] h. As the maximum feed rate to CDU101 is 12.5
kbbl/s, the feed rate of T1 to CDU101 can increase to 8.82
kbbl/s during [29.2, 72] h. The maximum additional amount
of crudes from T1 that can be charged to CDU101 is
max[210kbbl, (8.823.97)  42.8 kbbl] ¼ 207.5 kbbl. Thus,
any demand of CDU101 within [600, 720] kbbl can be satis-
fied by simply adjusting feed rate of T1 to CDU101 with the
robust schedule in Figure 8. It is not difficult to conclude
that any demand of CDU101 within [420, 600] kbbl can also
Figure 10. Feed rates to CDU101 and CDU102 for Exam- be satisfied by simply adjusting feed rates of T1 and T4 to
ple 1 from the robust schedule in Figure 8. CDU101 with the robust schedule in Figure 8. A similar
[Color figure can be viewed in the online issue, which conclusion can be made for CDU102. In Supporting Infor-
is available at wileyonlinelibrary.com.] mation, Appendix S3, we show that the schedule in Figure 8
is feasible for any demand realization within the range [480,
Computational Studies 720] kbbl by simply adjusting the feed rates from T1–T5 to
The proposed deterministic robust counterpart optimiza- CDU101 and CDU102.
tion formulations are applied to two examples. The complete Normal Distribution. We consider uncertainty with a
normal distribution in the demands of both CDU101 and
data for Example 2 are given in Supporting Information,
CDU102. The uncertainty level is e ¼ 15%; the infeasibil-
Tables S1–S3. They are solved using GAMS 22.6/CPLEX
TM ity tolerance is d ¼ 0%; and the reliability level j ¼ 10%.
11.0.0 on Dell OPTIPLEX 960 of Intel V R Xeon CPU 3.0
We solve model ROM and obtain the robust schedule as
GHz with 2 GB RAM running Linux. The computational
shown in Figure 11 within 72.7 CPU seconds. The corre-
results are presented in Tables 2 and 3.
sponding objective is $ 5614.974K, which is only 0.298%
different than the objective of the nominal schedule. From
Example 1 Figure 11, each CDU processes exact 600 kbbl of crudes.
Let us revisit Example 1 of Section ‘‘Motivating Exam- The inventory levels of T1–T5 and feed rates to CDUs are
ple’’. The nominal schedule with nominal demands of 600 illustrated in Supporting Information, Figures S1 and S2. A
kbbl for both CDU101 and CDU102 is given in Figure 3. similar analysis as bounded uncertainty can be done to con-
The objective is $ 5631.707K. clude that the inventory levels in some tanks do not reach
Bounded Uncertainty. The demand parameters of both their maximum and minimum capacities and the feed rates
CDU101 and CDU102 are assumed to be uncertain and vary in from some tanks to CDUs do not reach maximum and min-
a bounded interval [480,720] kbbl. When the demands of imum limits either. Therefore, demand uncertainty follow-
CDU101 and CDU102 increase to 705 kbbl, the nominal sched- ing normal distribution can be accommodated from the
ule is infeasible, as shown in Section ‘‘Motivating Example’’. robust schedule (Figure 11) by simply adjusting feed rates

Figure 11. Robust schedule for Example 1 from model ROM with demand uncertainty following normal distribution.

2386 DOI 10.1002/aic Published on behalf of the AIChE August 2012 Vol. 58, No. 8 AIChE Journal
Figure 12. Robust schedule for Example 1 from ROM-CWR with demand uncertainty following bounded uncertainty.

to CDUs. By executing this robust schedule, we can ensure Example 2


that the schedule is feasible by simply adjusting the feed This example is Example 21 from Li et al.3,34 It is the
rates from T1–T5 to CDU101 and CDU102 with a proba- largest industrial-scale problem involving one SBM pipeline,
bility of 90% in the presence of the 15% uncertainty in the three jetties, eight storage tanks (T101–T108), and three
demands of both CDU101 and CDU102. It is interesting to CDUs (CDU101–CDU103). The scheduling horizon is about
note that the same schedule is obtained with bounded 336 hrs (i.e., 14 days). The nominal demands for CDU101,
uncertainty. CDU102, and CDU103 are 1000 kbbl, 1000 kbbl, and 1000
The robust schedule for other known probability distribu- kbbl, respectively. The nominal schedule is illustrated in Fig-
tions such as uniform distribution and poisson distribution ure 13 with the objective of $ 4795.037K. When the
can also obtained from model ROM. demands of CDU101, CDU102, and CDU103 increase to
Remarks: We solve the deterministic robust counterpart 1100 kbbl, the nominal schedule is infeasible.
model ROM-CWR (see Supporting Information, Appendix Bounded Uncertainty. The demand parameters of
S2) and generate a schedule (Figure 12) with the objective CDU101, CDU102, and CDU103 are assumed to be uncertain
of $7248.618K. From Figure 12, it can be calculated that and vary in a bounded interval [850, 1150] kbbl. The model
the total amount fed to CDU101 and CDU102 are 785.00 ROM involves 160 binary variables, 2865 continuous variables,
kbbl and 825.58 kbbl respectively, which are greater than 9076 constraints, and 576 bilinear terms. The robust schedule
the nominal demand (600 kbbl). It is not difficult to con- (Supporting Information, Figure S5) from model ROM is
clude that each CDU processes the maximum amount of obtained within 51,600 CPU seconds and the objective is $
crude mixture. As the objective of the model ROM-CWR is 4780.727K, which is guaranteed to be within 2% of global
to maximize the total profit, the total amount fed to each optimality. More importantly, the objective of $ 4780.727K
CDU is maximized regardless of the specified demand for is only 0.257% different from that of nominal schedule. The
each CDU. Supporting Information, Figures S3 and S4 illus- inventory levels of T1–T8 and feed rates to CDUs are depicted
trate the inventory profiles of T1–T5 and feed rates to in Supporting Information, Figures S6 and S7. Similar to
CDU101 and CDU102. Example 1, it can be concluded that the inventory levels in

Figure 13. Nominal schedule for Example 2 from the proposed branch and bound global optimization algorithm of
Li et al.3

AIChE Journal August 2012 Vol. 58, No. 8 Published on behalf of the AIChE DOI 10.1002/aic 2387
Figure 14. Robust schedule for Example 2 from model ROM with demand uncertainty following normal distribution.

some tanks do not reach their maximum and minimum capaci- global optimization algorithm proposed by Li et al.3 was suc-
ties and the feed rates from some tanks to CDUs do not reach cessfully extended to solve the deterministic robust counter-
maximum and minimum limits either. Therefore, we can sim- part optimization model. The computational results show that
ply adjust feed rates to CDUs to accommodate demand uncer- the generated schedule is more robust than the nominal sched-
tainty based on the robust schedule. ule. In the future, we will extend the robust optimization
Normal Distribution. We also consider uncertainty framework to address other forms of uncertainty in ship
with a normal distribution in the demands. The uncertainty arrival, quality specification and some economic coefficients.
level is e ¼ 8%; the infeasibility tolerance is d ¼ 0% and
the reliability level j ¼ 5%. In the model ROM, we have
160 binary variables, 2577 continuous variables, 8884 con- Acknowledgments
straints, and 576 bilinear terms. The robust schedule (Figure The authors gratefully acknowledge support from the National Science
14) is obtained within 5630 CPU seconds and the objective Foundation (CMMI-08856021). Ruth Misener is further thankful for her
National Science Foundation Graduate Research Fellowship.
is $ 4789.280K, which is guaranteed to be within 2% of
global optimality. The objective function is also near to that
of nominal schedule, only 0.120% different from that of Literature Cited
nominal schedule. By executing this robust schedule (Figure 1. Pinto JM, Joly M, Moro L. Planning and scheduling models for
14), we can guarantee the schedule is feasible by simply refinery operations. Comput Chem Eng. 2000;24:2259–2276.
adjusting the feed rates from T1–T8 to CDU101, CDU102, 2. Li J, Karimi IA, Srinivasan R. Recipe determination and scheduling
and CDU103 within with a probability of 95% in the pres- of gasoline blending operations. AIChE J. 2010;56:441–465.
ence of the 8% uncertainty in the demands of CDU101, 3. Li J, Misener R, Floudas CA. Continuous-time modeling and global
optimization approach for scheduling of crude oil operations. AIChE
CDU102, and CDU103. The inventory levels of T1–T8 and
J., in press; DOI: 10.1002/aic.12623.
feed rates to CDUs are illustrated in Supporting Information, 4. Kelly JD, Mann JL. Crude-oil blend scheduling optimization: an
Figures S8 and S9. application with multi-million dollar benefits—Part 1. Hydrocarbon
Similar to Example 1, the schedules in Figure 14 and Support- Process. 2003;82:47–53.
ing Information Figure S5 are more robust compared to the 5. Floudas CA, Lin X. Continuous-time versus discrete-time
nominal schedule in Figure 13. The robust schedules for other approaches for scheduling of chemical processes: a review. Comput
Chem Eng. 2004;8:2109–2129.
known probability distributions such as uniform distribution and
6. Floudas CA, Lin X. Mixed integer linear programming in process
Poisson distribution can also be obtained from the model ROM. scheduling: modeling, algorithms, and applications. Ann Oper Res.
2005;139:131–162.
7. Ierapetritou MG, Floudas CA. Effective continuous-time formulation
Conclusions for short-term scheduling: 1. Multipurpose batch processes. Ind Eng
Chem Res. 1998;37:4341–4359.
In this article, we addressed the problem of scheduling of 8. Ierapetritou MG, Floudas CA. Effective continuous-time formulation
crude oil operations under demand uncertainty. The novel for short-term scheduling. 2. Continuous and semi-continuous pro-
unit-specific event-based continuous-time MINLP formulation cess. Ind Eng Chem Res. 1998;37:4360–4374.
developed by Li et al.3 and the robust optimization framework 9. Ierapetritou MG, Hene TS, Floudas CA. Effective continuous-time
formulation for short-term scheduling. 3. Multiple intermediate due
developed by Lin et al.48 and Janak et al.49 were successfully
dates. Ind Eng Chem Res. 1999;38:3446–3461.
utilized and applied to develop robust optimization models, 10. Lin X, Floudas CA. Design, synthesis and scheduling of multipur-
where a new approach was proposed to convert demand pose batch plants via an effective continuous-time formulation.
equality constraints to inequalities. The branch and bound Comput Chem Eng. 2001;25:665–674.

2388 DOI 10.1002/aic Published on behalf of the AIChE August 2012 Vol. 58, No. 8 AIChE Journal
11. Lin X, Floudas CA, Modi S, Juhasz NM. Continuous-time optimiza- 37. Adhitya A, Srinivasan R, Karimi IA. Heuristic rescheduling of crude
tion approach for medium-range production scheduling of a multi- oil operations to manage abnormal supply chain events. AIChE J.
product batch plant. Ind Eng Chem Res. 2002;41:3884–3906. 2007;53:397–422.
12. Lin X, Chajakis ED, Floudas CA. Scheduling of tanker lightering 38. Li Z, Ierapetritou M. Processing scheduling under uncertainty:
via a novel continuous-time optimization framework. Ind Eng Chem review and challenges. Comput Chem Eng. 2008;32:715–727.
Res. 2003;42:4441–4451. 39. Li J, Karimi IA, Srinivasan R. Robust scheduling of crude oil oper-
13. Janak SL, Lin X, Floudas CA. Enhanced continuous-time unit-spe- ations under demand and ship arrival uncertainty. Presented at the
cific event-based formulation for short-term scheduling of multipur- AIChE Annual Meeting, San Francisco, CA, Nov. 12–17, 2006.
pose batch processes: resource constraints and mixed storage poli- 40. Cao CW, Gu XS, Xin Z. Chance constrained programming models
cies. Ind Eng Chem Res. 2004;43:2516–2533. for refinery short-term crude oil scheduling problem. Appl Math
14. Janak SL, Lin X, Floudas CA. Comments on ‘‘Enhanced continu- Model. 2009;33:1696–1707.
ous-time unit-specific event-based formulation for short-term sched- 41. Wang JS, Rong G. Robust optimization model for crude oil schedul-
uling of multipurpose batch processes: resource constraints and ing under uncertainty. Ind Eng Chem Res. 2010;49:1737–1748.
mixed storage policies.’’ Ind Eng Chem Res. 2005;44:426. 42. Ben-Tal A, Nemirovski A. Robust convex optimization. Math Oper
15. Janak SL, Floudas CA, Kallrath J, Vormbrock N. Production sched- Res. 1998;23:769–805.
uling of a large-scale industrial batch plant. I. Short-term and me- 43. Ben-Tal A, Nemirovski A. Robust solutions of uncertain linear pro-
dium-term scheduling. Ind Eng Chem Res. 2006;45:8234–8252. grams. Oper Res Lett. 1999;25:1–13.
16. Janak SL, Floudas CA. Improving unit-specific event based continu- 44. Ben-Tal A, Nemirovski A. Robust solutions of linear programming
ous-time approaches for batch processes: integrality gap and task problems contaminated with uncertain data. Math Prog A.
splitting. Comput Chem Eng. 2008;32:913–955. 2000;88:411–424.
17. Shaik MA, Janak SL, Floudas CA. Continuous-time models for 45. Ben-Tal A, Goryashko A, Guslitzer E, Nemirovski A. Adjustable robust
short-term scheduling of multipurpose batch plants: a comparative solutions for uncertain linear programs. Math Prog A. 2004;99:351–376.
study. Ind Eng Chem Res. 2006;45:6190–6209. 46. Ghaoui LHE. Robust solutions to least-square problems with uncer-
18. Shaik MA, Floudas CA. Improved unit-specific event-based model tain data. SIAM J Matrix Anal Appl. 1997;18:1035–1064.
continuous-time model for short-term scheduling of continuous proc- 47. Ghaoui LHE, Oustry F, Lebret H. Robust solutions to uncertain
esses: rigorous treatment of storage requirements. Ind Eng Chem semidefinite programs. SIAM J Opt. 1998;9:33–52.
48. Lin X, Janak SL, Floudas CA. A new robust optimization approach
Res. 2007;46:1764–1779.
for scheduling under uncertainty. I. Bounded uncertainty. Comput
19. Shaik MA, Floudas CA. Unit-specific event-based continuous-time
Chem Eng. 2004;28:1069–1085.
approach for short-term scheduling of batch plants using RTN
49. Janak SL, Lin X, Floudas CA. A new robust optimization approach
framework. Comput Chem Eng. 2008;32:260–274.
for scheduling under uncertainty. II. Uncertainty with known proba-
20. Shaik MA, Floudas CA. Novel unified modeling approach for short-
bility distribution. Comput Chem Eng. 2007:31:171–195.
term scheduling. Ind Eng Chem Res. 2009;48:2947–2964. 50. Verderame PM, Floudas CA. Operational planning of large-scale
21. Li J, Floudas CA. Optimal event point determination for short-term industrial batch plants under demand due date and amount uncer-
scheduling of multipurpose batch plants via unit-specific event-based tainty. I. Robust optimization framework. Ind Eng Chem Res.
continuous-time approaches. Ind Eng Chem Res. 2010;49:7446–7469. 2009;48:7214–7231.
22. Pham V, Laird C, El-Halwagi M. Convex hull discretization 51. Verderame PM, Floudas CA. Integration of operational planning and
approach to the global optimization of pooling problems. Ind Eng medium-term scheduling for large-scale industrial batch plants under
Chem Res. 2009;48:1973–1979. demand and processing time uncertainty. Ind Eng Chem Res.
23. Misener R, Floudas CA. Advances for the pooling problem: model- 2010;49:4948–4965.
ing, global optimization, and computational studies. Appl Comput 52. Verderame PM, Floudas CA. Multisite planning under demand and
Math. 2009;8:3–22. transportation uncertainty: robust optimization and conditional value
24. Hasan MMF, Karimi IA. Piecewise linear relaxation of bilinear pro- at risk framework. Ind Eng Chem Res. 2011;50:4959–4982.
grams using bivariate partitioning. AIChE J. 2010;56:1880–1893. 53. Li ZK, Ding R, Floudas CA. A comparative theoretical and compu-
25. Misener R, Floudas CA. Global optimization of large-scale general- tational study on robust counterpart optimization: I. Robust linear
ized pooling problems: quadratically constrained MINLP models. optimization and robust mixed integer linear optimization. Ind Eng
Ind Eng Chem Res. 2010;49:5424–5438. Chem Res. 2011;50:10567–10603.
26. Misener R, Gounaris CE, Floudas CA. Mathematical modeling and 54. Bertsimas D, Sim M. Robust discrete optimization and network
global optimization of large-scale extended pooling problems with flows. Math Prog B. 2003;98:49–71.
the (EPA) complex emissions constraints. Comput Chem Eng. 55. Bertsimas D, Sim M. The price of robustness. Oper Res.
2010;34:1432–1456. 2004;52:35–53.
27. Meyer CA, Floudas CA. Global optimization of a combinatorially 56. Bertsimas D, Pachamanova D, Sim M. Robust linear optimization
complex generalized pooling problem. AIChE J. 2006;52:1027–1037. under general norms. Oper Res Lett. 2004;32:510–516.
28. Karuppiah R, Grossmann IE. Global optimization for the synthesis 57. Verderame PM, Floudas CA. Integrated operational planning and
of integrated water systems in chemical processes. Comput Chem medium-term scheduling of a large-scale industrial batch plants. Ind
Eng. 2006;30:650–673. Eng Chem Res. 2008;47:4845–4860.
29. Wicaksono DS, Karimi IA. Piecewise MILP under- and overestima- 58. Verderame PM, Floudas CA. Operational planning of large-scale
tors for global optimization of bilinear programs. AIChE J. industrial batch plants under demand due date and amount uncer-
2008;54:991–1008. tainty. II. Conditional value-at-risk framework. Ind Eng Chem Res.
30. Gounaris CE, Misener R, Floudas CA. Computational comparison of 2010;49:260–275.
piecewise-linear relaxations for pooling problems. Ind Eng Chem 59. Verderame PM, Floudas CA. Operational planning framework for
Res. 2009;48:5742–5766. multisite production and distribution networks. Comput Chem Eng.
31. Bergamini ML, Grossmann IE, Scenna N, Aguirre P. An improved 2009;33:1036–1050.
piecewise outer-approximation algorithm for the global optimization 60. Susarla N, Li J, Karimi IA. A novel approach to scheduling multi-
of MINLP models involving concave and bilinear terms. Comput purpose batch plants using unit-slots. AIChE J. 2010;56:1859–1879.
Chem Eng. 2008;32:477–493. 61. Li J, Karimi IA, Srinivasan R. Efficient bulk maritime logistics for
32. Saif Y, Elkamel A, Pritzker M. Global optimization for reverse os- the supply and delivery of multiple chemicals. Comput Chem Eng.
mosis network for wastewater treatment and minimization. Ind Eng 2010;34:2118–2128.
Chem Res. 2008;47:3060–3070. 62. Li J, Karimi IA. Scheduling gasoline blending operations from rec-
33. Reddy PCP, Karimi IA, Srinivasan R. Novel solution approach for ipe determination to shipping using unit slots. Ind Eng Chem Res.
optimization crude oil operations. AIChE J. 2004;50:1177–1197. 2011;50:9156–9174.
34. Li J, Li WK, Karimi IA, Srinivasan R. Improving the robustness and ef- 63. Li J, Susarla N, Karimi IA, Shaik MA, Floudas CA. An analysis of
ficiency of crude scheduling algorithms. AIChE J. 2007;53:2659–2680. some unit-specific event-based models for the short-term scheduling
35. Verderame PM, Elia JA, Li J, Floudas CA. Planning and scheduling of noncontinuous processes. Ind Eng Chem Res. 2010;49:633–647.
under uncertainty: a review across multiple sections. Ind Eng Chem 64. Floudas CA. Nonlinear and Mixed-Integer Optimization: Fundamen-
Res. 2010;49:3993–4017. tals and Applications. New York, NY: Oxford University Press, 1995.
36. Janak SL, Floudas CA, Kallrath J, Vormbrock N. Production sched- 65. Floudas CA. Deterministic Global Optimization: Theory, Methods,
uling of a large-scale industrial batch plant. II. Reactive scheduling. and Applications; Nonconvex Optimization and It’s Applications.
Ind Eng Chem Res. 2006;45:8253–8269. Dordrecht, Netherlands: Kluwer Academic Publishers, 2000.

AIChE Journal August 2012 Vol. 58, No. 8 Published on behalf of the AIChE DOI 10.1002/aic 2389
Appendix A: Deterministic Mathematical Positive variables
Formulation of Li et al.3 EI,C(i,c,n) ¼ composition of crude c in tank i at the end of event
point n
SSP ¼ average safety stock at the end of each event point
Notation Tstart (n) ¼ start time of event point n on jetties
B
Sets B (n) ¼ end time of event point n on jetties
Tend
TCW(v) ¼ demurrage cost of vessel v
VP ¼ VLCC parcels
I;U (i,u,n) ¼ start time that tank i feeds CDU u during event point n
Tstart
JP ¼ jetty parcels
I;U (i,u,n) ¼ end time that tank i feeds CDU u during event point n
Tend
SF,I ¼ set of pairs (tank i, event point n) that the concentration Tstart (p) ¼ start time for parcel p unloading
P
P (p) ¼ end time for parcel p unloading
of tank i is the same as its initial composition during Tend
P;I (p,i,n) ¼ start time that parcel p is unloaded to tank i during
the possible event point n Tstart
SP,I ¼ set of pairs (parcel p, tank i) that tank i can receive event point n
P;I (p,i,n) ¼ end time that parcel p is unloaded to tank i during
parcel p Tend
SI,C ¼ set of pairs (tank i, crude c) that tank i can hold crude c event point n
SI,U ¼ set of pairs (tank i, CDU u) that tank i can feed CDU u Tstart (u,n) ¼ start time of event point n on CDU u
U
SU,C ¼ set of pairs (CDU u, crude c) that CDU u can process
U (u,n) ¼ end time of event point n on CDU u
Tend
crude c VP,I(p,i,n) ¼ crude amount transferred from parcel p to tank i during
Parameters event point n
VI(i,n) ¼ crude volume in tank i at the end of event point n
Dn ¼ 1 if a parcel is unloaded in multiple event points VI,C(i,c,n) ¼ volume of crude c in tank i at the end of event point n
CIVS ¼ inventory cost ($/unit/h) VI,U,C(i,u,c,n) ¼ amount of crude c fed from tank i to CDU u during
CPEN ¼ safety stock penalty ($/unit/h) event point n
CPROF(c) ¼ marginal profit ($/unit volume) from crude c VI,U(i,u,n) ¼ amount of crude that tank i feeds to CDU u during
CSEA ¼ demurrage or sea-waiting cost event point n
CSET ¼ cost (k$) per changeover VU(u,n) ¼ total amount of crudes fed to CDU u during event point n
CULD ¼ unloading cost ($/hr)
Dmin
U (u) ¼ minimum allowable crude processing rate of CDU u
Dmax
U (u) ¼ maximum allowable crude processing rate of CDU u
D(u) ¼ demand of each CDU u Constraints
Emin
I;C (i,c) ¼ lower limit on the composition of crude c in tank i
Emin ¼ lower limit on the composition of crude c in tank i at The event point (n þ 1) on unit m must start after the
I;C (i,c,n)
event point n event point n on this unit m ends.
I;C (i,c) ¼
Emax upper limit on the composition of crude c in tank i
I;C (i,c,n) ¼
Emax upper limit on the composition of crude c in tank i at TBstart ðn þ 1Þ  TBend ðnÞ 8n (A1a)
event point n
EP(p,c) ¼ fraction of crude c in parcel p TIstart ði; n þ 1Þ  TIend ði; nÞ 8i; n (A1b)
Emin
U (u,c) ¼ minimum allowable composition of crude c in feed to CDU u
Emax (u,c) ¼ maximum allowable composition of crude c in feed to CDU u
U
eC(c,k) ¼ index of property k in crude c TUstart ðu; n þ 1Þ  TUend ðu; nÞ 8u; n (A1c)
emin
U (u,k) ¼ minimum allowable index of property k in CDU u X X
emax
U (u,k) ¼ maximum allowable index of property k in CDU u Xðp; i; nÞ ¼ 1 8p; Dn ¼ 0 (A2a)
Fmin
I;U (i,u) ¼ minimum feeding rate of crude from tank i to CDU u n
i:ðp;iÞ2SPI
Vmin
I;U (i,u,n) ¼ minimum feeding amount of crudes from tank i to
CDU u at event point n X
Xðp; i; nÞ  1 8p; n; Dn ¼ 1 (A2b)
I;U (i,u) ¼
Fmax maximum feeding rate of crude from tank i to CDU u
I;U (i,u,n) ¼
Vmax i:ðp;iÞ2SP;I
maximum feeding amount of crudes from tank i to X
CDU u at event point n Xðp; i; nÞ  B 8i; n; B > 1 (A3)
Fmin
P;I (p,i) ¼ minimum unloading rate of crude from parcel p to tank i
p:ðp;iÞ2SP;I
Fmax
P;I (p,i) ¼ maximum unloading rate of crude from parcel p to tank i p2JP
Vmax
P;I (p,i,n) ¼ maximum unloading amount of crude from parcel p to X X
tank i at event point n xeðp; nÞ  2  Xðp; i; nÞ  Xðp; i; n þ 1Þ
H ¼ scheduling horizon i:ðp;iÞ2SP;I i:ðp;iÞ2SP;I
SS ¼ desired safety stock
ST ¼ minimum time for crude settling and brine removal 8p; n \ N; Dn ¼ 1 (A4a)
TARR(p) ¼ expected arrival time of parcel p X X
Vinit
I (i) ¼ initial crude volume in tank i xeðp; nÞ  Xðp; i; nÞ  Xðp; i; n þ 1Þ
Vinit
I;C (i,c) ¼ initial amount of crude c in tank i
i:ðp;iÞ2SP;I i:ðp;iÞ2SP;I
Vinit
P (p) ¼ initial crude volume of parcel p
Tmin
ULD (v) ¼ stipulated departure time in the logistics contract for 8p; n \ N; Dn ¼ 1 (A4b)
each vessel v
Vmin (i) ¼ minimum allowable crude inventory in tank i X
I
Vmin (i,n) ¼ minimum allowable crude inventory in tank i at event xeðp; nÞ  Xðp; i; nÞ 8p; n \ N; Dn ¼ 1 (A4c)
I
point n i:ðp;iÞ2SP;I
Vmax
I (i) ¼ maximum allowable crude inventory in tank i X
Vmax
I (i,n) ¼ maximum allowable crude inventory in tank i at event xeðp; nÞ ¼ 1 8p; Dn ¼ 1 (A5)
point n n
Binary variables h i
VP;I ðp; i; nÞ  Fmin
P;I ðp; iÞ  TP;I ðp; i; nÞ  TP;I ðp; i; nÞ
end start
X(p,i,n) ¼ 1 if parcel p is unloaded to tank i during event point n
Y(i,u,n) ¼ 1 if tank i is charging CDU u during event point n 8ðp; iÞ 2 SP;I ; n (A6a)
0–1 Continuous variables h i
xe(p,n) ¼ 1 if parcel p is completed at the end of event point n VP;I ðp; i; nÞ  Fmax
P;I ðp; iÞ  T end
P;I ðp; i; nÞ  T start
P;I ðp; i; nÞ
z(u,n) ¼ 1 if a tank switch on CDU u takes place at the end of
event n 8ðp; iÞ 2 SP;I ; n (A6b)

2390 DOI 10.1002/aic Published on behalf of the AIChE August 2012 Vol. 58, No. 8 AIChE Journal
X
VP;I ðp; i; nÞ ¼ VPinit ðpÞ  Xðp; i; nÞ 8ðp; iÞ 2 SP;I ; Dn ¼ 0 VI;U ði; u; nÞ ¼ VU ðu; nÞ 8u; n (A22)
(A7a) i:ði;uÞ2SI;U

 end 
U ðuÞ TU ðu; nÞ  TU ðu; nÞ  VU ðu; nÞ
Dmin 8u; n ðA23aÞ
start
VP;I ðp; i; nÞ  VP;I
max
ðp; i; nÞ  Xðp; i; nÞ 8ðp; iÞ 2 SP;I ; Dn ¼ 1
(A7b)  end 
VU ðu; nÞ  Dmax
U ðuÞ TU ðu; nÞ  TU ðu; nÞ
start
8u; n
X X
VP;I ðp; i; nÞ ¼ VPinit ðpÞ 8p; Dn ¼ 1 (A8) (A23b)
i:ðp;iÞ2SP;I n X
VU ðu; nÞ  Emin
U ðu; cÞ  VI;U;C ði; u; c; nÞ
TPstart ðpÞ  start
TP;I ðp; i; nÞ þ H½1  Xðp; i; nÞ 8ðp; iÞ 2 SP;I ; n i:ði;uÞ2SI;U

(A9a) 8ðu; cÞ 2 SU;C (A24a)


X
TPend ðpÞ  TP;I
end
ðp; i; nÞ  H½1  Xðp; i; nÞ 8ðp; iÞ 2 SP;I ; n VI;U;C ði; u; c; nÞ  VU ðu; nÞ  Emax
U ðu; cÞ
i:ði;uÞ2SI;U
(A9b)
8ðu; cÞ 2 SU;C (A24b)
TPstart ðp þ 1Þ  TPend ðpÞ 8p 2 VP (A10)
X X X
U ðu; kÞ 
emin VI;U ði; u; nÞ  eC ðc; kÞ
TPstart ðp0 Þ  TPend ðpÞ 8p; p0 2 JP; TARR ðpÞ\TARR ðp0 Þ; B ¼ 1 i:ði;uÞ2SI;U i:ði;uÞ2SI;U c:ði;cÞ2SI;C
(A11) VI;U;C ði; u; c; nÞ 8ði; uÞ 2 SI;U (A25a)
X
VI;C ði; c; nÞ ¼ VI ði; nÞ 8i; n (A12) X X
c:ði;cÞ2SI;C eC ðc; kÞ  VI;U;C ði; u; c; nÞ  emax
U ðu; kÞ
i:ði;uÞ2SI;U c:ði;cÞ2SI;C
X
VI ði; nÞ  Emin
I ði; cÞ  VI;C ði; c; nÞ 8ði; cÞ 2 SI;C ; n (A13a) VI;U ði; u; nÞ 8ði; uÞ 2 SI;U (A25b)
i:ði;uÞ2SI;U
VI;C ði; c; nÞ  VI ði; nÞ  Emax
I ði; cÞ 8ði; cÞ 2 SI;C ; n 0 1
X X
(A13b)
U ðu; kÞ
emin @ qc  VI;U;C ði; u; c; nÞA
i:ði;uÞ2SI;U c:ði;cÞ2SI;C
VI;U;C ði; u; c; nÞ ¼ Einit
I ði; cÞ  VI;U ði; u; nÞ
X X
(A14)  eC ðc; kÞ  qc  VI;U;C ði; u; c; nÞ
8ði; uÞ 2 SI;U ; ði; cÞ 2 SI;C ; ði; nÞ 2 SF;I ; n
i:ði;uÞ2SI;U c:ði;cÞ2SI;C

VI;U;C ði; u; c; nÞ ¼ EI;C ði; c; n  1Þ  VI;U ði; u; nÞ 8ði; uÞ 2 SI;U (A26a)


(A15) X X
8ði; uÞ 2 SI;U ; ði; cÞ 2 SI;C ; ði; nÞ 62 SF;I ; n eC ðc; kÞ  qc  VI;U;C ði; u; c; nÞ  emax
U ðu; kÞ
i:ði;uÞ2SI;U c:ði;cÞ2SI;C
VI;C ði; c; nÞ ¼ EI;C ði; c; nÞ  VI ði; nÞ 8ði; cÞ 2 SI;C ; n 0 1
X X
(A16) @ qc  VI;U;C ði; u; c; nÞA 8ði; uÞ 2 SI;U
i:ði;uÞ2SI;U c:ði;cÞ2SI;C
Xðp; i; nÞ þ Yði; u; nÞ  1 8ðp; iÞ 2 SP;I ; ði; uÞ 2 SI;U
(A17) (A26b)
X
Yði; u; nÞ  2 8i; n (A18a) X
u:ði;uÞ2SI;U VU ðu; nÞ ¼ DðuÞ 8u (A27)
X n
Yði; u; nÞ  2 8u; n (A18b) start
TP;I ðp; i; n þ 1Þ  TP;I
end
ðp; i; nÞ 8ðp; iÞ 2 SP;I ; n \ N (A28)
i:ði;uÞ2SI;U

h i start
TP;I ðp; i0 ; n þ 1Þ  TP;I
end
ðp; i; nÞ  H ½1  Xðp; i; nÞ
VI;U ði; u; nÞ  I;U ði; uÞ
Fmax  end
TI;U ði; u; nÞ  start
TI;U ði; u; nÞ
8ðp; iÞ 2 SP;I ; ðp; i0 Þ 2 SP;I ; n \ N; Dn ¼ 1 (A29)
8ði; uÞ 2 SI;U ðA19aÞ
start 0
h i TP;I ðp ; i; n þ 1Þ  TP;I
end
ðp; i; nÞ  H ½1  Xðp; i; nÞ
I;U ði; uÞ  TI;U ði; u; nÞ  TI;U ði; u; nÞ  VI;U ði; u; nÞ
Fmin 8ðp; iÞ 2 SP;I ; ðp0 ; iÞ 2 SP;I ; p 6¼ p0 ; n \ N
end start
(A30)
8ði; uÞ 2 SI;U (A19b) start
TP;I ðp; i0 ; n þ 1Þ  TP;I
end
ðp; i; nÞ þ H½2  Xðp; i; nÞ
VI;U ði; u; nÞ  max
VI;U ði; u; nÞ  Yði; u; nÞ 8ði; uÞ 2 SI;U Xðp; i0 ; n þ 1Þ
(A20) 8ðp; iÞ 2 SP;I ; ðp; i0 Þ 2 SP;I ; n \ N; Dn ¼ 1 (A31)
X
VI;U ði; u; nÞ ¼ VI;U;C ði; u; c; nÞ 8ði; uÞ 2 SI;U
c:ði;cÞ2SI;C TBstart ðnÞ  TP;I
start
ðp; i; nÞ þ H ½1  Xðp; i; nÞ
(A21) 8ðp; iÞ 2 JP; ðp; iÞ 2 SP;I ; n \ N; B > 1 (A32a)

AIChE Journal August 2012 Vol. 58, No. 8 Published on behalf of the AIChE DOI 10.1002/aic 2391
 
TBend ðnÞ  TP;I
start
ðp; i; nÞ  H½1  Xðp; i; nÞ TCW ðvÞ  CSEA  TPend ðpÞ  TARR ðpÞ  TULD
min
ðvÞ
8ðp; iÞ 2 JP; ðp; iÞ 2 SP;I ; n \ N; B > 1 ðA32bÞ 8ðv; pÞ 2 SLV;P (A42)
 
start
TI;U ði; u; n þ 1Þ  TI;U
end
ði; u; nÞ 8ði; uÞ 2 SI;U ; n \ N P P
VI ði; nÞ þ VI ðiÞ
init
(A33) SSP  SS  i n
(A43)
Nþ1
start
TP;I ðp; i; n þ 1Þ  end
TI;U ði; u; nÞ  H ½1  Yði; u; nÞ
8ðp; iÞ 2 SP;I ; ði; uÞ 2 SI;U ; n \ N (A34)
Objective
start
TI;U ði; u0 ; n þ 1Þ  end
TI;U ði; u; nÞ  H½1  Yði; u; nÞ XXXX
8ði; uÞ 2 SI;U ; ði; u Þ 2 SI;U ; u 6¼ u0 ; n \ N
0
(A35) PROFIT ¼ CPROF ðcÞ  VI;U;C ði; u; c; nÞ
i u c n
X XX
TUstart ðu; nÞ  TI;U
start
ði; u; nÞ  H ½1  Yði; u; nÞ  TCW ðvÞ  CPEN  SSP  H  CSET  zðu; nÞ (A44)
v u n
8ði; uÞ 2 SI;U ; n ðA36aÞ
Hard Bounds
TUstart ðu; nÞ  TI;U
start
ði; u; nÞ þ H ½1  Yði; u; nÞ
8ði; uÞ 2 SI;U ; n ðA36bÞ 0  xeðp; nÞ  1 8p; n (A45)

TUend ðu; nÞ  TI;U


end
ði; u; nÞ  H ½1  Yði; u; nÞ 0  zðu; nÞ  1 8u; n (A46)
8ði; uÞ 2 SI;U ; n (A37a) SSP  SS (A47)

TUend ðu; nÞ  TI;U


end
ði; u; nÞ þ H½1  Yði; u; nÞ I;C ði; c; nÞ ¼ 0; EI;C ði; c; nÞ ¼ 1
Emin max
8ði; cÞ 2 SI;C ; n
8ði; uÞ 2 SI;U ; n (A37b) (A48)
start
TI;U ði; u; n þ 1Þ  TP;I
end
ðp; i; nÞ þ ST  Xðp; i; nÞ VImin ði; nÞ ¼ VImin ðiÞ; VImax ði; nÞ ¼ VImax ðiÞ 8i; n
 H ½1  Xðp; i; nÞ 8ðp; iÞ 2 SP;I ; ði; uÞ 2 SI;U ; n \ N (A38) (A49)
X TARR ðpÞ  TPstart ðpÞ  H 8p (A50a)
VI;C ði; c; nÞ ¼ VI;C ði; c; n  1Þ þ VP;I ðp; i; nÞ  EP ðp; cÞ
X
p:ðp;cÞ2SP;C
TARR ðpÞ  TPend ðpÞ  H 8p (A50b)
 VI;U;C ði; u; c; nÞ 8ði; cÞ 2 SI;C ; n > 1 (A39a)
u:ði;uÞ2SI;U
start
TP;I ðp; i; nÞ  H; TP;I
end
ðp; i; nÞ  H 8ðp; iÞ 2 SP;I ; n
X (A51a;b)
VI;C ði; c; nÞ ¼ VI;C
init
ði; cÞ þ VP;I ðp; i; nÞ  EP ðp; cÞ
p:ðp;cÞ2SP;C
X TBstart ðnÞ  H; TBend ðnÞ  H 8n (A52a;b)
 VI;U;C ði; u; c; nÞ 8ði; cÞ 2 SI;C ; n ¼ 1 (A39b)
u:ði;uÞ2SI;U start
TI;U ði; u; nÞ  H; TI;U
end
ði; u; nÞ  H 8ði; uÞ 2 SI;U ; n
X 
TUend ðu; nÞ  TUstart ðu; nÞ ¼ H 8u (A40) (A53a;b)
n

zðu; nÞ  Yði; u; nÞ  Yði; u; n þ 1Þ 8ði; uÞ 2 SI;U ; n \ N TUstart ðu; nÞ  H; TUend ðu; nÞ  H 8u; n (A54a;b)
(A41a)

zðu; nÞ  Yði; u; n þ 1Þ  Yði; u; nÞ 8ði; uÞ 2 SI;U ; n \ N Bound updates based on topology


(A41b) First, the value of VImax ði; nÞ is given by

8 init P
>
> VI ðiÞ þ VPinit ðpÞ if Dn ¼ 0 and n ¼ 1
>
>
>
>
p:ðp;iÞ2SP;I
>
< VPinit ðpÞþVIinit ðiÞVImax ðiÞ
( )
VI ði; nÞ ¼
max
P 8i (A55)
>
> min VI ðiÞ; VI ðiÞ þ
max init
VP ðpÞ
init
if Dn ¼ 1 and n ¼ 1
>
>
>
>
>
: max
p:ðp;iÞ2SP;I

VI ðiÞ if n > 1

I;C ði; c; nÞ ¼ 0 and EI;C ði; c; nÞ ¼ 1


Initialization: Emin max

2392 DOI 10.1002/aic Published on behalf of the AIChE August 2012 Vol. 58, No. 8 AIChE Journal
8 8 9
>
> >
> >
>
>
> >
< >
=
>
> init
ði;cÞ init
ði;cÞ
>
> max V init ðiÞþ
VI;C
P ;
VI;C
if VIinit ðiÞ > 0 and Dn ¼ 0
>
> > VP ðpÞ
init V max ði;nÞ
>
>
> >
>
I I
>
>
< : p:ðp;iÞ2SP;I ; p:ðp;cÞ62SP;C ;
EI;C ði; c; nÞ ¼
min ðpÞþV ðiÞV ðiÞ 8ði; cÞ 2 SI;C ; n ¼ 1
init init max
8 V P I I 9
>
> < =
>
>
> ði;cÞ ði;cÞ
init init
>
> max V init ðiÞþ P
VI;C
;
VI;C
if VIinit ðiÞ > 0 and Dn ¼ 1
>
> : I VPinit ðpÞ VI ði;nÞ;
max
>
>
>
:
p:ðp;iÞ2SP;I ; p:ðp;cÞ62SP;C

0 if VIinit ðiÞ ¼ 0
(A56a)
(
I;C ði;c;n1ÞVI ði;n1Þ
Emin min
if VImax ði; nÞ > 0
I;C ði; c; nÞ ¼
Emin 8ði; cÞ 2 SI;C ; n > 1
VImax ði;nÞ (A56b)
0 if VImax ði; nÞ ¼ 0
8 init
½EI;C ði;cÞ1½VImin ðiÞ
r
>
> þ1 if VImin ðiÞ > 0
>
> ½VI ðiÞþVP ðpÞ
r
>
min init
>
>

< EI;C ði; cÞ


init
if VImin ðiÞ ¼ 0 and max VPinit ðpÞ ¼ 0 X
p:ðp;iÞ2SP;I ;p:ðp;cÞ2SP;C
I;C ði; c; nÞ
Emax ¼ 8ði; cÞ 2 SI;C ; n; r ¼ 1
>
> VPinit ðpÞþVIinit ðiÞVImax ði;nÞ

>
> p:ðp;iÞ2SP;I
>
> 1 if VImin ðiÞ ¼ 0 and max VPinit ðpÞ > 0
>
: p:ðp;iÞ2SP;I ;p:ðp;cÞ2SP;C
p:ðp;cÞ2SP;C

VPinit ðpÞþVIinit ðiÞVImax ði;nÞ

(A57)
8 8 init P 9
>
> >
> VI;C ði;cÞþ VPinit ðpÞ >
>
>
> >
> p:ðp;iÞ2SP;I ;p:ðp;cÞ2SP;C >
>
>
> < =
>
> VPinit ðpÞþVIinit ðiÞVImax ði;nÞ
P V init ði;cÞþV max ði;nÞV init ðiÞ
>
> min ; I;C I I
if VIinit ðiÞ > 0
>
> >
> VIinit ðiÞþ VPinit ðpÞ VImax ði;nÞ >
>
>
> >
> ;p:ðp;cÞ2S >
>
>
< : p:ðp;iÞ2S P;I P;C
;
V init ðpÞþV init ðiÞV max ði;nÞ
I;C ði; c; nÞ ¼

P I I
Emax
>
> 1 if VIinit ðiÞ ¼ 0 and max VPinit ðpÞ > 0
>
> p:ðp;iÞ2SP;I ;p:ðp;cÞ2SP;C
>
>
>
> VPinit ðpÞþVIinit ðiÞVImax ði;nÞ
>
>

>
>0 if VIinit ðiÞ ¼ 0 and VPinit ðpÞ ¼ 0
>
> max
>
: p:ðp;iÞ2SP;I ;p:ðp;cÞ2SP;C
VPinit ðpÞþVIinit ðiÞVImax ði;nÞ

8ði; cÞ 2 SI;C ; n ¼ 1; D ¼ 0 (A58a)


8 8 init P 9
>
> < VI;C ði;cÞþ VPinit ðpÞ =
>
> P
p:ðp;iÞ2SP;I ;p:ðp;cÞ2SP;C
init
VI;C ði;cÞþVImax ði;nÞVIinit ðiÞ
>
> min V init ðiÞþ ; if VIinit ðiÞ > 0
>
> : I VPinit ðpÞ VImax ði;nÞ ;
< p:ðp;iÞ2SP;I ;p:ðp;cÞ2SP;C

I;C ði; c; nÞ ¼
Emax

>
> 1 if VIinit ðiÞ ¼ 0 and max VPinit ðpÞ > 0
>
> p:ðp;iÞ2SP;I ;p:ðp;cÞ2SP;C
>
> init

>
> if VIinit ðiÞ ¼ 0 and VP ðpÞ ¼ 0
:0 max
p:ðp;iÞ2SP;I ;p:ðp;cÞ2SP;C

8ði; cÞ 2 SI;C ; n ¼ 1; Dn ¼ 1 ðA58bÞ

I;C ði; c; nÞ ¼
Emax
8 8 Emax ði;c;n1ÞV min ði;n1Þþ P 9
VPinit ðpÞ
>
> >
> I;C I >
>
>
> >
> p:ðp;iÞ2SP;I ;p:ðp;cÞ2SP;C >
>
>
> >
> >
>
> > ðpÞþV ði;n1ÞV ði;nÞ >
init min max
>
> <
V P P I I
; =
>
> VI ði;n1Þþ
min VP ðpÞ
init
>
> min if VImin ði; n  1Þ > 0
>
> >
>
p:ðp;iÞ2SP;I ;p:ðp;cÞ2SP;C >
>
>
> >
> VPinit ðpÞþVImin ði;n1ÞVImax ði;nÞ >
>
>
< >
> >
>
>
: I;C
E max
ði;c;n1ÞV min
I ði;n1ÞþV max
I ði;nÞV min
I ði;n1Þ >
;
> VI ði;nÞ
max
>
>

>
> if VImin ði; n  1Þ ¼ 0 and VPinit ðpÞ > 0
>
>
1 max
>
> p:ðp;iÞ2SP;I ;p:ðp;cÞ2SP;C
>
> VPinit ðpÞþVImin ði;n1ÞVImax ði;nÞ
>
>

>
>
>
> I;C ði; c; n  1Þ
Emax if VImin ði; n  1Þ ¼ 0 and max VPinit ðpÞ ¼ 0
>
: p:ðp;iÞ2SP;I ;p:ðp;cÞ2SP;C
VPinit ðpÞþVImin ði;n1ÞVImax ði;nÞ

8ði; cÞ 2 SI;C ; n > 1; Dn ¼ 0 (A59a)

AIChE Journal August 2012 Vol. 58, No. 8 Published on behalf of the AIChE DOI 10.1002/aic 2393
I;C ði; c; nÞ ¼
Emax
8 8 max P 9
>
> >
>
EI;C ði;c;n1ÞVImin ði;n1Þþ VPinit ðpÞ >
>
>
> >
> >
;>
;p:ðp;cÞ2S
>
> < P P;I
p:ðp;iÞ2S P;C
=
>
> V min ði;n1Þþ V init ðpÞ
>
> min I P
if VImin ði; n  1Þ > 0
>
> >
>
p:ðp;iÞ2S P;I ;p:ðp;cÞ2S P;C >
>
< >
> >
>
: Emax
I;C ði;c;n1ÞVI ði;n1ÞþVI
min max
ði;nÞVImin ði;n1Þ ;
> VI ði;nÞ
max
>
>

>
> if VImin ði; n  1Þ ¼ 0 and VPinit ðpÞ > 0
>
>
1 max
>
> p:ðp;iÞ2SP;I ;p:ðp;cÞ2SP;C
>
> init

>
: Emax ði; c; n  1Þ if VImin ði; n  1Þ ¼ 0 and max VP ðpÞ ¼ 0
I;C
p:ðp;iÞ2SP;I ;p:ðp;cÞ2SP;C

8ði; cÞ 2 SI;C ; n > 1; Dn ¼ 1 (A59b)

We take the maximum value among Eqs. A57–A59 as n o


max
VP;I ðp; i; nÞ ¼ min VP;I
max
ðp; iÞ  H; Vpinit 8ðp; iÞ 2 SP;I ; n
I;C ði; c; nÞ.
final Emax
(A60)
max
The value of VP;I ðp; i; nÞ is computed by
min
The values of VI;U ði; u; nÞ and VI;U
max
ði; u; nÞ are given by:

8 min
> VI;U ði; u; nÞ ¼ 0 8n
>
> n o
< max
VI;U ði; u; nÞ ¼ min Fmax
I;U ði; uÞ  H; VI  VI ði; nÞ; VI
init min max
ðiÞ  VImin ðiÞ if n ¼ 1
max 8ði; uÞ 2 SI;U (A61)
>
> FI;U ði; uÞ  H; VImax ði; n  1Þ  VImin ði; nÞ
>
: VI;U
max
ði; u; nÞ ¼ min if n > 1
VImax ðiÞ  VImin ðiÞ

The entire nonconvex MINLP problem denoted as MP is Appendix B: Branch and Bound Global
presented as follows: Optimization Algorithm

ðMPÞ Min PROFIT


s:t: Eqs: A1A43 Piecewise-Linear Underestimators
Eqs: A45A54 and Eqs: A55A61 In the branch and bound global optimization algorithm,
variables EI,C(i,c,n) are uniformly partitioned and the nf4r
The resulting mathematical model MP is a nonconvex formulation30 is used to piecewise-underestimate each of the
MINLP and the sources of nonconvexities are the distinct bilinear terms in the model. The details about nf4r formula-
bilinear terms (i.e., Eqs. A15 and A16) presented as follows: tion can be found in the paper of Gounaris et al.30 and
Misener and Floudas.25
VI;U;C ði; u; c; nÞ ¼ EI;C ði; c; n  1Þ  VI;U ði; u; nÞ The relaxation of the bilinear terms (Eqs. A15,A16) using
8ði; uÞ 2 SI;U ; ði; cÞ 2 SI;C ; ði; nÞ 62 SF;I ; n ðA150 Þ the nf4r formulation is given by,

VI;C ði; c; nÞ ¼ EI;C ði; c; nÞ  VI ði; nÞ 8ði; cÞ 2 SI;C ; n (A160 )

8
> P
GR
>
> V ði; u; c; nÞ  E ði; c; n  1Þ  V min
ði; u; nÞ þ EI;C ði; c; n  1; gr  1Þ  DVI;U ði; u; c; n; grÞ
>
>
I;U;C I;C I;U
>
> gr¼1
>
>
>
> VI;U;C ði; u; c; nÞ  EI;C ði; c; n  1Þ  VI;U max
ði; u; nÞþ
>
>
>
> n h io
>
> P
GR
>
> E ði; c; n  1; grÞ  DV ði; u; c; n; grÞ  ðV max
ði; u; nÞ  V min
ði; u; nÞÞ  k ði; c; n  1; grÞ
< gr¼1 I;C I;U I;U I;U I;C

>
> P
GR
>
> VI;U;C ði; u; c; nÞ  EI;C ði; c; n  1Þ  VI;U
min
ði; u; nÞ þ EI;C ði; c; n  1; grÞ  DVI;U ði; u; c; n; grÞ
>
>
>
> gr¼1
>
>
>
> VI;U;C ði; u; c; nÞ  EI;C ði; c; n  1Þ  VI;U
max
ði; u; nÞþ
>
>
>
> n h io
>
> P
GR
>
: EI;C ði; c; n  1; gr  1Þ  DVI;U ði; u; c; n; grÞ  ðVI;Umax
ði; u; nÞ  VI;U
min
ði; u; nÞÞ  kI;C ði; c; n  1; grÞ
gr¼1

8ði; uÞ 2 SI;U ; ði; cÞ 2 SI;C ; n ðB1Þ

2394 DOI 10.1002/aic Published on behalf of the AIChE August 2012 Vol. 58, No. 8 AIChE Journal
8
> P
GR
>
> VI;C ði; c; nÞ  EI;C ði; c; nÞ  VImin ði; nÞ þ EI;C ði; c; n; gr  1Þ  DVI ði; c; n; grÞ
>
>
>
> gr¼1
>
> VI;C ði; c; nÞ  EI;C ði; c; nÞ  VImax ði; nÞþ
>
>
>
> GR
P  

>
>
>
< EI;C ði; c; n; grÞ  DVI ði; c; n; grÞ  ðVImax ði; nÞ  VImin ði; nÞÞ  kI;C ði; c; n; grÞ
gr¼1
8ði; cÞ 2 SI;C ; n (B2)
>
> P
GR
>
> VI;C ði; c; nÞ  EI;C ði; c; nÞ  VImin ði; nÞ þ EI;C ði; c; n; grÞ  DVI ði; c; n; grÞ
>
>
>
> gr¼1
>
> VI;C ði; c; nÞ  EI;C ði; c; nÞ  VImax ði; nÞþ
>
>
>
> GR
P  

>
> EI;C ði; c; n; gr  1Þ  DVI ði; c; n; grÞ  ðVImax ði; nÞ  VImin ði; nÞÞ  kI;C ði; c; n; grÞ
:
gr¼1

When the number of grid points is equal to 2, the nf4r formulation is reduced to McCormick convex and concave envelopes
as follows,

VI;U;C ði; u; c; nÞ
8
>
>  EI;C ði; c; n  1Þ  VI;U
min
ði; u; nÞ þ Emin
I;C ði; c; n  1Þ  VI;U ði; u; nÞ  EI;C ði; c; n  1Þ  VI;U ði; u; nÞ
min min
>
>
<  E ði; c; n  1Þ  V min ði; u; nÞ þ Emax ði; c; n  1Þ  V ði; u; nÞ  Emax ði; c; n  1Þ  V min ði; u; nÞ
I;C I;U I;C I;U I;C I;U
8ði; uÞ 2 SI;U ; ði; cÞ 2 SI;C ; n
>
>  EI;C ði; c; n  1Þ  VI;U ði; u; nÞ þ EI;C ði; c; n  1Þ  VI;U ði; u; nÞ  EI;C ði; c; n  1Þ  VI;U ði; u; nÞ
max min min max
>
>
:
 EI;C ði; c; n  1Þ  VI;U
max
ði; u; nÞ þ Emax
I;C ði; c; n  1Þ  VI;U ði; u; nÞ  EI;C ði; c; n  1Þ  VI;U ði; u; nÞ
max max

(B3)
8
>
>  EI;C ði; c; nÞ  VImin ði; nÞ þ Emin
I;C ði; c; nÞ  VI ði; nÞ  EI;C ði; c; nÞ  VI ði; nÞ
min min
>
>
<  E ði; c; nÞ  V min ði; nÞ þ Emax ði; c; nÞ  V ði; nÞ  Emax ði; c; nÞ  V min ði; nÞ
I;C I I;C I I;C I
VI;C ði; c; nÞ 8ði; cÞ 2 SI;C ; n (B4)
>
>  E ði; c; nÞ  V max
ði; nÞ þ E min
ði; c; nÞ  V ði; nÞ  E min
ði; c; nÞ  V max
ði; nÞ
>
>
I;C I I;C I I;C I
:
 EI;C ði; c; nÞ  VImax ði; nÞ þ Emax
I;C ði; c; nÞ  V I ði; nÞ  E max
I;C ði; c; nÞ  V I
max
ði; nÞ

The piecewise-linear relaxation of problem MP (see branching and branching on VI(i,n) where (i, n) contributes
Appendix A) is denoted as problem (RMP) and defined as to the greatest discrepancy between the auxiliary and origi-
follows, nal problem variables:
8 9
ðRMPÞ Min PROFIT < X  =
s:t: Eqs: A1A14; A17A43; and max V^I;C ði; c; nÞ  E^I;C ði; c; nÞ  V^I ði; nÞ 8i; n
B1B2 or B3B4 i;n :c:ði;cÞ2S ;
I;C
Eqs: A45A54 and Eqs: A55A61
(B5)

where, (V^I , V^I;C , and E^I;C ) are the optimal solutions gener-
Branching Strategy ated from the piecewise-linear relaxation for a given node.
After solving a relaxation of each node using the piece- The branching point on VI(i,n) is 10% away from its optimal
wise-linear underestimator, the variable VI(i,n) is selected for solution:

8  
^ if 1:1  V^I ði; nÞ\ 12 VImin ði; nÞ þ VImax ði; nÞ
branching < 1:1  VI ði; nÞ
>
 
VIpoint ði; nÞ ¼ 0:9  V^I ði; nÞ if 0:9  V^I ði; nÞ > 1 V min ði; nÞ þ V max ði; nÞ (B6)
>
: 1  min  2 I I

2 VI ði; nÞ þ VI ði; nÞ
max
otherwise

Solution Improvement Strategy discrepancy. Then, the binary variables from solution S1 is
34
The refinement strategy from Li et al. is used to improve fixed to get a NLP. This alternating series of MILP and NLP
the quality of each solution from Pool-2. Let S denote any continues, until the solutions of successive NLPs converge.
solution from Pool-2. First, the compositions of tanks from S All improved solutions for Pool-2 form another solution pool
are extracted and fixed in the MINLP to get an MILP. A so- (denoted as Pool-3). The best solution in Pool-3 is selected
lution S1 to this MILP is a schedule with no composition as the final UB.

AIChE Journal August 2012 Vol. 58, No. 8 Published on behalf of the AIChE DOI 10.1002/aic 2395
Optimality-Based Tightening Lower and Upper Bounds Additional Strategy
The lower and UBs for variables EI,C(i,c,n), VI(i,n), and When solving an MILP relaxation using the piecewise-lin-
VI,U(i,u,n) can be further tighten by solving LP minimization ear underestimators at each node, the relative convergence of
and maximization problems with Eqs. A1–A44 and McCor- this MILP relaxation is adjusted based on the difference
mick convex and concave envelopes (i.e., Eqs. B3 and B4) between the best lower and UBs in the tree. It is required to
to relax the bilinear terms (i.e., Eqs. A15 and A16). The be less than some ratio denoted as c (e.g., 0.05) of the cur-
minimization and maximization problems can be stated as rent relative gap between the LB and UB bounds. In other
follows words,

ðMPLÞ Min z obbt  


UB  LB
s:t: Eqs: A1A14; and A17A43; Relative Gap ¼ c    (B7)
A45A54; A55A61; and B3B4 LB 
0  Xðp; i; nÞ; Yði; u; nÞ  1
ðMPUÞ Max z obbt Note that as the MILP relaxation at each node may not be
s:t: Eqs: A1A14; and A17A43; A45A54; solved to optimality, the ‘‘best estimate’’ is always considered
A55A61; and B3B4 as the results of each node in the branch and bound tree.
0  Xðp; i; nÞ; Yði; u; nÞ  1
where z_obbt becomes VI(i,n), EI,C(i,c,n), and VI,U(i,u,n),
respectively, to update VImin ði; nÞ; Emin I;C ði; c; nÞ VI
max
ði; nÞ; Manuscript received Apr. 18, 2011, revision received July 19, 2011, and final
EI;C ði; c; nÞ; and VI;U ði; u; nÞ accordingly.
max max
revision received Sept. 9, 2011.

2396 DOI 10.1002/aic Published on behalf of the AIChE August 2012 Vol. 58, No. 8 AIChE Journal