Está en la página 1de 31

This article was downloaded by:[University of Waterloo]

On: 10 October 2007


Access Details: [subscription number 769429802]
Publisher: Taylor & Francis
Informa Ltd Registered in England and Wales Registered Number: 1072954
Registered office: Mortimer House, 37-41 Mortimer Street, London W1T 3JH, UK

Engineering Optimization
Publication details, including instructions for authors and subscription information:
http://www.informaworld.com/smpp/title~content=t713641621
A ROLLING HORIZON HEURISTIC FOR REACTIVE
SCHEDULING OF BATCH PROCESS OPERATIONS
Ali Elkamel a; Atul Mohindra b
a
Department of Chemical Engineering, University of Kuwait, Safat, Kuwait
b
The Foxboro Company, Foxboro, MA, USA

Online Publication Date: 01 August 1999


To cite this Article: Elkamel, Ali and Mohindra, Atul (1999) 'A ROLLING HORIZON
HEURISTIC FOR REACTIVE SCHEDULING OF BATCH PROCESS
OPERATIONS', Engineering Optimization, 31:6, 763 - 792
To link to this article: DOI: 10.1080/03052159908941396
URL: http://dx.doi.org/10.1080/03052159908941396

PLEASE SCROLL DOWN FOR ARTICLE

Full terms and conditions of use: http://www.informaworld.com/terms-and-conditions-of-access.pdf


This article maybe used for research, teaching and private study purposes. Any substantial or systematic reproduction,
re-distribution, re-selling, loan or sub-licensing, systematic supply or distribution in any form to anyone is expressly
forbidden.
The publisher does not give any warranty express or implied or make any representation that the contents will be
complete or accurate or up to date. The accuracy of any instructions, formulae and drug doses should be
independently verified with primary sources. The publisher shall not be liable for any loss, actions, claims, proceedings,
demand or costs or damages whatsoever or howsoever caused arising directly or indirectly in connection with or
arising out of the use of this material.
Downloaded By: [University of Waterloo] At: 18:30 10 October 2007

Eng. Opt.. 1999. Vol 31. pp. 763-792 0 1999 OPA (Omwar Publishers Association) N.V.
Reprints available dirmly from !he publisher Publirhcd by liccnw undcr
Photocopying pcrmill~dby licrnw only the Gordon and Breach Scicncc
Publishers imprint.
Printed in Malaysia.

A ROLLING HORIZON HEURISTIC


FOR REACTIVE SCHEDULING
OF BATCH PROCESS OPERATIONS
ALI ELKAMEL" and ATUL MOHINDRA
" Departmenr of Chemical Engineering. University of Kuwait,
P. 0. Box 5969. 13060 Safat, Kuwait;
The Foxboro Company, 33 Commercial Sireel, C41-2G.
Foxboro, MA 02035, USA

(Received 24 February 1998)

Batch chemical plants are dynamic processing facilities where static production
schedules can rarely be adhered to due to market and operating uncertainties. On-line
schedule modification of a prior; timing assignments and resource allocations in response
to unantipicated disruptions is done through a decomposition heuristic that uses a
rolling horizon implementation policy. An attempt is made to minimize the impact of the
disruptions on the original schedule near the point of each deviation while exploiting the
combinatorial flexibility of task and resource reassignments in future scheduling time
windows. The problem is addressed as a multiobjective optimization problem involving
completion time criteria, relative customer importance, and production cost considera-
tions.
A rigorous analysis of problem sensitive parameters, including penalty weights and
subhorizon length, is conducted. A model plant case study is performed. Variations on
storage availability and task flexibility are investigated in an attempt to characterize
dominant effects of the weighting parameters. Results indicate that user preference can
serve as a strong guide for obtaining near optimal reactive scheduling solutions. It is
shown that the combinatories can be controlled and that costly and inefficient full scale
rescheduling of multipurpose production facilities can be avoided.

Keywords: Chemical plant; scheduling; batch processing

INTRODUCTION

Batch production on a large scale has long been of interest to the


chemical process industry. Since most products are usually developed
in batches a t the bench level, direct scale-up can prove to be an
Downloaded By: [University of Waterloo] At: 18:30 10 October 2007

764 A. ELKAMEL A N D A. MOHINDRA

enticing option. This is especially true for pharmaceuticals and high-


value-added speciality chemicals that provide high rates of return per
unit sold.
The most commonly found facilities for batch manufacturing are
multiproduct and multipurpose plants. Multiproduct facilities are cap-
able of producing a number and variety of products in noninteracting
production lines. Processing units and storage vessels are typically
dedicated to the production of a single product or several closely
related products, only one being produced during a given campaign on
a given line. Multipurpose plants are also capable of producing a
number and variety of products but with the additional flexibility of
having nondedicated processing and storage units. Production lines
interact through the use of vessels to process and store different .
materials, but cleaning cost considerations are significant.
Plant operations are governed by production planning and
scheduling. Planning is a long term activity that is usually combined
with corporate goals and economic forecasting. Scheduling is more of
a short-term activity in that it dictates event times and sequences on
the plant floor for day to day operations. Reactive scheduling involves
the short term adjustment of the master schedule in response to
unanticipated deviations in plant operating parameters. Typically, the
scheduling of process operations has been viewed as a priori time
assignment of tasks to units and materials to storage vessels subject to
the assunlption of deterministic plant parameters. However, in a real
manufacturing facility, there is a limit to the certainty of various
parameters such as market demand, processing time, and resource
availability. This uncertainty, which is characteristic of dynamic
production environments, is the motivating force for developing a
rigorous framework for reactive scheduling. Full scale rescheduling of
an entire plant at the point of every disturbance would be disruptive,
costly, and highly inefficient in terms of preserving smooth operation.
Reactive scheduling represents a n extension of the general schedul-
ing problem. In addition to making decisions about the allocations
of resources in the time and space domains, additional constraints as
to which allocations are preferable are imposed. A complex set of
secondary decisions is introduced into an already involved problem.
The heirarchy of decisions that are involved in scheduling and reactive
Downloaded By: [University of Waterloo] At: 18:30 10 October 2007

REACTIVE BATCH SCHEDULING 765

scheduling are illustrated in Figure I . At the highest level, an


assessment of what to produce and how to produce it is made. These
issues are settled before master schedule generation. Decisions must
then be made about the number and types of orders to consider. These
are based on conditions such as market demands, plant capabili-
ties and capacity, and raw material availability. Once this is done,
decisions must be made about when products can be completed. Next,
the prioritization of different orders is considered. It must be decided

M&r ScluduUw S&v RedwScb&&#SIqr Over@ Re&"

FIGURE I Decision levels in reactive scheduling.


Downloaded By: [University of Waterloo] At: 18:30 10 October 2007

766 A. ELKAMEL A N D A. MOHINDRA

which customers should be given priority when setting due dates in


the master schedule. The allocation of resources available in the plant
must then be decided. Finally, in the purely reactive stages of
scheduling, decisions involving unit replacements and time shifting of
tasks that could not be executed a t predesignated times must be
made. Simultaneously, decisions about resource reallocation and
possible due date modification need to be considered. The overlap
region in the scheduling and reactive scheduling framework, indicated
by the shading in Figure 1, is clearly of significance. Decisions made
a t the higher scheduling level about due-dates, priorities, and resource
utilization have an impact on future decisions that must be made in a
reactive context. Through the appropriate modelling of uncertainties
in meeting due-dates, resource availabilities, and customer prefer-
ences a t the master scheduling level, it is possible that a reasonable
amount of flexibility could be left in the schedule so as to reduce the
disruptive nature of making these difficult choices during reactive
scheduling.
The scheduling problem of process operations has been addressed
by numerous researchers over the years (see for instance Refs. [I -81).
A robust and comprehensive approach has been the State Task Net-
work (STN) developed by Kondili et al. [9]. The major drawback of
this approach is in the uniform discritization of time which lead to
models with a large number of discrete variables. Elkamel et al. [lo]
suggested a decomposition strategy to remedy this situation. Later,
Elkamel and Al-Enezi [ I I] proposed valid inequalities to the Kondili
E I al. model [9] that strengthen the relaxation solutions to the model
and reduce the computational effort.
Even though reactive scheduling has been the subject of much
research by the operations research community [12- 141, only little
work has been reported in the chemical process industry. Cott and
Macchietto [I51 proposed a n Earliest Finishing Unit (EFU) heuristic
as part of an integrated framework for process monitoring, diagnosis,
and control. EFU operates by making use of a number of rules that
enable it to shift task starting times on affected processing units so as
to complete production as quickly as possible without reassigning
batches to other processing units. The authors demonstrate the
applicability of the heuristic to some small and medium sized
problems, and conclude its general validity. N o measure of optimality
Downloaded By: [University of Waterloo] At: 18:30 10 October 2007

REACTIVE BATCH SCHEDULING 767

was taken by this approach and the impact of change on the old
schedule was not considered.
Kanakamedala et a/. [16] present an approach that goes beyond
E F U in terms of finding good feasible solutions to reactive scheduling
problems. They present a Least Impact (LI) heuristic that attempts to
minimize the number of deviations from the original schedule. This
heuristic is based on a beam search and gave a significant improve-
ment over the EFU heuristic for a relatively large plant structure. This
heuristic, however, is not readily adaptable to handle other plant
structures, and the authors call for further work in mathematical
programming that may result in a more rigorous reactive scheduling
framework.
The heuristic techniques described previously for solving the
reactive process scheduling problem tend to be either too broad-
based, such as that developed by Fox er al. [13], resulting in poor
solution quality, or too myopic, such as that developed by
Kanakamedala et a / . [16], resulting in a good, but problem specific
solution. In this paper, a mathematical programming based heuristic
for reactive scheduling is illustrated. This heuristic focuses on the
generation of quality solutions for a broad range of problems. It
combines heuristic rules with a mathematical scheduling technique
known to guarantee optimal solutions through an exact algorithm,
namely, branch and bound. At the time of a disturbance, or set of
disturbances, to the original schedule, the remainder of the
scheduling horizon is divided up into subhorizons of nonuniform
length. Each scheduling subhorizon is represented by an MlLP
formulation of the sequencing constraints and order due dates that
govern that time interval with an objective of keeping production
costs at a minimum.
The disturbances are absorbed into a schedule by being sequentially
forward shifted in time until they can be fully accommodated. This
leads to a multiobjective trade-off cost problem: preserve the schedule
as much as possible, subject to cost considerations, yet also allow
enough changes so that due dates are not missed and the associated
penalty costs are not incurred. This rolling horizon strategy makes it
possible to minimize the cost of changes. This is done by applying a
linear penalty function to the objective function of the M l L P
subproblems near the disturbances. When weighted properly, the
Downloaded By: [University of Waterloo] At: 18:30 10 October 2007

768 A. ELKAMEL A N D A. MOHlNDRA

penalty function can be used to generate a least impact schedule


for the subhorizon near any given disturbance. Further, in more dis-
tant future time periods, where schedules are not as rigidly fixed as in
the present, the subsequent MILP subproblems have diminished pen-
alty terms allowing tasks to be represented into a minimum cost
configuration.
The next section describes the reactive scheduling model in the form
of a mixed integer linear programming formulation and details the
pcnalty function that is applied to the objective function in order to
control reassignments. The following sections describe the rules for
making subhorizon cuts and the selection criteria for the penalty
terms. Later, an example case study is considered and the rolling
horizon heuristic is illustrated. A sensitivity analysis on the various
penalty coefficients is carried out. Finally, conclusions about the
performance of the heuristic are given along with future improvement
considerations.

MATHEMATICAL MODEL FOR REACTIVE


SCHEDULING REPRESENTATION

The development of the reactive scheduling model is based on the state


task network (STN) methodology of Kondili er ul. [9] as revised and
amplified by Elkamel e/ al. [17]. This formulation, in its general form,
can be used to schedule the general chemical plant since constraints
involving continuous units are included in addition to those governing
batch operations. The key assumption in this model is the uniform dis-
critization of time. The model defines the following variables with
respect to the type of unit within the processing facility.

Batch Units

I if task istarts on batch unitjat the beginning of time period I;


w, = 0 otherwise

Biil = Amount of material that task i starts processing in batch unit j


a t the start of time period t.
Downloaded By: [University of Waterloo] At: 18:30 10 October 2007

REACTIVE BATCH S C H E D U L l N G 769

These variables serve the purpose of indicating when a batch recipe


task is assigned, to which unit it is assigned, and how much material is
to be processed during the assignment.

Continuous Units

yg~= { 1 if task i is being processed on continuous unit j


a t the start of time period t;
0 otherwise

Qg, = Amount of material finished being processed by task i in


continuous unit j a t the end of time period I .
These variables are the continuous unit analogs of the above variables
for batch processing units. Q,, denotes material that has completed a
processing step since continuous processing is rate dependent.

Storage Vessels

1 if storage unit j is used to store states during time interval r ;


z,3j,= 0 otherwise

F,,, = Amount of state s that is being stored in storage vessel j during


time period I.
In the non-reactive model, constraints were written for the
allocation of units to tasks and for the nonpreemption of tasks.
Various types of units were considered including purely batch units,
continuous units, purely storage units, and batch units with allowed
storage. The main constraints in the model are the overall material
balance constraints. These constraints included terms for the pur-
chasing of raw materials and the delivery of finished goods. Various
other constraints were also employed in the model. The capacity
limitation constraints state that the batch sizes and the processing rates
may range between some minimum production level and some upper
physical bound. The resource constraints employed two types of re-
sources: renewable and non-renewable resources. The equipment
cleaning constraints were designed to handle the cleaning of equipment
items.
Downloaded By: [University of Waterloo] At: 18:30 10 October 2007

770 A. ELKAMEL AND A . MOHlNDRA

The aim of solving the non-reactive scheduling problem was taken


to be to minimize the cost of production. The major costs considered
were the costs of feedstock, inventory costs, deadline violation costs,
and machine operating costs. Additionally, there are rewards (profits)
for satisfying orders. There are additional goals which arise in the
reactive scheduling problem, and these involve penalty functions for
attaining a least-impact schedule that still meets as many order due
dates as possible. Flexibility in selecting the weights also enables the
introduction of other objectives such as order prioritization and pre-
ferential task-unit-time reassignments. The reminder of this section
defines the components of the linear penalty function that are added to
the objective function of the non-reactive scheduling model. In addi-
tion, the modifications to the constraint set for establishing control
over lot size changes are defined.
In the approach of Kanakamedala er 01. [I61 least impact reactive
scheduling is merely defined as an attempt to minimize the number
of time shifts and unit replacements in the reactive solution from what
was originally scheduled. Along these lines, a simple linear penalty
function can be devised which only considers the least impact schedule
modification.objective. For batch units, the function takes the form:

and can be added to the objective function of the non-reactive


scheduling model. Penalty terms for other units can also be written in
a similar fashion. The goal of this function is simply to attempt to
preserve all the batch task-unit-time assignments in a given schedule.
By assignment is meant the configuration specified in the original
schedule. Criteria for selecting the penalty weight, au, can be as simple
as

I. = { >< 00 if Wi,, = I in the original schedule


if Wvl = 0 in the original schedule (2)

Clearly the effect of these penalty terms would be to force the solver to
keep assignments as much as possible the same as originally planned.
However, such a simple function would oversimplify the reactive
Downloaded By: [University of Waterloo] At: 18:30 10 October 2007

REACTIVE BATCH SCHEDULING 77 1

schedule modification problem. The lumping of penalties for making


various potential readjustments into one term, a*, inherently assumes
that all schedule changes are of equal impact. In practice however, it
may, for example, be far less costly to make a time shifting adjustment
than a unit replacement. Another shortcoming of this function is the
fact that, although positive values of aij, could represent real costs of
schedule modifications, negative values d o not have any physical
significance. There is no cash payback in reality for not making a
schedule modification. Thus, selection of negative coefficients would
be purely heuristic and could easily lead to solutions without any
measure of the trade-offs associated with keeping that task-unit-time
assignment as opposed to making an alternate assignment.
In order to introduce full flexibility in reactive scheduling, a measure
of the trade-off costs associated with choosing penalty weights to pre-
serve schedules should be determined according to an appropriate
weighting of the following objectives
1. Simple time shifting of a task-unit assignment only.
2. Unit replacement for a task-time assignment only.
3. Batch size preservation for a particular task-unit-time assignment.
4. Resource purchase modifications and reallocations over the schedul-
ing subhorizon.

Assignment Penalties

These are the penalties on binary task-unit-time allocation variables.


A high penalty cost on a particular variable will have the effect of driving
it to zero. Physically, this is the equivalent of saying "Avoid making this
task-unit-time assignment because a high cost will be incurred".

(i) Task Time Shifting Pcnaltj~


Batch Units
Downloaded By: [University of Waterloo] At: 18:30 10 October 2007

772 A. ELKAMEL AND A . MOHINDRA

Continuous Units

where
t,, = Start time of event disruptive to the original schedule.
/I1= Subhorizon length of desired penalized interval.
(,: = Weighted penalty cost of starting task i o n batch u n i t j a t time t .
,JI; = Weighted penalty cost of starting task i o n continuous unit j a t
time I.
The constants oT and y: equal zero for the original assignment
times. In other words (17
= 0 at those times when Wv, = 1, and yT = 0
when Yii, = I in the original formulation solution. For those schedules
that are not generated using the same scheduling model, the original
assignment times are assigned formulation variable values correspond-
ing to a model solution. This assignment procedure is independent of
the algorithm o r heuristic rules used to generate the original schedule.

(ii) Processing Unit Replacement Penalty

Batch Units

Continuous Units

whcre
B, = Set of allowed replacement units for each batch task i at each
time I.
C, = Set of allowed replacement units for each continuous task i at
each time t .
Downloaded By: [University of Waterloo] At: 18:30 10 October 2007

REACTIVE BATCH SCHEDULING 773

a" = Weighted penalty cost of performing task i on batch unit j at


I
time t .
$ = Weighted penalty cost of performing task i o n continuous unit j
at time 1.

As is the case with a: and "r,


a,? = 0 and $ = 0 for the original
task unit time assignments. Clearly, a: and ay, as well as the con-
tinuous unit variations, are independent parameters that enable more
control over task-unit-time reassignment. Often, in industrial settings,
the cost associated with finding a replacement unit is much higher than
that of maintaining modest holding times for displaced materials. The
segregation of the associated penalty costs is meant to reflect this fact.

(iii) Storage Time Shifr Penalty


Storage vessels are subject to the same types of penalty terms as
processing units. Both time shifting and unit replacement considera-
tions can be imposed by similarly weighted terms.

where
zT = Weighted penalty cost of storing state s in vessel j a t time 1.
I

(iv) Storage Unit Replacement Penalty

where
SV, = Set of allowed storage vessels for each state s a t each time t .
7"
-I
= Weighted penalty cost of storing state s in unit j a t time 1.
Downloaded By: [University of Waterloo] At: 18:30 10 October 2007

774 A. E L K A M E L AND A. MOHINDRA

Batch Size Preservation Penalties


In order to consider the issue of batch size preservation in the new
schedule, introduction of a penalty term is slightly. more involved.
Previously, representation of a linear penalty for attempting to pre-
serve batch size was not clarified. The standard approach would be to
minimize

where
B z w = Continuous variable representing batch size in the new
schedule.
B$* = Continuous value of theoriginal batch size in the unperturbed
schedule.
Equation (9) is nonlinear and would force the problem to become a
MINLP, which is usually more difficult to solve. In order to avoid the
nonlinearity, a construction is applied for linearizing absolute values
in the objective functions of optimization problems, as detailed by
Nemhauser and Wolsey [IS]. Two new continuous variables are added
to the objective function and one additional constraint is included
in the problem specification. Specifically, to overcome the difficulty
of preserving batch size, in addition to task-unit assignments, the
following additions to the previously described penalty function can be
made

where
cry
Vf
= Weighted penalty coefficient representing the cost of increasing
the size of a specific batch from its a priori assigned size.
a;: = Weighted penalty coefficient representing the cost of decreasing
the size of a specific batch from its a priori assigned size.
Positive continuous variable representing positive deviation
from original batch size.
Downloaded By: [University of Waterloo] At: 18:30 10 October 2007

REACTIVE BATCH SCHEDULING 775

itij1 = Positive continuous variable representing negative deviation


from original batch size.
In addition, the following constraints are imposed:

Resource Reallocation and Inventory Control


The parameters introduced here are meant to allow more control over
scheduling conditions than merely penalizing assignment changes.
Though resource profiles are often thought of as secondary to
assignments, they govern the feasible range of scheduling solutions.
Proper adjustment of resource profiles during reactive scheduling
can play a crucial role in meeting desired rescheduling goals and satisfy-
ing order delivery profiles.

(i) Storage Levels

T o the objective function, add

where
ST,T= Set of stable states with allowed storage capacity in the plant.
zf;' = Weighted penalty coefficient representing the cost of storing an
amount of a specific state greater than a user specified amount.
2
-:, = Weighted penalty coefficient representing the cost of storing an
amount of a specific state less than a user specified amount.
p: = Positive continuous variable representing positive deviation from
desired storage level.
n: = Positive continuous variable representing negative deviation
from desired storage level.
Downloaded By: [University of Waterloo] At: 18:30 10 October 2007

776 A. ELKAMEL AND A. MOHINDRA

And to the constraint set add

desired
C C F ~-
rEST, 1
17: = F,,

Where
F:$" = Continuous variable representing amount of states stored a t
time t in the new schedule.
~ d c s i r e d= Constant specifying preferred storage level for state s at time
51

I , possibly that of the original schedule but not necessarily


SO.

(ii) Purclrasing of Raw Materials


The objective function must contain

where
S T / = Set of states that can be purchased as feedstocks.
o $ = Weighted penalty coefficient representing the additional expense
incurred for purchasing an amount of state s greater than some
user specified amount at time r .
cr;; = Weighted penalty coefficient representing the additional expense
incurred for purchasing an amount of state s greater than some
user specified amount at time t .
p:, = Positivecontinuousvariable representingpositivedeviation from
desired purchasing level.
n f l = Positive continuous variable representing negative deviation
from desired purchasing level.
The constraint set must again be expanded to include

H'

C C p::"
TEST, I
+dl- = P,,
desired
Downloaded By: [University of Waterloo] At: 18:30 10 October 2007

REACTIVE BATCH S C H E D U L I N G 777

where
P T W = Continuous variable representing amount of state s pur-
chased at time t in the new schedule.
p desired = purchasing level specified for states a t time t (possibly that of
.TI

the original schedule but not necessarily).

SUBHORIZON LENGTH DETERMINATION

At the time of a disturbance, o r set of disturbances the scheduling


horizon is divided up into different subhorizons. The reason for
dividing up the schedule is threefold. First, doing so provides a mean
of preserving the old schedule as much as possible near the time of the
disturbance. Since sudden changes can be highly costly and potentially
disruptive to smooth operations, segregating the costs of making re-
assignments at different future times is important. Another reason for
decomposition of the schedule is that most realistic problems based on
actual scheduling data usually involve thousands of binary variables
and are intractable by present solution methods [lo]. Thirdly, in any
production scenario, consideration must be given to customers.
Deciding which task assignments and resource allocations are most
flexible with respect to reassignment is likely a direct function of rela-
tive customer importance.
Selection of the subhorizon for reactive scheduling is a sensitive
issue. Where the "cuts" are made can radically effect the solution out-
come because of the sequential nature of product recipes. Since recipes
are connected networks, any form of discretization which attempts to
partition the network risks a loss of information. This loss is attributed
directly to the fact that when a cut is made, information about future
events are not incorporated into the smaller time window. The rules
presented in this section attempt to minimize the effects of information
loss between one scheduling subproblem and the next.
The first issue to be decided before making a cut in the time horizon
is purely logical. Given a priori knowledge of a scheduling algorithm,
such as branch and bound for MILP optimization, and the computing
tools available to execute the algorithm, a decision must be made on
the largest problems of practical size that can be solved. If the
Downloaded By: [University of Waterloo] At: 18:30 10 October 2007

778 A. ELKAMEL AND A. MOHINDRA

available algorithm and tools can optimally schedule the entire


horizon, with reasonable computational effort, then there may not
be a need to make any cuts and risk information loss [lo].
Once the cut is made, backtrack the flow of material from order due
dates and see where that material is supposed to be, according to the
original schedule, a t the time of the cut [lo]. Set demands for the in-
termediates that should be met at the end of the subhorizon and a t the
start of the next time window. If unstable intermediates are present at
the cut, formulation devices must be employed if the cut is to be
retained. The final storage level (target inventory) for the unstable
material, Fq,,,, should be set to the desired value and a n artificial
demand for the material should also be employed. The intermediary
instability can still be denoted by setting'c;: F,, = 0. This is done
only to prevent a loss of information that would otherwise result in the
unstable material not being made.

IMPLEMENTATION STRATEGY

A key aspect to the effective implementation of any heuristic is


automation. This is especially important in an iterative scheme such as
that presented by the rolling horizon heuristic. Solution representation
is also important since an inadequate display of results will lead to a
dilution of the value of the generation technique. The framework for
implementing the heuristic in this work appears in Figure 2. It is the
aim of the reactive scheduling system to be highly interactive. The
operator should, with ease, be able to update information about what
is occurring on the plant floor. The means of doing so, as detailed
below, is a Scheduling language interface using the language RCSPec
developed by Zentner et al. [19]. Followed by the information gathe-
ring stage, the decision making framework that will set subhorizon
lengths, make penalty assignments, and interact with an MlLP solver
to get solution information. The final stage is also interactive in that
graphical output, depicting the new schedule and inventory profiles, is
returned to the operator.
The RCSPec language for representing process scheduling problems
is a natural language means for representing problems without the
need for knowledge about algorithms and heuristics. Details of the
Downloaded By: [University of Waterloo] At: 18:30 10 October 2007

REACTIVE BATCH SCHEDULING

IMPLEMENTATION

%'
Scbcduliig Language

FIGURE 2 Implementation strategy.

language and its inception can be found in the original manuscript. Its
user friendliness is the primary reason for incorporating the language
into the reactive scheduling interface. The main features of the
language that make its application practical are the ability to add new
keywords, the flexibility and ease of modifying problem parameters,
and its suitability to a decomposition strategy. Adding keywords that
have the effect of changing the objective function and constraint set
are essential to the MILP modification scheme of the preceding sec-
tions. The ease of modifying the scheduling parameters is the key to
the success of a rigorous sensitivity analysis. It facilitates modification
of penalty weights, as well as enabling storage, connectivity, and inter-
mediate stability specifications to be altered to force a given plant struc-
ture to have dominant features attributable to the problem classes
Downloaded By: [University of Waterloo] At: 18:30 10 October 2007

780 A. ELKAMEL A N D A. MOHINDRA

defined. Changing the scheduling horizon and setting demand levels


within the language context is also trivial. This makes the language
ideally suitable to a decomposition strategy since the length of time
windows is easily modifiable. The reactive scheduling problem can be
viewed in two distinct stages: a pre processing stage and a maintenance
stage. During preprocessing, information about the original schedule is
determined followed by collection of information regarding the priori-
tization of orders, and calculations to make n priori reassignment
penalty cost assessments. During the maintenance stage the schedule is
regularly updated according to need.
Once solution generation has been completed, a cost of assessment
is made by checking to see how many penalty costs have been
incurred. This is done by comparing the a priori determined M I L P
variable values for the original schedule with the resulting variable
values of the reactive solution. Based on this information, a table is
generated containing the number of time shifts, number of batch size
changes, and number of unit replacements.

ILLUSTRATIVE CASE STUDY


A N D SENSITIVITY ANALYSIS

The purpose of this section is to demonstrate the ability of the rolling


horizon heuristic to solve reactive scheduling problems involving vari-
ous, at times simultaneous, conflicts. A model plant structure con-
taining features not previously considered in the reactive scheduling
literature serves as the basis of the study.

Test Problem
The state task network representation for this plant appears in Fig-
ure 3, and details of the plant tasks and equipment structure are outlines
in Table I. A large horizon length is chosen in order to demonstrate the
rolling horizon methodology. Note fromTable I that the processing time
is a function of the unit on which the task is performed. Each of the three
mixing tasks, three reaction tasks, and two drying tasks are allowed to
have two dimerent processing times. This feature adds to the
combinatorial complexity of the scheduling problem.
Downloaded By: [University of Waterloo] At: 18:30 10 October 2007

REACTIVE BATCH SCHEDULING 78 1

FIGURE 3 State task network for test problem.

As the STN shows, there are 8 processing tasks, 3 cleaning tasks, 3


feeds, 2 intermediates, one of which is unstable, and 3 shared secon-
dary resources. The secondary resources are specified to have cyclical
availability profiles. From the STN it can also be noted that there are
two products produced jointly in one interacting production line. The
relevant information about resources appears in Table 11, and storage
vessel data is presented in Table 111. Table IV lists the demand profile
used in this study. A Gantt chart of the original optimal schedule for
this demand pattern is illustrated in Figure 4.
In order to generate disturbances for testing, it was necessary to
devise a generator that could randomly select from a large set of
possibilities. The generator, using a formula that mathematically
computes random numbers based on a random initial seed, was
required to make the following determinations: time of disturbance,
type of disturbance, time of notification, and, in the case of unit
unavailabilities and processing time deviations, length of disturbance.
Downloaded By: [University of Waterloo] At: 18:30 10 October 2007

782 A. ELKAMEL A N D A. MOHINDRA

TABLE I Test plant task and equipment information


T ~ s no~ue
k Allowed unit Unit size Processina time
Mixingl MI6 1250 2
(ml) MI^ 750 4
MI8 1000 4
Mixing2 M 16 1250 2
(m2) MI7 750 4
MI8 1000 4
Mixing3 MI6 1250 2
(m3) MI^ 750 4
MI8 1000 4
React l RI 2000 6
(Rxl) R2 2500 6
R3 1500 4
React2 RI 2000 6
(Rx2) R2 2500 6
R3 1500 4
Rencl3 RI 2000 6
(Rx3) R2 2500 6
R3 1500 4
Dryingl D25 750 3
(dl) D26 1000 2
Drying2 D25 750 3
(d2) D26 1000 2

TABLE I1 Resource information for test case


Resource Type Stability Initial Purchasable Cycle rime . Value
C NRR Unlimited 10000 No - 2.5
D NRR Unlimited 10000 Yes - 0.5
E NRR Unlimited 10000 No - 10
A BC NRR Unlimited 0 No - 2
X NRR Zero Wait 0 No - 0
Y NRR Unlimited 0 No - 586
Z NRR Unlimited 0 No - 925
Elec RR Unlimited 1000 Yes 15 0.145
Stenm RR Unlimited 1000 Yes 15 3.5
Water RR Unlimited 1000 Yes 20 0.08
NRK
RR --Nonrenewable Resource.
Renewable Resource.

The possible disturbances considered in this study were unit un-


availabilities, processing time deviations, new orders, order cancella-
tions, order priority changes, and deadline advances. Table V list the
randomly generated disturbances that were introduced to the schedule.
Downloaded By: [University of Waterloo] At: 18:30 10 October 2007

TABLE 111 Storage information for test case


Vessel State Capacity
MI6 C 1250
D 1250
E
ABC
C
D 750
E 750
ABC 750
MI8 C 1000
D 1000
E 1000
ABC 1000
D25 Y 750
Z 750
D26 Y 1000
z 1000
Vessel C C 10000
Vessel D D 10000
Vessel E E 10000
Warehouse Y Y 10000
Warehouse Z Z 10000

TABLE IV Original demand profile for test case


Producl Due dale Amounr demanded
Downloaded By: [University of Waterloo] At: 18:30 10 October 2007

784 A. ELKAMEL AND A. MOHINDRA

FIGURE 4 Original schedule for test case

TABLE V Disturbance details

3 3 Order Cancellation Third order for Order size: 300


product Z
16 New order 700 of product Y Due dole 16
33 Due Date Advance Thirteenth order New due dale 33
for product Z
29 29 Processing time Drying task on Duration of 5 extra
deviation unit D25 time periods
29 29 Unit Unavnilabilitv Unit R3 From time 29 to 32

Note that disturbances that d o not have the same notification time are
not considered simultaneously.
At time 3, the first two disturbances are reported. The reactive
scheduling algorithm goes into effect and generates a new schedule to
accommodate the disruptions. The subhorizon length selection yielded
a time window of 23 time periods, which was set to be the flexible
region, subject to reassignment penalty costs. The remainder of the
schedule was solved without penalty costs in a time window of length
22. The reactive scheduling solution then becomes the operating
schedule for all times after time period 3. At time period 9, the next
disturbance is noticed and the heuristic is used again with the solution
Downloaded By: [University of Waterloo] At: 18:30 10 October 2007

REACTIVE BATCH SCHEDULING 785

from the first disturbances as its basis. The reactive scheduling for this
instance was done with two time windows of length 23, and 17, re-
spectively. Once again, the solution that is found becomes the plant
schedule and a final iteration of the procedure is simulated for a
simultaneous conflict of processing time deviation and unit unavail-
ability at time 29, which was solvable in one time window of length 22.
Figures 5 through 7 show the scheduling solutions obtained by the
heuristic procedure. Storage vessels are not represented in the pre-
ceding Gantt charts. Storage in processing vessels is illustrated by
shaded regions with the name of the stored state. Plain shaded regions
represent cleaning tasks. The tags on the processing tasks are the name
of the task that takes place and the corresponding batch size.
Note from Figure 5 that the new order disturbance is essentially
accommodated by executing some additional processing tasks after the
end of the original horizon. This can be attributed to the fact that the
model plant was already operating at bottleneck capacity and could
not acconlmodate a new order by increasing some intermediary lot
sizes. The second disturbance essentially introduces an urgency con-
dition on the production of a particular order. This immediately has
the effect of causing the solver to execute processing tasks as quickly as
possible to meet the new deadline. Note the increased use of storage

FIGURE 5 Solution after disturbance I .


Downloaded By: [University of Waterloo] At: 18:30 10 October 2007

786 A. ELKAMEL A N D A. M O H l N D R A

for intermediate product ABC in Figure 6 . This is an indication that


mixing tasks are being done faster than the bottlenecked reactors can
accommodate.
Figure 7 demonstrates the resolution of two simultaneous conflicts.
Reactor R3 becomes unavailable from time 29 to 32 and the pre-
empted task is subsequently restarted. The dark shading indicates the
unavailability. Simultaneously, the half shaded box on unit D25's time-
line indicates a processing time deviation. Observe that since these
changes happened so close to the end of the preplanned horizon, rela-
tively larger lot sizes and timing adjustments were made in order to
minimize late deliveries. Tables VI through VllI list the schedule
changes that occurred a t each reactive scheduling point. In these tables
a reassignment is defined as a task-unit-time assignment in the new
schedule. Similarly, a new assignment is taken to be one that cannot be
matched to any assignment in the original schedule. The net penalty
cost is the sum cost of schedule modifications plus the cost of due data
violations. Unit replacements did not take place in these cases because
of unit replacement penalties that were larger than time shift penalties.
This serves as an indicator of the types of trade-off issues that arise in
reactive scheduling, and of the sensitivity of the penalty function model
to parameter selection.
Downloaded By: [University of Waterloo] At: 18:30 10 October 2007

REACTIVE BATCH SCHEDULING

FIGURE 7 Solution after disturbance 3

TABLE VI Schedule modifications made after the first disturbance set (Net Penalty
Cost = 2.9899e+06)
Task Unir Start rime Amouni Time shift Barcll size chanpe

Reassigned
Reassigned
Reassigned
New task
New task
New task
New task
New task
New task
Downloaded By: [University of Waterloo] At: 18:30 10 October 2007

TABLE Vll Schedule modifications made after the second disturbance set (Net
Penalty Cost = 3.10459e+09)
Tnsk Unir Slnrt r i m Anrounr Time shilr Barcl~size chanee
Rxl '
Rx2'
ml'
m3
Rx3'
dl'
ml'
1112'
dl
d2
Rx l
Rx3
m2
dl
Rx3
d2 -400
m3' Reassigned
m3 Reassigned
Rx2 Reassigned
dl Reassigned
dl Reassigned
rnl New task
Rx l New task
d2 New task
d2 New task
d2 New task

TABLE Vlll Schedule modifications made after the third disturbance set (Net Penalty
Cost = 2.5133e+09)
T d Unir Srart rin~c Amount Time slrfi Barch size cl~nnge
-

dl D25 36 200 2
dl' D25 39 750 8
d2' D26 31 900 - 450
Rx I' R1 33 1875 I 875
Rx3' R3 32 1000 5 500
Rx2' R2 33 1875 - -625
m3' MI8 34 1000 I 625
m2' MI7 38 750 2
Rx3' R3 38 I500 6
Rxl' RI 41 2000 2 125
Rx3 R3 43 750 6 -750
d2 D26 36 300 - Re;~ssigned
d2 D26 39 500 - Reassigned
d2 D26 41 300 - Reassigned
d2 D26 45 375 - Reassigned
1111 MI6 31 1250 - New task
dl D25 42 300 - New task
dl D25 47 550 - New task
Downloaded By: [University of Waterloo] At: 18:30 10 October 2007

REACTIVE BATCH S C H E D U L I N G

Sensitivity Study
The purpose of this section is to discuss the effect of various model
parameters on the solutions obtained by the suggested rolling horizon
technique. The parameters that require testing are the subhorizon
length and the penalty parameters. In order to check the sensitivity of
solutions to the selection of each subhorizon cut, tests were performed
for the first disturbance set with the minimum cut taken to be the one
used to obtain the actual solution. Several runs were done, each time
increasing the subhorizon length of the first subproblem. It was found
that the length of the time window selected affects the outcome of the
reactive schedule modification. The main trend is that as the sub-
horizon length increases, the number and magnitude of task-time-unit
shifts decreases. It is also of note that complete task-unit-time reassign-
ments d o not take place as often in the penalized region. Rather, the
solver opts to make larger modifications in the future time windows
with the lowest penalty costs.
The sensitivity with respect to penalty parameters is checked by
definingschedulingproblem classes in order to distinguish those schedul-
ing problems that will have certain "dominating" features that should
make them insensitive to certain penalty parameter variations during
reactive scheduling. These are: the storage, the batch size, the resource
constrained the assignment constrained, and due-date dominated
classes. The storage dominated class represents schedules where storage
capacity is available and nearly all materials in the plant are stable over
nearly the entire scheduling horizon. The batch size dominated class
refers to right schedules where batch sizes are close to unit or bottleneck
capacities and/or fixed due to constraints on the merging and splitting of
batches. The resource constrained class represents schedules where
additional outside raw material purchasing is not allowed once
execution of the plan is underway. The assignment constrained class
involvesschedulesin which unit replacementsare not allowed and/or are
extremely costly. Finally, the due-date dominated class corresponds to
scheduling problems where the costs of missing order deadlines are
extremely high compared to other costs in the objective function.
The portion of the schedule after the third disturbance was selected
for sensitivity analysis. This portion is different from other points in
the schedule because it lies a t the edge of the flexible region and the
Downloaded By: [University of Waterloo] At: 18:30 10 October 2007

790 A. ELKAMEL AND A. MOHINDRA

beginning of the free region. The scheduling subproblem at this point


was solved repeatedly while varying each of the penalty parameters
independently. Plant parameters were also modified and the solution
runs repeated for cases of each problem class defined earlier. The main
results are summarized below. For each problem class, solutions were
found first with mixed penalties. Then each individual parameter was
tested at a high value and a low value corresponding to the priority
bounds on the given parameter.
From the numerical trials for the assignment constrained class, no
sensitivity was exhibited to the penalty terms for batch size preser-
vation. Storage control parameters also had no impact on the problem
class. The only parameter that exhibited noticeable sensitivity was a:.
Increasing the unit replacement penalties only had the effect of
increasing the objective cost.
Results for the batch size dominated class exhibited greatest sensi-
tivity to timeshifting penalties. Unitreplacement and batchsize penalties
did not show any impact upon adjustment. Fixed changeover costs that
were not part of the penalty terms may have been the reason why unit
replacements did not occur. Increased storage penalties only resulted in
increased computation time but not in a different solution outcome.
Time shifting, batch size, and storage penalty terms all exhibited
significant sensitivity for the due data dominated class. Unit replace-
ment penalties again were not a cause of sensitivity. This strongly
sensitive class confirms that trade-off considerations are significant in
the objective function. Delivery deadlines represented by objective
weighted terms have a strong impact on the amount of reactive
scheduling that may take place.
For the resource constrained class, only unit replacement terms did
not demonstrate a great deal of sensitivity. All other terms had clear
effects. This is of note because it correlates with the hypothesis that
there is a strong relation between resource specifications and problem
difficulty. When such a high degree of sensitivity exists, even slight
perturbations to the objective function coefficient could lead to
exponential increases in problem difficulty.
As was expected, all parameters for the storage dominated class
exhibited a great deal of sensitivity with regard to solution control.
Because of the enhanced assignment flexibility created by the addition
of more storage space in the plant, the number of feasible assignments
Downloaded By: [University of Waterloo] At: 18:30 10 October 2007

REACTIVE BATCH SCHEDULING 79 1

grows dramatically and the solver is able to search many reassignment


combinations.

CONCLUSION

In this paper, a mathematical programming based heuristic which uses a


rolling horizon planning strategy to solve reactive scheduling problems
has been presented. The consideration of user preference and relative
customer importance has led to a penalty function approach. Weights
on certain assignment variables are incorporated into the objective func-
tion of an MILP formulation with the aim of both preserving and
modifying the original schedule in a near optimal fashion. The main
shortcoming of this approach has been the larger and varied number
and types of user specified, subjective parameters, in addition to a
number of model induced parameters, such as subhorizon length. In
order to characterize the sensitivity of the model to these parameters,
initial studies of some of the parameters have been done on a test case.
Visible response to subhorizon length variation has been observed and it
appears that by lookingat a longer time horizon, an improved quality of
reactive modifications can be made, especially when faced with multiple,
simultaneous disruptions. Indications from the results are that
parameter sensitivity in the model is a function of plant characteristics.
DiKerent penalty weights appear to have a stronger impact when the
plant has certain dominant characteristics.

References

[I] Reklaitis, G . V. (1992). Overview of scheduling and planning of batch process


operations. NATO AS/, Borclr Processing Systems Engineering, Anlalya, Turkey.
[2] Zentner, M. G . and Reklaitis, G . V. (1992). An interval-based mathematical model
for the scheduling of resource-constrained batch chemical processes. NATO AS/,
Barck Processing Sysrems Engineering, Antnlya, Turkey.
[3] Zentner, M . G., Pekny, J. F., Raklaitis, G . V. and Gupta. J. N. D. (1994). Practical
considerations is using model based optimization for the scheduling and planning
of batch/semicontinuous processes. Process C o ~ ~ r r o4(4),
l . 259-280.
[4] Pekny, J. F. and Miller, D. L. (1991). Exact solution of the no-wait flowshop
scheduling problem with a comparison to heuristic methods. Compurers orrd
Clremicul Engineering, lS(1 I), 741 -748.
151 Kudva, G., Elkamel, A., Pekny, J. F. and Reklaitis, G. V. (1994). A heuristic
algorithm for scheduling batch and semi-continuous plants with production
deadlines, intermediate storage limitations, and equipment changeover costs.
Compr. C / ~ o ? rEng.,
. 18(9), 859-875.
Downloaded By: [University of Waterloo] At: 18:30 10 October 2007

792 A. ELKAMEL AND A. MOHINDRA

[6] Musier. R. Fitt and Evans, L. B. (1989). An approximate method ior the
production scheduling of industrial batch processes with parallel units. Conipt.
Chmrl. Big., 13, 229-238.
171 Pekny, J . F., Miller, D. L. and McRae, G. 5. (1990). An exact parallel algorithm for
scheduling when production costs depend on consecutive system states. Compr.
Cl~err~.Erig., 14, 1009- 1023.
[8] Wellons, M. C. and Reklaitis, G. V. (1989). Optimal schedulegeneration fora single-
product production, I-problem formulation. Compr. Chern. Dig., 13,201 -212.
[9] Kondili, E., Pantelides, C. C. and Sargent, R. W. H. (1988). A general algorithm
for scheduling batch operations. Proceedings, Third Inreniutionul Syriipo.~ilrn~ or!
Proccss Sj~sre~iis Erigirieering, Sydney. Australia, pp. 62-75.
[lo] Elkamel, A., Zentner, M., Pekny, J. F. and Reklaitis, G. V. (1997). A de-
composition heuristic for scheduling the general batch chemical plant, Erig. Opr.,
28, 299-330.
[I I] Elkmnel. A, and Al-Enezi, G. (1998). Structured valid inequalities and separation
in optimal scheduling of the resource-constrained batch chemical plant. Murh.
Engrifi I d . . h(4), 29 1-3 18.
[I21 Nof, S. Y., Rajan, V. N. and Frederick, S. W. (1990) Knowledge-based dynamic
real-time scheduling and rescheduling: A review and some annotated references.
Researcli Me~irororrrl~rrr~ No. 89-16, School of Industrial Engineering, Purdue
University, West Lahyelte, IN.
1131 Fox, M. S. and Smith, S. F. (1984). ISIS-A knowledge-based system for Factory
scheduling. E.vperr Sysrnris, I(]), 25-49.
[I41 Ow, P. S., Smith. S. F. and Thiriez, A. (1988). Reactive plan revision. Proc. Se~wirli
Nat7 Cortf. A / , MIT Press, Cambridge, Mass., pp. 77-82.
[I51 Cott. B. J. and Macchietto, S. (1989). Strategies for operating batch plants subject
to variability - 21 performance assessment. C l i o i ~ Erig.
. Res. Des., 67(6), 593-599.
1161 Kanakamedala, K. B., Reklaitis, G . V. and Venkata Subramanian, V. (1994).
Reactive scheduling modification in multipurpose batch chemical plants. lr~rlusrrial
ar~dEr~girleerhlgClremistry Research, 33(1), 77-90.
[I71 Elkamel, A,, Zentner, M. G., Pekny, J. F. and Reklaitis, G. V. (1992). An enhanced
uniform discritimtion model for the batchlsemi-continuous chen~ical plant
scheduling problem. C l P A C Report. School of Chemical Engineering, Purdue
University, West Lafayette, IN 47907. USA.
[I X] Nemhnuser, G. L. and Wolsey, L. A. (1988). Integer arrdComhinarori~~lOp~in~izatio~~.
Wiley, New York.
[I91 Zentner. M. G . , Elkamel, A,, Pekny, J . F. and Reklaitis. G. V. (1997). A language
for describing process scheduling problems. Conip. Chenl. E I I ~ .22(1-2),
, 125- 145.

También podría gustarte