Copyright © IFAC Dynamics and Control of Process Systems,

Corfu, Greece, 1998

BA TCH PROCESS ENGINEERING REVISITED: ADDING NEW

SPICE TO OLD RECIPES

Sandro Macchietto

Process Systems Elllerprise Ltd. - London W6 9DA - UK. [email protected]

Abstract: The paper presents a personal overview of the recent advances in one of the oldest
processing technologies. Particular attention is given to developments since the landmark 1992
NATO ASI on Batch Process Engineering. with specific focus on plant-wide aspects and
integration. New techniques are noted for modelling. simulating and optimising batch processes,
for the design and validation of automation systems. and integration of batch operations
planning and execution . Some of these technologies are already helping leading companies to
dramatically improve the performance of batch manufacturing operations in existing plants, the
return on assets for new plants , and the responsive operation for make-to-order mode . The
significant economic impact of these technologies is noted with reference to a number of
industrial case studies. It is concluded that a radical revolution is now possible in Batch Process
Engineering. Copyright © 1998 IFAC

Keywords : Batch Processing, Simulation, Scheduling , Dynamic Optimisation , Integrated

Operations. procedural control.

I. INTRODUCTION
Twenty years ago there were just a few individuals with little) sense as one of continuous process engineenng.
research interest In batch processing , which was Accordingly, in the enclosed some brief and highly
considered largely the domain of practitioners. The selective personal views are presented of just some
difficulty associated with non linear, time-varying development in a few key areas. The paper covers in some
dynamics , tight coupling of design and operations and more detail some of the less heavily beaten paths.
discontinuous events meant that most problems of interest particularly discrete control automation and integrated
were just too difficult. Starting in the mid seventies and operation of batch processes.
early eighties some interest in this area grew. stimulated
by the move in the chemical industry towards speciality
2. MODELLING AND SIMULATION
chemicals. A NATO ASI meeting in 1992 brought
together (they would fit on the sides of a small pool) a Successful results on a variety of demanding applications
good part of the research community from academia and show that the modelling and simulation of hybrid systems
industry involved in batch processes for an assessment of and complex operating procedures. capturing the detailed
the field (Reklaitis et al.. 1996). A few years hence. is physics and dynamics of batch processes, have come of
clear from even a cursory examination of literature that age. A new generation of general-purpose software tools
there has been a very rapid growth in the field. with such as gPROMS (1998) and its research derivative form
significant development in fundamental methods and MIT have emerged, which specifically incorporate the
applications in most directions . I feel a general review of needs of batch processes. These tools result from a solid
batch process engineering makes now as much (indeed, as fundamental analysis of the engineering problem domain

535
and the design of an appropriate modelling framework. already enables this to some extent (see above) , but
They have placed strong emphasis on the development of significant more progress in inter-operability of software
suitably robust algorithms and numerical solution engines. and specially applications is required . Slow but significant
They also deal with the required complexity through a progress in this area is being made in industry -led
strong object orientation in the modelling language and activities such as the European CAPE-Open project and its
software architecture design , using techniques such as announced world-wide follow-up .
hierarchical model decomposition , incapsulation. etc.
The main progress has been in the new ways of
3. CONTROL AND DYNAMIC OPTIMISATION
modelling complex procedural aspects (tasks) in addition
to physical systems with discontinuities. The present In "traditional" batch control applications (control of
ability to deal easily with distributed systems (partial single operations such as reaction and distillation), much
differential equations) and the corresponding boundary progress has been made in addressing not just regulatory
conditions , and population balances mean s that a whole control and process performance optimisation, but also
range of key processes , from crystallisation to issues such as safety , reliable operation over a range of
chromatographic separations and tubular reactors, can now process and disturbance conditions, and design-control
be finally incorporated within process models with some integration. Excellent recent reviews are available (Berber,
ease. This has been enabled by the considerable progress 1995. Bonvin, 1997, Abel et aI., 1997) and the interested
in solution techniques capable of dealing with large-scale reader is referred to them. From my perspective, the more
systems of differential and algebraic equations. Such interesting developments seem to all come from model-
modelling and simulation tasks can now be routinely based applications of optimisation with mechanistic
performed . Other interesting work aims at easing further models (most batch reaction , distillation. etc. are usually
the model development of general process (e.g. MODEL- significantly nonlinear) . On one hand these models can
LA, Stephanopoulos et al. 1990 and MODKIT, Bogush et represent process behaviour over a wide range with some
aI., 1996) or is specially tailored for recipe-oriented confidence. On the other, the optimisation will take full
process descriptions such as Batches (from Purdue advantage of this complex behaviour to achieve extra
University) and BASiP (Engell et aI. , 1998). Interestingly, performance , flexibility , robustness, etc. as required .
all have emerged from the academic community. The In this optimisation framework many "classical"
traditional vendors have recognised the opportunity and a problems can now be more elegantly and broadly
number of commercial offerings are at various stages of reformulated and solved. For example , the problem of
development (e .g. BDK from Hyprotech and Batch Plus input-output controllability (the ability to achieve
from ASPEN Tech.) although both of these appear to desirable control performance for a range of disturbances)
target a very much narrower problem domain. As with in a batch reactor may be quite naturally restated as
high performance cars, some start from a robust and finding the best input controls and trajectories (typically
reliable engine and then worry about aerodynamics and within bounds) that keep the trajectories and final values
looks, others start from the latter and then strive for the of outputs (and states , if required) within given bounds
required horsepower and reliability. An interesting around a reference trajectory, for given (bounded)
observation is that MODE-LA, MODKIT and BASiP all disturbances . " Best" here may be defined in any suitable
utilise gPROMS as a solution engine. way, for example taking into account not just process
In the wider modelling area, there are two further metrics (e.g . integral of the error) but directly economics
trends that will no doubt have a significant impact on the (e .g. value of quality losses) . In fact, given a good model
analysis of batch processes. The first stems from the need of the process. it is not even necessary to specify a priory
to consider models at a rather more detailed level that has reference output trajectories. as these will be a direct result
traditionally been done in batch simulators. For example , of the optimisation . Control settings may similarly become
understanding and simulating detailed miXIng and optimisation parameters. as long as the controller itself is
hydrodynamic effects is essential for scaling up a design included in the model. Similarly , it is possible to
and quality control of many batch processes. This has incorporate uncertainties in the model parameters and
traditionally been dealt with outside and separately from multiple operation scenaria (e .g. processing of two
overall simulation. for example using CFD tools. The mixtures in the same batch column).
ability to merge heterogeneous models, applications and A number of applications have shown the power of
calculations engines within and open "Object" framework such approach (e.g. Louleh . et aI. , 1996, Mohideen et aI.,
(Kakhu et al. 1988) will no doubt enable the integration of 1996 and Bansal et al. 1998 , albeit for a continuous
such more detailed effects into design and control work. process). One of the interesting points is the increased
The second trend is a move towards so called "model confidence obtained in scaling up from small , laboratory
servers" (Pantelides and Britt, 1994), whereby the experiments to full industrial size, as in the safety study by
modelling capability and solution engines of various Toulouse et al. (1996).
(typically heterogeneous) application packages can be Such optimisation approaches typically rely on the
tapped on demand from other applications and users, no parameterisation of some control or disturbance profiles to
doubt over the web. Progress in interfacing technology transform an otherwise infinite dimensional optimisation
into a tractable one . Of course these parameterisations ii) Interesting applications have been presented where
will have to be chosen wisely to produce meaningful a small number of integer decision variables are also
results. included (for example. to represent sequencing or
In general. dynamic optimisation is now strongly allocation choices, or measured-controlled variable
emerging as a central methodological framework for pairings). Algorithmically, this involve solving a MINLP
detailed control and design of hybrid/continuous batch master problem (this was simply an NLP in the previous
systems . Indeed, it has been argued that a whole range of case) with embedded an integration of the DAE equations.
problems arising In the batch area, from control The resulting mixed-integer dynamic optimisation
procedures generation to safety verification and problems are still difficult to solve and there is some good
simultaneous optimisation of recipe structure. operating theoretical and numerical work to be done. but the results
conditions and schedule are just dynamic optimisation from the few initial applications indicate returns in terms
problems (Barton et al.. 1998). These authors present a of process improvements and economics can be quite
very nice and lucid analysis of the particular problems that impressive (e.g . Mohideen et al. 1997).
arise when extending optimisation from purely continuous iii) Finally, some problems may be formulated
systems to those with discontinuous behaviour. Three involving a large number of integer decision variables in
situations may be distinguished . addition to continuous ones. For this case the situation is
i) Problem formulations in the form of a mulTistage not so rosy. in the sense that existing theory and methods
opTimal control problem have now been presented for a appear to be still inadequate. and the more promising
variety of useful applications. These result in an extension avenues on the short term should include re-formulations
of the continuous dynamic optimisation problem and (i.e. to avoid the problem in the first place) or hierarchical
sufficiently general, robust and efficient algorithms based decomposition and/or partitioning techniques. For
on control vector parameterisation have recently been example, Georgiadis et a!. (1998a,b) developed
developed. Indeed, the first commercial solution is already simulation models for tubular and plate heat exchangers
available and extensively used in gPROMS. subject to severe fouling. The model. incorporating a
Computational requirements are not excessive. and detailed reaction mechanism of deposition on the walls,
applications with detailed dynamic models (several transport phoenomena and hydrodynamic profiles,
thousands of state variables) and tens of decision variables reproduces very well experimental values of the mass
can be solved in a few hours in a standard workstation. deposits and exit temperature profiles, for a variety of
One of the main advantages of the approach is that it complex co- and counter-current configurations and feed
applies to generic sets of DAE equations, and therefore is mixtures. In the shell and tube case, the detailed PDE
not limited to individual unit operations. For example. model with fouling could be used directly to obtain control
Charalambides (1996) showed that a batch reactor, the policies (as well as a couple of key exchanger design
subsequent batch distillation and all off-cut recycling can parameters) that optimise the operation-cleaning cycle.
be optimised as an integrated whole. The optimal results taking into effect the substantial economic impact of shut
balance the reaction (longer time leads to higher down periods (Georgiadis et a!. 1998c).
conversion and hence easier separation. but also higher If we wish to find the optimal operation-cleaning
by-products) with the separation (increased impurities schedule for a whole network of such slowly fouling
increase loads on separation and affect the amounts of off- exchangers using detailed dynamic models. the problem
cuts recycled to the reactor). Detailed profiles of reactor becomes computationally intractable. However, the
and column controls (time-varying temperature and reflux optimal time-decaying profile of each exchanger may be
ratios . respectively). the duration of all operation steps. approximated by a suitable parameterisation. the problem
and amounts to be recycled were all optimally calculated. cleverly reduced to a MILP formulation and solved quite
The economic impact of such techniques is very large. In a efficiently even for rather large networks (Georgiadis et al.
significant industrial application, described by lshikawa et 1998d) . The optimal operation-cleaning schedules thus
al. (1997) the optimal operation of a complex batch found reflect the rate of fouling. the trade-off between
reactor-separation system. already well engineered over longer operation and longer cleaning times, and the
many years. lead to a doubling of profit margins in an possibility of manipulating the heating medium within its
existing plant and excellent insights for a new design. control range.
Prior to dynamic optimisation, a high fidelity model was Together with the recent work of Jain and Grossman
developed with the use of some additional experiments for (1998) for continuous processes with time-decaying
the kinetics and fully validated against laboratory and performance. this shows that for special cases it is now
plant data. becoming possible to integrate quite detailed process
Future developments in this area need to deal with dynamics and control with scheduling. when there is
some remaining theoretical problems to do with "high strong coupling between the two. There is clearly the need
index" DAE systems, discrete switching and path to generalise these initial results into more general-purpose
constraints, although the latter do not really appear to techniques . Applications awaiting include all kinds of
create many problems in practice. decay-regeneration cycles and periodic maintenance, in
addition to strictly batch processes.

537
 Studio (PRESTO) in the Centre for Process Sy stems
Engineering at Imperial College . The economic analysis
4. DESIGN OF SEQUENTIAL CONTROLLERS
was carried out by a PRESTO team member with a PhD in
As noted in many occasions. 80% of batch control IS management. The supplier company has automation and
sequential. and 80-90iJc of that is to handle abnormal control engineers in several countries and was interested in
eventualities (possible failures. emergencies. etc.). The standardising and where possible , automate . its procedures
consequence of errors and omissions, whether in the initial for design . commissioning and lifetime maintenance of
control logic design or implementation into control code, sequential control code . A representative plant (Plant X)
range from simple delays. to costly rework, delays in was selected which had already been proposed internally
commissioning and occasional catastrophic failure. as for a benchmarking exercise. Following the company" s
amply documented in the literature. The design and standard internal procedures, a team of engineers
validation of sequential control systems is therefore an developed the process specifications, from which the
area of great practical and theoretical interest. detailed control sequence and logic design were developed
This area has only recently captured the attention of and finally implemented into a complete set of control
the chemical engineering community, however there is code . The work of the team for each of the activities
now some serious work ongoing, utilising a variety of involved (specification , implementation, testing , etc.) was
formalisms ranging from Petri Nets (PN) to Finite State closely observed. The results of each development step
Machine (FSM) and Temporal Logic (see review by was analysed in detail to establish the number and type of
Engell et al. .1997). errors introduced and detected (or not) at each step , the
The approach we have followed over the past few amount of rework and corrections required , and the
years utilises FSMs to model the purely discrete states of a number and type of errors that would have filtered
system, and a special form of Temporal Logic to express undetected to the final control code delivered for
in unambiguous terms the specifications posed on the commissioning. A cost model of these activities at the
sy stem. The method permits checking whether the company was also developed. Similar data were collected
specifications are consistent with the model and amongst for a second. live project in the company , and from two
themselves, and to constructively develop the logic other control integrators, which permitted a broad
required (a procedural controller) so that the controlled validation of the economic model for two situations (new
system behaves according the specifications. A procedural and retrofit projects, respectively). In parallel. a team was
controller will have a number of interesting and quite put in place that developed control logic and code, starting
strong properties , for example the guarantee that no from the same engineering specifications but utilising the
uncontrolled events will occur which will take the system formal PCT analysis and synthesis tools. An estimate was
outside of a desired state set. This is analogous to a then made of the work and cost required if the formal
stability property in continuous systems. Procedural procedures were implemented within the organisation in a
controllers may be automatically translated into the syntax production mode . From the cost, rework etc . data from the
of a target control system language. other two automation suppliers (which were largely
The overall approach , called Procedural Control similar) an estimate was developed for the potential impact
Theory (PCT) is described by Sanchez et aI. , (1998a) and in the industry as a whole.
Sanchez (1996). with details in a series of related papers. The details of the analysis are covered by
The main theoretical problem associated with this confidentiality but the broad conclusions are available .
type of method s has been the so called "state explosion" For a new plant, the automation costs were found to be
which results from the combination of even a small around 7-8 % o f the total plant cost. A number of errors in
number of components into a composite system ( a system both specification and implementation were detected when
of 10 two-state components will have 2 '11 possible the usual development procedures were followed, most of
combinations) . This is being addressed by a variety of which were found at some point during development. A
techniques , mainly through hierarchical decomposition few however would have trickled down to the code
into small subsystems. We have been able to extend delivered. All of these errors were either not generated or
theoretical results such as controllability and stability to detected much earlier when the PCT method was used.
hierarchical structures of subsystems in general series- The formal procedure would require rather more work up
parallel combinations. Interestingly. we can therefore front , but would produce some large savings overall.
follow the same structuring as recommended by the ISA These would be particular large in any subsequent retrofit.
S88.01 standard for design of batch control code. A Overall , it was estimated that introducing this
similar state reduction goal is also achieved by separating. technique would save the company 30iJc of its control
as far as possible , control for abnormal and emergency software development cost, equivalent to 10ge of its
operations from control for the normal case. turnover and 240 engineers. The reduced cost base would
To assess both the technical and economic potential of have made new markets worth £500M accessible to the
these methods , fairly early on we carried out a company. The total savings in automation supplied to the
collaborative project with an automation supplier and UK process industry was estimated at £200M p .a. Some
several operating companies. within the Process Concept immediate benefit resulted from modifications of the
standard specification sheets to eliminate possible long as the solution remains firmly model -based and
ambiguities. configuration and data driven. The Integrated Batch
As part of the work, we also developed a full dynamic Operations section later in this paper deals more in detail
model in gPROMS of the uncontrolled Plant X process (a with this second approach .
small milk pasteurisation plant). The control logic
produced for each control sequence using PCT was in the
6. OFF-LINE PLANNING AND SCHEDULING
form of a Finite State Machine representation. This was
automatically translated into gPROMS control tasks that Considerable progress has been made In the
could drive the simulation according to the required logic development of off-line scheduling methods for planning
(details in Sanchez et aI. , 1998b). This provided a way to and scheduling batch operation. A thorough and up-to-
demonstrate the correct working of the procedural control date review of this area is given by Shah (1998) . Advances
sequences thus developed. range from much more general formulations , which enable
Finally , based on our experience of the "control capturing a wide range of applications and constraints, to
design process", we have started putting together an much improved solution techniques. An important
engineering environment to help regular engineers underlying contribution has been the rapid and wide
(without a PhD in formal methods) design such provably acceptance of the elegant State Task Network (STN)
correct sequences and code (Marikar et al.. 1998). representation , initially introduced by Kondili et al. (1988)
Indications are that it should be possible to achieve this as the way to describe batch recipes and ancillary
goal. For example. nineteen engineers in a MSc course , processes , and associated mathematical notations and
with no previous experience of sequential control , were MILP formulations. Later generalised by Pantelides
able to put together, after a short training and in just a few (1994) into the so called Resource Task Network (RTN)
days. a good part of the control sequences for our batch and expanded by many others, these have provided a
pilot plant. much needed common language to the process scheduling
community. Another happy development, largely
associated with the above, has been the slow demise of
5. ON-LINE SUPERVISORY CONTROL
assumptions such as No and Infinite Intermediate Storage,
The regulatory control of batch processes (executing a Zero Wait. which have caused so much confusion and
batch) is well catered for by a large number of industrial wasted research (to this date I have yet to find a batch
batch control systems. However, the on-line plant-wide process application where such unlikely idealisations
supervisory control of a batch plant remains a challenge, adequately represent reality). For example, it may be
involving on one hand production scheduling decisions possible to "store" material (a task) between two process
(such as batch sequencing. equipment and other resource "stages" (tasks? vessels?) without an intermediate "store"
allocation - notoriously complex due to unavoidable (definitely a vessel), as long as the upstream or
combinatorial aspects), and on the other real-time downstream equipment can carry out the "store" task . It is
operations management and control. On the medium term exactly in such issues that STN and RTN , by cleanly
horizon, barring a substantial breakthrough in scheduling separating resource networks from recipes and process
technology (through radically improved formulation networks, have enabled a new clarity of discourse .
and/or solution methods) , there appear to remain two A key issue in scheduling is the representation of
practical approaches: i) the development and exploitation, time . Representation as discrete intervals is typically
as far as possible. of specialised algorithms tailored to widely used. This has the two main advantages that i) it is
problem structure or knowledge. ii) some kind of problem possible to describe very wide types of constraints (such as
decompositions , either hierarchical (with aggregation and equipment and pre and post conditions) and ii} that the
disaggregation), temporal (rolling horizon) or resulting models are typically linear (MILP) and hence in
geographical (decentralised). Our experience in this area theory may be solved to global optimality. Although it is
is that neither approach is without pitfalls. It is always possible to tackle and solve usefully complex applications
possible to tailor a specific solution around a single (for example, involving up to a few thousand integer
application. However in batch applications problems decision variables) , the solution of really large-scale
rarely remain the same for long (just another product, problems remains a challenge . It is envisaged that progress
please ... . or equipment, constraint, etc.) and there are here will come from improved formulations (with fewer
simply too many individual problems around. integer variables, tighter relaxations and fewer
Decompositions have often been used to slay complexity degeneracies). improved bounding procedures to eliminate
(divide and conquer) at the cost of new demands for much searching, and use of known structural features
ensuring co-ordination . consistency and optimality. My arising from knowledge of the problem. in pre-processing
personal vIew IS that taking the latter avenue and during the search . Alternative stochastic optimisation
(decomposition) is preferable as long as the techniques methods are also been explored with some success,
used are indeed applicable to a general enough class of although it is not clear at this point whether they can
applications and co-ordination and consistency-checking successfully overcome the combinatorial aspects inherent
methods can be automatically implemented. That is, as in the problem. As a small note , we have been able to

539
generate a suitable encoding for such an evolutionary experience has shown that in all the cases we have been
search method starting from the same engineering problem involved with it has been possible to increase process
representation that is used for generating the MILP manufacturing capacity by at least IO-20'7c through
representation . This will enable us to provide shortly a improved scheduling at the supervisory level alone and in
realistic comparison of the techniques. which is so far still many cases by significantly more.
missing (it is trouble enough to produce so lution s one In spite of the obvious industrial need and Interest.
way. let alone a comparison. and most researchers in this there has heen only limited academic work in this area.
area are religiously in one camp or the other anyway). Some re-scheduling techniques has been presented.
More interestingly. it should be possible to devise and ranging from fairly simple heuristic and searches (Cott and
study hybrid solution method s. Macchietto. 1989. Kanakadamela et a!. 1994) to more
An interesting aspect of stochastic solution methods ambitious ones based on mathematical programming
is that they can deal with model nonlinearities. such as (Rodrigues et al.. 199-1. Schilling. 1997).
those arising, for example . when processing times and No doubt one of the reasons for the few publications
yields depend on batch size. as long as they are properly is that any implementation and experimental work requires
included in the problem encoding. access to rather large amount of infrastructure (a batch
Notwithstanding the progress noted at the end of the plant. its control system. an optimal scheduler. etc.) which
Dynamic Optimisation section. substantial work remains even many companies do not have in (one) place. Another
to be done on incorporating of more complex task models is possibly that for the simple plants that can be easily
into general schedule optimisation methods. described in an academic paper. these technique s are
simply not needed . while getting together the data required
for realistica lly complex cases is not easy. The rece nt
7. INTEGRATED BATCH OPERATIONS
availability of generic simulation tools suitable for batch
The real challenge in batch operations is getting plants (and realistic Virtual Plants. see below) should at
planning, scheduling and control to work together in an least provide useful simulated test beds on the desktop .
integrated framework such as that schematically We have been working for some time on a generic
represented in Figure I . The most important items in this integration framework (e.g. Cott and Macchietto , 1989)
figure are the feedback arrows. As it has been observed and our approach is briefly summarised here . The main
many times. an optimal plan with no or infrequent issues involved and some of the lessons learned are also
feedback and schedule re-calculation cannot meet the highlighted.
challenge of constantly changing conditions. This should
come to no surprise to any control engineer'
8. INTEGRATION OF ON-LINE SCHEDULING AND
CONTROL

With regards to the integration loop between on-line

scheduling/supervisory management and control execution
(sequential and regulatory) in Figure I, we have now
achieved some practical results for quite general batch
models . We have been demonstrating our own batch pilot
plant running in closed loop for a few years now using the
SUPERBATCH software. within the framework of an
advanced operations support system. Some of the
modelling and design issues involved were described in
Macchietto (1997) and the latest update is given in Liu and
.... Macchietto (1998). so reference is made to those papers .
An installation on a full-scale industrial dairy plant (also
briefly described In Macchietto, 1997) has been in
operation for over a year, with a second commercial
Fig I . Function hierarchy of operation problems. installation due to be commissioned shortly.
As far as I am aware. these still remain the only
applications reported that use generic model -based
The economic value to systems and solution providers techniques.
of this integrated process operation market has been
estimated by ARC. one of the most respected analysts in
the sector. at 8.8 billion dollars per annum in '97. doubling 9. INTEGRATION OF SCHEDULE OPTIMISATION
in the next 3 years. Figures for the operating companies AND ON -LINE SCHEDULING
are much harder to obtain. Benson and Perkins (1997)
quote a capacity loss for average performance in the UK We assign to the off-line schedule optimisation box in
process industry of 8'7c on changeovers alone. Our own the hierarchy the function of calculating the optimal

540
number of batches required to meet given demand profiles for whatever reason. The other option is to return to the
for each product over a certain horizon . the optimal batch optimiser with further constraints.
size and the corresponding equipment allocation. Recipes The supervisory system manages the execution of all
are represented in the form of State Tasks Networks the batches posted to it over the scheduling horizon . It
(Shah, 1998). Typically , all major resources and tasks are initiates execution of various tasks as they become due by
included (with approximately one to one match between communicating with the available control system (or
tasks and "operations" in the S88 .01 control sense). Unless simulation). monitors the execution of the various tasks
strictly necessary details of material transfer operations are and revises its schedule as required using the above
typically ignored here. Such an optimiser typically also algorithm . once a minute. until there is nothing left to do.
determines the timing of each batch (start and finish of its The optimal scheduler can be re-run at any time . Of
various tasks) and the use of other resources required course. it must be re-initialised to reflect the current state
within the problem constraints. These results may be of the plant. as represented by the most up-to date
obtained in principle with any scheduling technology. We supervisory schedule. This amounts to a "hot restart" of
utilise gBSS, our own generic, model-base scheduling the optimiser. taking into account any tasks already started
software based on mathematical programming (gBSS , and any subsequent ones which are committed as a
1998). The system can also be used for long term consequence. The re-initialisation requires particular care
campaign planning, planning/scheduling of multi-plants where there exist possible model mismatches . The
and finally to carry out design optimisation. optimisation model will also be changed to incorporate
The integration between the schedule optimisation new demands. resources availability . etc . Whether some
function (carried out by gBSS) and on-line supervisory future but yet uncommitted operation is still to be carried
batch management (carried out by SUPERBATCH) is out as before. or whether it is legal for the optimiser to
addressed as follows . First an optimal schedule is change the corresponding decision variables, will dictate
calculated by gBSS using its own (rather aggregated) which degrees of freedom are available to the optimiser.
model. The values of some of the decision variables thus This choice may likely involve planning/operation people
calculated by the optimiser are retained . These include the (e .g . retain all existing allocations until my shift is
batch size and equipment allocation , and take up the finished) . or be done automatically (e .g. anything not
corresponding degrees of freedom . Other optimisation committed may be changed). Whatever the choice , a new
results are used but only in a different form and some optimisation is solved and the upper loop is closed. The
others are ignored altogether. For example, the relative frequency of such updates will typically be once every few
precedence of the batches is retained (batch sequence , or hours. depending on the rate of change of process and
"partial order" in mathematical term s ) and the timing demands (the update rate is of course limited by the time
information are ignored. The list of batches and their for the optimisation solution).
relative precedence, their size and allocation become the This work is still ongoing but some results for a
specifications for the supervisory scheduler and are passed realistic fine chemicals application from an industrial case
to it. This re-calculates an operationally feasible schedule , study are available (Rickard et al. 1998). The process
using its own , rather more detailed model of the plant and involves 11 products and 9 intermediates on a rather
other constraints, utilising just time as the remaining heavily resource-constrained plant, and some pretty radical
degree of freedom and a set of recursion formulae. The change in product demands some time into production.
starting and finishing time of the various activities for all with a two-week horizon schedule . The initial optimal
batches are calculated starting from the first batch in the schedule is successfully refined by SUPERBATCH and
list so as to meet all operational constraints. then the next executed on-line. until the change of demands. At that
and so on until all have been scheduled. A number of time the new optimal schedule is calculated. reflecting the
feedback loops allow changing some of the earlier current state of production and so on. The on-line system
decisions so as to meet detailed operational constraints. actually runs a simulation of the plant in gPROMS .
Only when this is done successfully. the specification is As there are all kind of opportunities for mismatch
accepted and passed on to the on-line portion of between the optimisation and supervisory models. it is
SUPERBATCH (the monitor) for schedule execution . Of very important that these be carefully chosen so as to
course. since two models at different levels of aggregation allow an easy mapping between them . This is true with
are used at the optimisation and supervisory levels. there is respect to the definition of all resources . tasks and
no theoretical guarantee that the more detailed schedule operations. recipes. etc. Based on experience with several
produced by the latter will still be optimal . or for that attempts we now have a set of guidelines to do so. An
matter. feasible . To this purpose some constraints must be interesting aspect of the current work is that we start from
relaxed at the supervisory level . for example due dates . just a single. composite model of the batch process and
These are however checked on the resulting schedule and procedures. defined in engineering and operational terms.
any violation is reported. There is of course the possibility From this both the optimisation and supervisory model are
for supervisory and operations personnel to intervene and automatically generated in such a way as to ensure
modify the supervisory schedule interactively at any time . consistency is built-in .

541
 offering integrated systems with clearly defined functions
under a variety of acronyms (Enterprise Resource
10 ASSESSING A SCHEDULE
Planning. ERP seems to be the current favourite). In
An interesting integration aspect is the assessment of a addition to accounting and financial tracking functions.
schedule. Off-line planners and schedulers tend to think in these typically include plant maintenance , quality
economic terms (maximum profit). while operations management , inventory and warehouse management,
functions measure performance in terms of equipment logistic and material tracking . One of the main effects of
utilisation , down time , how well they produce the required such top down systems is that they impose the need to
orders. OTIF (on-time. in-full) delivery , etc . formally define the internal "processes" within an
It is rather obvious (but in my experience hardly ever organisation and associated data flows. together with
the practice) that for the planning/scheduling and standard , modular definitions of the various functions.
operations functions to work in the same direction they Hence the data and IT infrastructure environment is far
should at least share the same performance metrics . This is better than it used to be . There are nonetheless several
much more complex that it seems since most economic challenges here . One is the transformation of transaction
performance data needed for optimisation are typically data into real-time data. In relation to scheduling , a
only available much later, in aggregate form (e.g. weekly transactional view of the world requires that a workflow
averages) and for accounting purposes. Our own work in be defined where the order of transactions is defined a
this direction has been to develop , in collaboration with priory and linked to time or other events in a sequential
the Management School at Imperial College, an Activity fashion. On the other hand scheduling typically consists in
Based Cost model suitable for the purpose (Lionis et aI., the simultaneous processing of a great number of
1998). ABC attributes an economic value or cost to each concurrent decisions, the sequence of which will possibly
and all resources utilised by production activities, whether be changed at every schedule update.
materials , personnel. utilities or equipment. For the latter. The other is that the paradigms imposed by such
capital costs are taken into account. We have found that a standard functions not always match well requirements
very good mapping can be established between the batch from the process point of view . Nonetheless a number of
operations models required for scheduling and an activities are under way to interface that commercial
appropriate activity based model. As a result. the detailed scheduling systems with EPR packages. There are also
scheduling information now available from the cross-industry activities aiming at making some order and
supervisory model can be utilised directly , in conjunction possibly standardise definitions, functions and conceptual
with the cost model, to give a whole range of economic models to facilitate integration specifically with process
performance indicators, which reflect directly scheduling scheduling and control. One such effort is the BA TCIME
decisions . Indicators go from the cost, profit and profit project (CEC Project 21215 , http://www.iwi.uni-
margins of individual batches, to value of material losses, sb.delbatcime), supported by the European Batch Forum
etc. Equipment capital costs in the form of assets and the European Union. Within other partners in this
depreciation can be apportioned to various products and project, we have prepared a detailed model of a complete,
individual batches according to their scheduled and actual representative batch plant (from raw materials intake to
use . The cost of unutilised capacity, unscheduled product despatch) to be used as a reference for scheduling
equipment breakdowns, deviations from nominal recipes , and integration exercised. It is intended that this will be
etc . thus becomes immediately visible . While such publicised through the EBF. In the States, Committee
financial performance metrics are of course of interest to SP95 of the Instrument Society of America (ISA) is also
plant management, the opportunity here is that such actively working on producing a standard
measures can be produced directly on-line every time a Enterprise/Control Systems Integration, with the support
schedule is modified or updated . At the supervisory level. of the World Batch Forum (contact www.whf.org).
these metrics may be used to assess and monitor schedule
deterioration due to unplanned events. delays, etc. The
12. VIRTUAL PLANT
selected metrics can also be used to compare on the same
basis schedules produced by the optimisation and Visualisation ' of batch processes is a difricult arca.
supervisory models. As every good control engineer involving disparate sets of time varying data with complex
knows, it will then be possible to use these "economic intenkpendencies. As already noted. the introduction of a
sensors" for closed rather than open loop production Statc Task Network to represent batch recipes and
control. standardisation in batch control nomenclature introduced
by the ISA .88 .0 I standard have already produced major
advantages in communication and interoperahility . The
11. INTEGRA nON OF OPTIMAL AND ON-LINE
traditional Gantt chart representation for schedules also
SCHEDULING WITH ENTERPRISE SYSTEMS
appear to be a somewhat standardised tool or the trade.
This is clearly the next goal and an important one . The animation of simulations is a very effective way to
Here things are facilitated by the rapid emergence over the convey very complex sequential events . Onc of the
last few years of a number of strong commercial players challenges. as always . is that one would like to avoid

542
devl~ing speci;.ti animati()n tool s just i"or thi ~ purpose . plant for ingredient. product and solve nt transfers and
Furthermorc. such 3nimati(ln should he \·iewed in the vari ous reacti ons and separati on tasks inv olving different
hnlader C()l1te .'(t of" ~upp(lning all kind (li" engineering utilities usage and control configurations at variou s times
3ct ivities. from day-to-da) con trol to safety and h3!oj1 (Figure 2) . For vi suali sati on. a mapping was defined
~tudics (typicall y dtll1l' hy a te«m). t() \ isualisatJ(111 or between di sc rete states of the process and co ntrol system
rem(lte operations and trall1ing oi" per~onnel throu gh (already available from the finite state data in the
~Imulat()rs. To addre~~ these needs. in thc past year wc gPROMS mode ls) and the attribute li st of the geometric
ha\c hecn pn1duccd and demonstrated a (()l1ce pt ca lled models of the corresponding equipment. The values of
Virtual Plant. This wa~ done through a collah(lrativc the se attributes change dynamically as the process is
proj e( t invtllvin g thc Centre f(lr Process Systems carried out. So. for example. it is possible to highlight on
Engineering (in particul ar. through its Process Concept th e geometric model the pieces of the plant undergoing
Studio l. Silicon Graphics in the US al1d Swillerland. spec ific activities (e.g. pipes involved in a transfer
Cadcentre in Camhridge. UK and () ur~ehes at PruCl'SS operation) as events occur. Similarly. continuous variables
Systems Enterpri~e Ltd . such as temperatures. pressures and concentrations were
In a nutshell. Virtual Plant brings together advanced mapped to sec tions of the outer shell of the corrcsponding
3D geometric models and dynamic proce ss simulation equipment. so that we could immediately "see" the
model s of plants in a virtual reality environment and over concentration of a reactant as the batch re ac ti o n proceeds.
the Web . Users may then "walk through " the virtual plant. thc concentration fronts of the two main spccies as they
while initiating operations and visualising its complex travel at different specd and get separated down the
dynamic process behavi our simulated in real time. from cromatographic column. etc.
anywhere In the world through a standard weh browser.
Thc technology used is Cadcentrc ' s PDMS . Review
Reality and Hyperplant for geometric Virtual Reality
modelling and Web access, and Process Systems
Enterprise ' s gPROMS for dynamic process modelling and
simulation, with Silicon Graphics supplying visualisation,
high performance computing , and web and network
infrastructure technology .
Review Reality enables a fly-through visualisation of
a 3D engincering modcl of a plant built in PDMS. a CAD
package. both of which are typically used to support plant
design and commissioning activities by contractors and
owners-operators. Hyperplant gives a web acccss interface
to the same geometric models and engineering data.
The specific example we built centred on our (real)
multipurpose batch pilot plant at Imperial College. The
dynamic model is thc same that we use in our operations
support environment described above. and includes a Fig. 2 Virtual Plant allows 3D visualisation of plant
simulation of the complete sequential control system and geometry and " walk through" in virtual reality . A detailed
advanced control of the reactor. for production of any real-time simulation of the process brings it to life by
number of batches of scveral dextrin products. For animating the VR model.
demonstration purposes. the plant model was extended
with the inclusion of a separation sec tion (comprising a
complex chromatographic column and associated final For batch plants in particular the visualisation
product and solvent circulation equipment) downstream of achieved is per-se a potent communication device . There
the reactor for purification of the desired sugar products. are many other interesting appli cations in support of both
The dynamic model of the chromatographic column de sign and operations. The most obvious opportunities for
includes a full distributed model with axial and radial Virtual Plant may be see n as:
diffu sion (modelled in gPROMS as partial differential i) A sophisticated concurrent design
equation). and all associated control and operation development. enabling geographically
sequences (modelled in gPROMS as tasks). dispersed design teams to examllle
The 3D geometric model of the plant reflects actual interaction between process and geometric
geometry and was derived from the actual CAD design design. layout and operation s
data used to build the plant. For presentation purposes the ii) An advanced operations support
plant was virtually re-located into a green field. blue sky environment. again for distributed teams. for
site . rather prettier than Imperial College ' s basement. remote operation. concurrent operations
The complex activities during process include with access from several sites. virtual
activating various flexibly configurable routes through the

543
 control room. operator training. model- independently from each other and communicate and keep
based hazard and operability analysis. synchronised through a standard web communication
remote emergency management. architecture.
At present. the hardware required to support advanced
For example. in batch plant design and retrofit. MILP virtual reality visualisation and high fidelity dynamic
based techniques may be used for selecting the optimal modelling as discussed above may be substantial for a full
type. number. size and connectivity of process equipment scale large plant. Nonetheless the performance of a rather
(e .g. Barbosa and Macchietto. 1997). The layout of such small two-processor workstation is already sufficient to
equipment on 2 and 3 dimensions may then be optlmised run our Virtual batch pilot plant in real time. With
(Georgiadis et al.. 1(97). As the next step. a more detailed multiprocessor hardware becoming both commonplace
geometric model could be easily built and standard and cheaper by the day. hardware requirements will not be
facilities of CAD packages used to automatically route the limiting bottleneck. In any case. it is likely that future
pipes and obtain actual pipe run lengths. These will IT architectures will probably consign such computation
determine. for example. product losses during changeovers workhorses to some company server hidden somewhere in
and cleaning. time delays and dynamic responses relevant the background and accessed through an intranet.
to control design and performance. etc.
What makes Virtual Plant technically possible is the
13. CONCLUSIONS
very strong object and engineering orientation of the
geometric and dynamic simulation models. In Review
Reality. a pipe is not just a wire frame. but a complete There are very strong economic drivers that require
engineering model of the artifact with a variety of that the batch process plants be viewed in a broad rather
associated data. from size to materials of construction. than narrow sense. as plant-wide. integrated systems.
connections to other parts of the plant. maintenance Specific requirement are for manufacturing facilities to
record. etc. On the other side, as already observed. the respond to rapid product and market innovation. to
process models of not just equipment but of task and produce flexibly a variety of products. while making
procedural controllers are also complex objects whIch efficient use of material and equipment resources. The
know their status and other engineering attributes, and can field is maturing rapidly and very substantial progress has
be similarly linked and traced into complex networks of been achieved over the last few years.
plants and procedures. Modelling and simulation of batch operations from
complex unit operations to entire plants can now be
routinely undertaken.
Dynamic optimisation IS becoming a key
methodology in aid of design and control activities.
Further work is necessary to extend the current methods to
include many discrete variables.
Some techniques for the design of procedural
controllers are starting to become useful in practical
applications. but this is still a very much open and
challenging area. The economic assessment presented
above indicates that this area should present some terrific
research opportunities.
As we move from control of a single batch to control
of batch manufacturing. schedul ing of operations and
integration of scheduling and control become important.
Again. much has been achieved over the last few years but
there is scope for further work. in particular for large scale
systems and on integration methodologies.
Fig. 3 Worldwide. instant access to all engineering and Many other areas which are not discussed here are
process simulation data is obtained over the WEB. clearly important. from monitoring and hazop analysis to
design.
Overall. it appears that a variety of new
The required software integration was achieved with methodologies and solution tools are now available and
no change at all to the underlying gPROMS model ready to be used to investigate new concepts in design.
through its standard open communication interfaces. On control and operations management of batch processes.
the geometric software side. only small changes were These will no doubt include modular. highly flexible plant
required to accommodate some new attributes (e.g. the designs with "smart" processing units having a degree of
temperature of a vessel and its colour mapping) previously self configuration and monitoring. flexibly linked in
irrelevant. The varIOUS applications run happily virtual supply chains. with no stocks and instant response

544
to new demands and products. Products. plants and Exchangers under Milk Foulmg. AIChE J .. 44. 4. 959-
processes will be designed in two weeks in a virtual 971.
Georgiadis. M.C. . G.E. Rotstein and S. Macchietto (I 988b).
prototyping environment and have zero defects by design.
Modelling and Simulation of Complex Plate Heat
... Well there is still some way to go, but it seems to me Exchangers Arrangements under Milk Fouling .
we are well on the right direction now. Batch processing . Computers & Chem . Engnrg . 22S. 331-338. ~
for so long the Cinderella of process engineering is ready Georgiadis. M.C.. G.E. Rotstein and S. Macchietto (1988c).
for a radical revolution. Optimal Design and Operation of Heat Exchanl!ers
under Milk Fouling. AIChE J .. 44. 9. 2099-2111. ~
Moreover. concepts and techniques developed in this
Georgiadis. M.C.. L.G. Papageorgiu, G.E. Rotstein and S.
field are exactly what is required to also transform stodgy Macchietto (1998d). Optimal Cleaning Policies in Heat
one-product, continuous plant into lively, flexible, agile Exchanger Networks under Milk Foulinl! . submitted
and responsive manufacturing nodes for the global process for publication. ~
supply chain. gBSS (1998) . User Manual. Process Systems Enterprise Ltd ..
London UK, on www . p5~nterpri .~.co m .
gPROMS 1.6 (1998). User Manual. Process Systems Enterprise
Ltd .. London UK. on www . ps~Ilterpri se.c om .
14. ACKNOWLEDGMENTS lshikawa T. . Y. Natori . L. Liberis and c.C.Pantelides (1997).
Modelling and Optimisation of an Industrial Batch
The work described was supported by EPSRC and Process for the Production of Dioctyl Phtalate .
several industrial organisations . The contribution of many Technical Report. Centre for Process Systems
Engineering, Imperial College.
students and collaborators to the development of the ideas
Jain V. and LE. Grossmann (1998). Cyclic Scheduling of
presented is gratefully acknowledged. Continuous parallel processes units with decaying
performance. AIChE J .. in press.
Kakhu . A .. B. Keeping. Y. Lu and c.c. Pantelides (1998). An
15. REFERENCES Open Software Architecture for Process Modelling and
Model-based Applications. To appear. Proceedings
Abe!. 0. , A. Helbig, W. Marquardt (1997) . Technical Report FOCAPO' 98, Snowbird, USA. July 5- \0.
LPT-1997-20, RWTH Aachen. Kanakamedala K.B., G.V. Reklaitis and V. Venkatasubramanian
Bansa!. V. R. Ross, J.D. Perkins and E.N . Pistikopoulos (1998). (1994). Reactive Schedule Modification in
Interactions of Design and Control : Double effect MUltipurpose Batch Chemical Plants. Ind. Eng. Chem.
distillation systems . Proceedings. DICOPS-5, Corfu, Res., 33, 77-90.
June 8-10. Kondili E., c.c. Pantelides and R.W.H. Sargent (1988). A
Barbosa. A.P. and S. Macchietto (1997). Design of Multipurpose General Algorithm for Scheduling of Batch Operations .
Batch Plants with Operational Pre-conditions. Proceedings of 3rd Intnl. Symp . On Process Systems
Investigayao Operacional , vol.l7. 2, 113-136 (in Engng., Sydney. 62-75 .
english.) Lionis V .. G.E. Rotstein. E. Mohamed and R. Weiss (1998) .
Barton P., R.J. Allgor, W.F. Feehery and S. Galan (1998). Integration of Production Scheduling with Activity
Dynamic Optimisation in a Discontinuous World. Ind. Based Costing Models. To appear. Proceedings
Eng. Chem. Res .. 37, 3, 966-981. FOCAPO' 98, Snowbird, USA. July 5-10.
Berber. R. (1995). Control of Batch Reactors: A Review. In Liu Z . and S. Macchietto (1998). An Advanced Operations
" Methods of Model-based Control" , R. Berber Ed .. Support System for a Batch Pilot Plant. Proceedings.
NATO-ASI Series, Kluver Press. 459-494. DICOPS-5, Corfu , June 8-10.
Benson, R. and J.D. Perkins (1997). The future of Process Louleh, Z .. M.V . LeLann. A. Chamayou and G. Casamatta
Control - A UK Perspective. In "Chemical Process (1996). A new heating-cooling system to improve
Control - V" (1. Kantor, C.E. Garcia and B. Camahan controllability of batch reactors. Chem. Eng. Sci. , 51 .
Eds) . CACHE, p.197. 11. 3163-3168.
Bonvin . D. (1997). Optimal Operation of Discontinuous Macchietto. S. (1997). Batch Food Processing : The proof is in
Reactors A Personal View. Proceedings the eating. In "Chemical Process Control - V" (1.
ADCHEM'97, Banff. Canada. June 9-11. 155-169. Kantor. C.E. Garcia and B. Carnahan Eds.), CACHE.
Charalambides M. (1996). Optimal design of Integrated Marikar M.T. , G. Rotstein . A. Sanchez and S. Macchietto
Processes. PhD Thesis. University of London. (1998). Computer Aided Analysis and Synthesis of
Cott . B. J. and S. Macchietto (1989). An Integrated Computer Procedural Controllers. Accepted. Procedingsof lEE IV
Aided System for the Operation of Batch Plants . Workshop on Discrete Event Systems . WODES.
Comput. and Chem. Eng. , 13. 11112, 1263-1272. Cagliary , Italy. August.
Engel!. S .. M. Fritz and K. Woellhaf (1998). Object-Oriented Mohideen M.J. J.D.Perkins and E.N . Pistikopoulos (1996 ).
Modeling and Simulation of MUltipurpose Batch Optimal design of dynamic systems under uncertainty.
Plants . to be published in "Computer Aided and AIChE J .. 42.8,2251 -2272.
Integrated Manufacturing Systems Techniques and Mohideen M.J . J.D.Perkins and E.N. Pi stikopoulos (1997).
Applications". Gordon & Breach. Towards an efficient numerical procedure for mixed
Engel!. S .. S. Kowalewski and B. Krol!h (1997). Discrete Events integer optimal control. Comput. chem. Engng ..
and Hybrid Systems in Proc~ess Control. In "Chemical 21.S457-S462.
Process Control - V" (1 . Kantor. C.E. Garcia and B. Rickard, Macchietto and Shah (1998). Intel!rated Decision
Carnahan Eds.). Support for Flexible MUltipurpose Plants. Technical
Georgiadis, M.C. . G.E. Rotstein and S. Macchietto (l988a) . Report, May .
Modelling and Simulation of Shell and Tube Heat Pantelides . c.c. (1994) .. In "Procedings of FOCAPD ' 94". L.T.
Biegler and M.F. Doherty Eds .. CACHE Corp.

545
Pantelides . c.c. (1996) . An Advanced Tool for process
Modelling, Simulation and Optimisation . Proceedings
of Computers Europe 111. Frankfurt , Chemical
Engineering Eds.
Pantelides , c.c. and H. Britt (1995 ) Multipurpose Process
Modelling Environments. In "Procedings of FOCAPD
' 94", L.T. Biegler and M.F. Dohert y Eds .. CACHE
Corp .. 128-141.
Reklaitis. G .. A. Suno!. D. Rippin and O. Hotacsu Eds. (1996) .
Batch Process Systems Engineering . Springer Verlag.
Rodrigues M.T.M .. L. Gimeno . C.A.s. Passos and M.D. Campos
(1996). Reactive Scheduling Approach for
Multipurpose Batch Chemical Plants. Computers
chem. Engmg .. S20, 1215-1226.
Sanches A. (1996). Formal Specification and synthesis of
Procedural controllers for Process Systems. Lecture
Notes on control and Information Sciences V. 212 ,
Springer Verlag.
Sanchez A .. G.Rotstein , N. Alsop and S. Macchietto (1998a) .
Synthesis and Implemetation of Procedural Controllers
for Event-driven Operations in Chemical Processes.
Submitted for publication .
Sanchez AC. Gollain and S. Macchietto ( 1998b ). Rapid
development of structured simulation models for the
verification of discrete-event control sy stems.
proceedings DYCOPS -5, Corfu, June 6-10.
Shah, N. ( 1998). Single- and Multi-site Planning and
Scheduling: Current Status and Future Challenges. To
appear, Proceedings FOCAPO' 98. Snowbird. USA,
July 5-10.
Schilling, G.H . (1997). Algorithms for Sort-term and Periodic
Process Scheduling and Rescheduling. PhD Thesis,
University of London.
SUPERBATCH 2.0 (1998). User Manual. Process Systems
Enterprise Ltd., London UK, on
www.psenterprise.com.
Toulouse, c. , J. Cezerac, M . Cabassoud, M.V. Lelann and G.
Casamatta (1996). Optimisation and scale-up of Batch
Chemical Reactors - Impact on safety constraints.
Chem. Eng. Sci., 51. 10, 2243 -2252.
Bogusch. R., B. Lohmann, W . Marquardt (1996). Computer-
Aided Modelling with MOD KIT. Technical Report
LPT-1996-08, RWTH Aachen .
StephanopoulosG. , G. Henning and H. Leone (1990) . MODEL-
LA A Language for Process Engineering. Part I and 11.
Computers chem. Engng . 14. 813 -869.

546