Content uploaded by Dominique Bonvin

Author content

All content in this area was uploaded by Dominique Bonvin on Sep 12, 2017

Content may be subject to copyright.

Control and Optimization of Batch Chemical Processes

Dominique Bonvin1,∗and Gr´egory Fran¸cois2

1Laboratoire d’Automatique, Ecole Polytechnique F´ed´erale de Lausanne

EPFL - Station 9, CH-1015, Lausanne, Switzerland.

2Institute for Materials and Processes, School of Engineering

The University of Edinburgh, Edinburgh EH9 3FB, UK.

∗corresponding author : dominique.bonvin@epﬂ.ch

November 30, 2016

Abstract

A batch process is characterized by the repetition of time-varying operations of ﬁnite duration. Due to the

repetition, there are two independent “time” variables, namely, the run time during a batch and the batch

index. Accordingly, the control and optimization objectives can be deﬁned for a given batch or over several

batches. This chapter describes the various control and optimization strategies available for the operation

of batch processes. These include online and run-to-run control on the one hand, and repeated numerical

optimization and optimizing control on the other. Several case studies are presented to illustrate the various

approaches

Keywords

Batch control, predictive control, iterative learning control, run-to-run control, batch process optimization,

dynamic optimization, optimizing control, run-to-run optimization.

1

1 Introduction

Batch processing is widely used in the manufacturing of goods and commodity products, in particular in

the chemical, pharmaceutical and food industries1. Batch operation diﬀers signiﬁcantly from continuous

operation. While in continuous operation the process is maintained at an economically desirable operating

point, in batch operation the process state evolves from an initial to a ﬁnal time. In the chemical industry for

example, since the design of a continuous plant requires substantial engineering eﬀort, continuous operation is

rarely used for low-volume production. Discontinuous operations can be of the batch or semi-batch type. In

batch operations, the products to be processed are loaded in a vessel and processed without material addition

or removal. This operation permits more ﬂexibility than continuous operation by allowing adjustment of

the operating conditions and the ﬁnal time. Additional ﬂexibility is available in semi-batch operations,

where reactants are continuously added by adjusting the feedrate proﬁles, or products are removed via some

outﬂow. We use the term batch process to include semi-batch processes.

In the chemical industry, many batch processes deal with reaction and separation operations. Reac-

tions are central to chemical processing and can be performed in an homogeneous (single phase) or hetero-

geneous (multi-phase) environment. Separation processes can be of very diﬀerent types, such as distillation,

absorption, extraction, adsorption, chromatography, crystallization, drying, ﬁltration and centrifugation.

The operation of batch processes follows recipes developed in the laboratory. A sequence of operations is

performed in a pre-speciﬁed order in specialized process equipment, yielding a certain amount of product.

The sequence of tasks to be carried out on each piece of equipment, such as heating, cooling, reaction,

distillation, crystallization and drying, is predeﬁned. The desired production volume is then achieved by

repeating the processing steps on a predetermined schedule.

This chapter describes the various control and optimization strategies available for the operation of

batch processes. Section 2 discusses the main features of batch processes, while Section 3 presents the two

types of models that are available for online and run-to run operation, respectively. The rest of the chapter

contains two parts, namely Part A comprising Sections 4-7 and concerned with control (online control, run-

2

to-run control, batch automation, and control applications), and Part B comprising Sections 8-10 and dealing

with optimization (numerical optimization, real-time optimization, and optimization applications). Finally,

a summary and outlook are provided in Section 11.

2 Features of Batch Processes

Process engineers have developed considerable expertise in designing and operating continuous processes.

On the other hand, chemists have been trained to develop new routes for synthesizing chemicals, often

using a batch or semi-batch mode of operation. The operation of batch processes requires considerable

attention, in particular regarding the coordination in time of various processing tasks such as charging,

heating, reacting, separation, cooling, and the determination of optimal temperature and feeding proﬁles.

Process control engineers have been eager to use their expertise in controlling and optimizing continuous

processes to achieve comparable success with batch processes. However, this is rarely possible, the reason

being signiﬁcant diﬀerences between continuous and batch processes. The two main distinguishing features

are discussed ﬁrst 2:

•Distinguishing feature #1:No steady-state operating point. In batch processes, chemical and physical

transformations proceed from an initial state to a very diﬀerent ﬁnal state. In a batch reactor, for

instance, even if the reactor temperature is kept constant, the concentrations, and thus also the reaction

rates, change signiﬁcantly over the duration of the batch. Consequently, there does not exist a steady-

state operating point around which the control system can be designed. The decision variables are

inﬁnite-dimensional time proﬁles. Furthermore, important process characteristics—such as static gains,

time constants and time delays—are time varying.

•Distinguishing feature #2: Repetitive nature. Batch processing is characterized by the frequent rep-

etition of batch runs. Hence, it is appealing to use the results from previous runs to optimize the

operation of subsequent ones. This has generated the industrially relevant topics of run-to-run control

and run-to-run optimization.

3

In addition to these two features, a number of issues tend to complicate the operation of batch processes:

•Nonlinear behavior. Constitutive equations such as reaction rates and thermodynamic relationships

are typically nonlinear. Since a batch process operates over a wide range of conditions, it is not possible

to use, for the purpose of control design and optimization, models that have been linearized around a

steady-state operating point, as this is typically done for continuous processes.

•Poor models. There is little time in batch processing for thorough investigations of the reaction, mixing,

heat- and mass-transfer issues. Consequently, the models are often poor. For example, it may well

happen that, in the production of specialty chemicals, the number of signiﬁcant reactions is unknown,

not to mention their stoichiometry or kinetics.

•Few speciﬁc measurements. The sensors that allow measuring concentrations online are rare. Chemical

composition is usually determined by drawing a sample and analyzing it oﬄine, that is, by using invasive

and destructive methods. Furthermore, the available measurements—often physical quantities such as

temperature, pressure, torque, turbidity, reﬂective index and electric conductivity—might exhibit low

accuracy due to the wide range of operation that the measuring instrument has to cover.

•Constrained operation. A process is typically designed to operate in a limited region of the state space.

In addition, the values that the inputs can take are upper and lower bounded. The presence of these

limitations (labeled constraints) complicates the design of operational strategies for two main reasons:

i) even if the process is linear or has been linearized along a reference trajectory, constraints make it

nonlinear, and ii) controllability might be lost when a manipulated variable hits a constraint. Due to

the wide operating range of batch processes, it is rarely possible to design the process so as to enforce

feasible operation close to constraints, as this is typically done for continuous processes. In fact, it has

been our experience that safety and operational constraints dominate the operation of batch processes.

•Presence of disturbances. Operator errors (e.g. wrong stirrer or solvent choice, incorrect material

loading) and processing problems (e.g. fouling of sensors and reactor walls, insuﬃcient mixing, incorrect

feeding proﬁles, sensor failures) represent major disturbances that, unfortunately, cannot be totally

4

ruled out. There are other unmeasured disturbances entering the process as the result of upstream

process variability such as impurities in the raw materials.

•Irreversible behavior. In processes with history-dependent product properties, such as polymerization

and crystallization, it is often impossible to introduce remedial corrections once oﬀ-speciﬁcation ma-

terial has been produced. This contrasts with continuous processes, where appropriate control action

can bring the process back to the desired steady state following an upset in operating conditions.

•Limited corrective action. The ability to inﬂuence the process typically decreases with time. This,

together with the ﬁnite duration of a batch run, limits the impact of corrective actions. Often, if a

batch run shows a deviation in product quality, the charge has to be discarded.

3 Models of Batch Processes

Model-based control and optimization techniques rely on appropriate mathematical representations. The

meaning of the qualiﬁer “appropriate” depends on the system at hand and the objectives of the study. Some

modeling aspects of batch processes are addressed next3.

3.1 What to Model?

Consider the example of a batch chemical reactor. The reactor system comprises the reactor vessel and

one or several reactions. For reasons mentioned above, the reactions are often poorly known in the batch

environment. Even when the desired reaction and the main side reactions are well documented, there

might exist additional poorly known or totally unknown reactions. For example, it is frequently observed

that the desired products and the known side products do not account for all the transformed reactants.

Hence, since the stoichiometry of the reaction system is not completely known, the kinetic models are often

of aggregate nature, that is, they encompass real and pseudo (lumped) reactions in order to describe the

observed concentrations. This way, the number of modeled reactions, and thereby the number of kinetic

5

parameters to be estimated, can be kept low. Furthermore, to determine the operational strategy for a

reactor, it is necessary to consider the reaction kinetics, reactor dynamics and operational constraints.

In industrial situations, the dynamics associated with the thermal exchange between the reactor and the

jacket are often dominant and must necessarily be included in the model. The reasoning developed here

for batch reactors4,5 extends similarly to other batch units such as batch distillation columns6,7 and batch

crystallizers 8,9 .

3.2 Model Types

The models used can be of several types, the characteristics of which are detailed next.

•Data-driven black-box models. Despite their simplicity, empirical input-output models are often able

to satisfactorily represent the relationship between manipulated and observed variables. Linear and

nonlinear ARMAX-type models are readily used to represent dynamic systems10,11. Neural network

methods have also been proposed to model the dynamics of batch processes12. More recently, re-

searchers have investigated the use of multivariate statistical partial least-squares (PLS) models as a

means to predict product quality in batch reactors13. Experimental design techniques can help as-

sess and increase the validity of data-driven models14. Although simple and relatively easy to obtain,

data-driven input-output models have certain drawbacks:

1. They often exhibit good interpolative capabilities, yet they are inadequate for predicting the

process behavior outside the experimental domain in which the data were collected for model

building. Such a feature signiﬁcantly limits the applicability of these models for optimization,

that is, the determination of better proﬁles that have not been seen before.

2. Input-output models represent a dynamic relationship only between variables that are manip-

ulated or measured. Unfortunately, some key variables, such as concentrations, often remain

unmeasured in batch processes and thus cannot be modeled via data-driven models.

•Knowledge-driven white-box models. A mechanistic state-space representation based on energy and

6

material balances is the preferred approach for modeling batch processes. The rate expressions describe

the eﬀect that the temperature and the concentrations have on the various rates. The model of a

given unit relates the independent variables (inputs and disturbances) to the states (concentrations,

temperature and volume) and the corresponding outputs.

The dynamic model of a batch process can be written as follows15:

˙xk(t) = Fxk(t), uk(t), xk(0) = x0,k ,(1)

yk(t) = Hxk(t), uk(t),(2)

zk=Zxk[0, tf], uk[0, tf],(3)

where tdenotes the run time, kthe batch or run index, xthe n-dimensional state vector, and uthe m-

dimensional input vector. There are two types of measured outputs, namely, the p-dimensional vector

of run-time outputs y(t) that are available online and the q-dimensional vector of run-end outputs z

that are available at the ﬁnal time tf. Note that zis not limited to quantities measured at the ﬁnal

time but can include quantities inferred from the proﬁles xk[0, tf] and uk[0, tf], such as the maximal

temperature reached during the batch.

In addition to the run-time dynamics speciﬁc to a given run, one has the possibility of updating the

initial conditions and the inputs on a run-to-run basis as follows:∗

x0,k+1 =Ixk[0, tf], uk[0, tf], x0,0=xinit ,(4)

uk+1[0, tf] = Kxk[0, tf], uk[0, tf], u0[0, tf] = uinit[0, tf].(5)

Note that, in this formulation, the ﬁnal time tfis the same for all batches.

Mechanistic models are typically derived from physico-chemical laws. They are well suited for a wide

range of process operations. However, they are diﬃcult and time-consuming to build for industrially

relevant processes. A sensitivity analysis can help evaluate the dominant terms in a model and retain

those that are most relevant to the processing objectives 16 . No realistic model is purely mechanistic,

as a few physical parameters typically need to be estimated from process data. As with data-driven

∗The initial conditions and inputs can be updated on the basis of several prior batches; the relationships proposed here consider

only the previous batch for simplicity of notation.

7

models, experimental design techniques17 are useful tools for building sound models from a limited

amount of data.

•Hybrid grey-box models. As a combination of the two extreme cases listed above, hybrid models can

be very helpful in certain situations. They typically possess a simple structure that is based on some

qualitative knowledge of the process 18 . The model parameters are particularly easy to identify if the

model structure can be reduced to an ARMAX-type form 19. For example, reaction lumping is often

exercized. Because of the aggregate structure of the model, it is necessary to adjust certain model

parameters online as, for example, the rate parameters might represent a diﬀerent aggregation initially

than toward the end of the batch. This can be done continuously or in some ad-hoc fashion, but

clearly at a rate slower than that of the main dynamics. Furthermore, tendency models have been

proposed 20. These models retain the physical understanding of the system but are expressed in a

form that is suitable to be ﬁtted to local behavior. Although tendency modeling appears to be of high

industrial relevance, it has received little attention in the academic research community.

3.3 Static View of a Batch Process

Due to the absence of a steady state, most signals in a batch process evolve with time. Consequently, the

decision variables (here the input proﬁles u[0, tf] for a given batch run) are of inﬁnite dimension. It is

however possible to represent a batch process as a static map between a ﬁnite number of input parameters

(used to deﬁne the input proﬁles before batch start) and the run-end outputs (representing the outcome at

batch end). The key element is the parameterization of the input proﬁles as

u[0, tf] = Uπ, t,(6)

with the input parameters π. This parameterization is achieved by dividing the input proﬁles into time

intervals and using a polynomial approximation (of which the simplest form is a constant value) within each

interval. The input parameter vector πcan also include switching times between intervals21 .

With such a parameterization, the batch process can be seen as a static map between the ﬁnite set

8

of input parameters πand the run-end outputs z:

z=Mπ.(7)

This static model indicates that, once the input parameters πhave been speciﬁed, it is possible to compute

the run-end outputs z. For this, one needs to generate the input proﬁles u[0, tf] using Eq. (6), integrate

Eq. (1) and generate zvia Eq. (3). Alternatively, with an experimental set-up, one can generate the input

proﬁles u[0, tf], apply them to the batch process and collect information regarding the run-end outputs z.

Although the static model (7) looks simple, one has to keep in mind that the batch dynamics are hidden by

the fact that only the relationship between the input parameters (before the run) and the run-end outputs

(after the run) are considered. The dynamics are re presented implicitly in the static map between πand

z. These two diﬀerent ways of looking at the operation of a given batch are illustrated in Figure 1. Very

naturally, the “Dynamic view” will be relevant for online operations, while the “Static view” will be more

suited for run-to-run operations.

x(t)

u(t) y(t)

z

x[0,tf]

u[0,tf]y[0,tf]z

π

Static view

Dynamic Model

(1) - (3)

Dynamic Model

(1) - (3)

Dynamic view

Static Model (7)

Input Parameterization

(6)

Figure 1: Two diﬀerent views of the same batch run. Dynamic view: the dynamic model describes the

behavior of the states x(t), the run-time outputs y(t) and the run-end outputs zas functions of the inputs

u(t). Static view: the static model z=Mπexplicits the relationship between the input parameters πand

the run-end outputs z.

9

PART A. CONTROL

Batch process control has gained popularity over the past forty years. This is due to the development of novel

algorithms with proven ability to improve the performance of industrial processes despite uncertainty and

perturbations. But the development of new sensors and of automation tools also contributed to this success.

At this stage, it is important to distinguish between control and automation. In practice, industry often

tends to consider that control and automation are two identical topics and use the two words interchangeably.

Although arguable, we will consider hereafter that control corresponds to the theory, that is, the research

ﬁeld, whereby new control methods, algorithms and controller structures are sought. On the other hand,

automation represents the technology (the set of devices and their organization) that is needed to program

and implement control solutions.

The control of batch processes diﬀers from the control of continuous processes because of the two

main distinguishing features presented in Section 2. First, since batch processes have no steady-state operat-

ing point, the setpoints to track are time-varying trajectories. Second, batch processes are repeated over time

and are characterized by two independent variables, the run time tand the run index k. The independent

variable kprovides additional degrees of freedom for meeting the control objectives when these objectives do

not necessarily have to be completed in a single batch but can be distributed over several successive batches.

Batch process control aims to track two types of references:

a. The run-time references yr ef (t) or yref [0, tf] are determined oﬄine either from accumulated experience

and knowledge of the process, or by solving a dynamic optimization problem (see Section 8.1). Exam-

ples of tra jectories to be tracked include concentration, temperature, pressure, conversion and viscosity

proﬁles. In general, trajectory tracking can be achieved in two ways: (i) via “online control” in a single

batch, whereby input adjustments are made as the batch progresses based on online measurements, or

(ii) via “run-to-run control” over several batches, where the input trajectories are computed between

batches using information gathered from previous batches. It is also possible to combine approaches

(i) and (ii) as will be detailed later.

10

b. The run-end references zref are associated with the run-end outputs that are only available at the end

of the batch. The most common run-end outputs are product quality, productivity and selectivity. As

with run-time references, the run-end references can be tracked online or on a run-to-run basis.

With these two types of control objectives and the two diﬀerent ways of reaching them (online and over

several runs), there are four diﬀerent control strategies as illustrated in Figure 2. It is interesting to see what

view (dynamic or static) each control strategy uses. It turns out that Strategies 1-3 use a dynamic view

of the batch process, with only Strategy 4 based on the static view. In the following, Section 4 addresses

online control (the ﬁrst row in Figure 1, Strategies 1 and 2), while Section 5 deals with run-to-run control

(the second row in Figure 1, Strategies 3 and 4).

4 Online Control

The idea of “online control” is to take corrective action during the batch, that is, in run-time t. It is possible

to do so with two objectives in mind, namely, the run-time reference trajectories yref (t) and the run-end

references zref . This will be discussed in the next two subsections.

4.1 Feedback Control of Run-time Outputs (Strategy 1)

We will address successively the application of conventional feedback control to batch processes, the thermal

control of batch reactors, and the stability issue for batch processes.

4.1.1 Conventional feedback control

The application of online feedback control to batch processes has the peculiarity that the setpoints are often

time-varying trajectories. Even if some of the controlled variables, such as the temperature in isothermal

operation, remain constant, the key process parameters (static gains, time constants and time delays) can

11

Run-end references

zref

Run-time references

yref (t) or yref [0,tf]

Control objective

Feedback control

1

Iterative learning control

2

Run-to-run control

4

Implementation

aspect

Online

--> u(t), u[t, tf]

Run-to-run

--> u [0,tf]

Predictive control

3

u ( t ) y(t)

FbC

P

u[0, y[0,

tf]

tf]

ILC with run update

P

[0, z

tf]

u

RtR with run update

π

U (π)

P

E

x(t)

^

MPC

u[t,

tf]

P

y(t)

z

pred

(t)

Figure 2: Control strategies for batch processes. The strategies are classiﬁed according to the control

objective (horizontal division) and the implementation aspect (vertical division). Each objective can be met

either online or on a run-to-run basis depending on the type of measurements available. The signal u(t)

represents the input vector at time t,u[0, tf] the corresponding input trajectories, y(t) the run-time outputs

measured online at time t, and zthe run-end outputs available at the end of the batch. P stands for “plant”,

E for “state estimator”, FbC for “feedback control”, MPC for “model predictive control”, ILC for “iterative

learning control”, and RtR for “run-to-run control”. When online and run-to-run control are combined, the

run-to-run part u[0, tf] can be considered as the feedforward contribution, while the online part u(t) is the

feedback part.

12

vary considerably during the duration of the batch.

Feedback control is implemented using PID techniques or more sophisticated alternatives such as

cascade control, predictive control, disturbance compensation (feedforward control) and time-delay compen-

sation22,23. The online feedback controller can be written formally as

u(t) = Ky(t), yref (t),(8)

where Kis the online control law for run-time outputs as illustrated in Figure 3.

Controller K

Batch Process y(t)

u(t)

yref(t)

Figure 3: Online feedback control of the run-time outputs yin a batch process.

One way to adapt for variations in process characteristics is via adaptive control24 . However,

adaptive control requires continuous process identiﬁcation, which is often impractical in batch processes due

to the ﬁnite batch duration. Instead, practitioners use conventional feedback control and schedule the values

of some of the controller parameters (mostly the gains, thus leading to gain scheduling 25 ). Appropriate

values of the control parameters can be computed oﬄine ahead of time for the various phases of the batch

run and stored for later use.

A few industrial applications of advanced feedback control are available in the literature. For exam-

ple, a successful implementation of temperature control in an industrial 35-m3semi-batch polymerization

reactor using a ﬂatness-based two-degree-of-freedom controller has been reported26. Also, the performance

of four diﬀerent controllers (standard PI control, self-tuning PID control, and two nonlinear controllers) for

regulating the reactor temperature in a 5-liter jacketed batch suspension methyl methacrylate polymerization

reactor has been compared27. As expected, the performance of the standard PI controller was the poorest

since the controller parameters were ﬁxed and not adapted to changing process characteristics. The self-

tuning PID control performed better because available measurements were used to adapt the controller to

13

the varying process characteristics. The two nonlinear controllers, which were based on diﬀerential-geometric

techniques requiring full-state measurement28, showed excellent performance despite signiﬁcant uncertainty

in the heat-transfer coeﬃcient.

4.1.2 Control of heat generation

The control of exothermic reactions in batch reactors is a challenging problem that is tackled using one of

various operation modes 29 :

1) Isothermal batch operation. Most reaction systems are investigated isothermally in the laboratory.

For safety and selectivity reasons, pilot-plant and industrial reactors are also run isothermally at the same

temperature speciﬁcations. However, the transfer function between the jacket inlet temperature and the

reactor temperature depends on the amount of heat evolved, and it can become unstable for certain combi-

nations of the reaction parameters29. Adaptation of the controller parameters is therefore necessary. Most

often, this is done by using pre-computed values, that is, via gain scheduling. An energy balance around the

reactor is sometimes used to estimate the heat generated by the chemical reactions, which represents the

major disturbance for the control system. This estimate can then be used very eﬀectively in a feedforward

scheme. It is important to mention that isothermal batch operation often exhibits low productivity because

it is limited by the maximal heat-generation rate (corresponding to the maximal heat-removal capacity of

the reactor vessel), which typically occurs initially for only a short period of time.

2) Isothermal semi-batch operation. The productivity of isothermal batch reactors can be increased

through semi-batch operation. This way, the reactant concentrations can be increased once the initial

phase characterized by the heat-removal limitation is over. For strongly exothermic reactions, semi-batch

operation has the additional advantage of improving the eﬀective heat-removal capacity through feeding of

cold reactants. This, in turn, increases the productivity of the reactor by allowing higher temperatures.

Hence, isothermal semi-batch operation is often the preferred way of running discontinuous reactors in

industry. The temperature is controlled by manipulating the feedrate of the limiting reactant. However, the

14

system exhibits non-minimum phase behavior, that is, a ﬂowrate increase of the cold feed ﬁrst reduces the

temperature before the temperature raises due to higher reaction rates, which signiﬁcantly complicates the

control30. Furthermore, in most semi-batch operations, it is useful to ﬁnish oﬀ with a batch phase so as to

fully consume the limiting reactant. Without a kinetic model, it is not obvious when to switch from the

semi-batch to the batch mode. Finally, although isothermal, the operation of a semi-batch reactor should

be such that the heat-removal constraint is never violated. This has forced industrial practice to be rather

conservative, with a temperature setpoint often chosen from adiabatic considerations.

3) Constant-rate batch operation. Since a key objective in the operation of batch reactors is to master

the amount of heat produced by the chemical transformations, it is meaningful to maintain a constant reaction

rate. The temperature is kept low initially when the concentrations are high, to be increased with time as

the reactants deplete. This way, the reactor productivity is considerably higher than through isothermal

operation. However, large temperature excursions can have a detrimental eﬀect on product selectivity. If

reaction kinetics are known, it is a simple matter to compute the temperature proﬁle that results in the

desired heat generation. If the heat evolution can be measured or estimated online, feedback control can

maintain it at the (constant) setpoint. Laboratory-scale reaction calorimeters have been devised to measure

or estimate the heat-evolution proﬁles of (possibly unknown) reaction systems31. The temperature proﬁles

obtained in the laboratory can then be used in production environments, at least in qualitative terms,

provided that suﬃcient heat-removal capacity is available.

4.1.3 Run-time stability of batch processes

Stability is of uppermost importance to prevent runaway-type behavior. However, run-time asymptotic

stability is of little value since tfis ﬁnite. Hence, one can say that “instability” can be tolerated for a batch

process because of its ﬁnite duration.

Yet, an important issue in run time is reproducibility, which addresses the question of whether the

trajectories of various runs with suﬃciently close initial conditions and identical input proﬁles will remain

15

close during the run. The problem of stability is in fact related to the sensitivity with respect to perturbations.

The basic framework for assessing this sensitivity is as follows:

•the system is perturbed by either variations of the initial conditions or external disturbances that aﬀect

the system states,

•sensitivity is assessed as a norm indicating the relative eﬀect of the perturbations.

For a system to be stable, this norm must remain bounded for a bounded perturbation, and it must go to

zero with time for a vanishing perturbation. However, when dealing with ﬁnite-time systems, it is diﬃcult

to infer stability from this norm since, except for some special cases such as ﬁnite escape time, boundedness

is guaranteed. Also, the vanishing behavior with time cannot be analyzed since tfis ﬁnite. The element of

interest is in fact the numerical value of the norm, and the issue becomes quantitative rather than binary

(yes-no). Details can be found elsewhere 15.

4.2 Predictive Control of Run-end Outputs (Strategy 2)

With a suﬃciently accurate process model, and in the absence of disturbances, tracking the proﬁles de-

termined oﬄine is often suﬃcient to meet the batch-end product quality requirements32. However, in the

presence of disturbances, following pre-speciﬁed proﬁles is unlikely to lead to the desired product quality.

Hence, the following question arises: Is it possible to design an online control scheme for eﬀective control of

run-end outputs using run-time measurements? Since such an approach amounts to controlling a quantity

that has not yet been measured, it is necessary to predict run-end outputs to compute the required corrective

control action. Model predictive control (MPC) is well suited for the task of controlling future quantities that

need to be predicted 33 . This approach to batch control can, therefore, be formulated as an MPC problem

with a shrinking prediction horizon (equal to the remaining duration of the batch) and an objective function

that penalizes deviations from the desired product quality at batch end34.

16

4.2.1 Model predictive control

With the MPC strategy, two steps need to be implemented online at each sampling time, namely, the

estimation of the current states and the computation of piecewise-constant proﬁles of future control moves.

First step: Online estimation of x(t). This is done via state estimation using Model (1)-(2) and measure-

ments of the run-time outputs y(t)35,36. The resulting state estimate is denoted ˆx(t). However, this task

is diﬃcult to implement with suﬃcient accuracy because of some of the limiting characteristics of batch

processes, in particular their nonlinear behavior, the presence of large model inaccuracies and the lack of

speciﬁc measurements.

Second step: Online prediction of zand computation of u[t, tf]. In the context of batch operation, predic-

tive control consists in determining the inputs u[t, tf] to meet zr ef by solving at time ta ﬁnite-horizon

open-loop optimal control problem. Concretely, we consider the dynamic model given by Eqns (1) and (3)

and solve the following optimization problem at the discrete time instant ti:

min

u[ti,tf]J:= 1

2∆zT(ti)P∆z(ti) + 1

2Ztf

ti

uT(t)R u(t)dt (9)

s.t. ˙x(t) = Fx(t), u(t), x(ti) = ˆx(ti),(10)

zpred(ti) = Zx[0, tf], u[0, tf],(11)

x(tf)∈ X ,(12)

where zpred(ti) are the run-end outputs predicted at time ti,∆z(ti) := zpred(ti)−zref the predicted run-end

errors, Pand Rare positive-deﬁnite weighting matrices of appropriate dimensions, Xis the bounded region

of state space where the ﬁnal state should lie, tiis the current discrete time instant at which the optimization

is performed, and tfthe ﬁnal time. Note that the proﬁles u[0, tf] and x[0, tf] include (i) the parts u[0, ti]

and x[0, ti] that were determined prior to the time ti, and (ii) the the parts u[ti, tf] and x[ti, tf] that are

computed at ti. The integral term in the objective function (9) is optional and serves to penalize large

control eﬀorts. Numerical optimization yields the control sequence, u∗[ti, tf], of which only the ﬁrst part,

u∗[ti, ti+h], is applied to the plant in an open-loop fashion, where his the sampling period. Numerical

optimization is then repeated at every sampling instant. Note that it is important to include a bounded

17

region for the ﬁnal states for the sake of stability33.

The control law can be written formally as

u[t, tf] = Pzpred(t), zr ef ,(13)

where Pis the predictive control law for run-end outputs and zpred (t) is the prediction of zavailable at time

tas illustrated in Figure 4.

Model Predictive

Controller P

Batch Process y(t)

z

ref

State Estimation

x(t)

‹

z

u[t, tf]

Figure 4: Online predictive control of the run-end outputs zin a batch process. The model in the “Model

Predictive Controller” block is used to compute zpred(t) from ˆx(t).

Linear MPC is widely used in industry, more so than nonlinear MPC. However, one may ask whether

linear MPC is suﬃcient to deal with highly nonlinear batch processes? Note that nonlinear MPC is an active

area of research. For example, a nonlinear MPC scheme has been proposed to regulate run-end molecular

properties within bounds in the context of batch polymerization reactors 37.

4.2.2 Alternative ways of meeting zr ef online

In MPC, zis predicted online as zpred(t) from the state estimates ˆx(t), and then the control proﬁle u∗[t, tf]

is computed via numerical optimization. This is rather demanding, in particular regarding the availability

of an accurate process model for both state estimation and input computation. This requirement can be

reduced signiﬁcantly with the approaches proposed next.

Use of data-driven PLS models. When the detailed dynamic model (1) is not available, it is possible to use

partial least-squares (PLS) regression models identiﬁed from historical data and a few additional experi-

18

ments38 . Then, in real time during the progress of the batch, it is possible to use the PLS model to estimate

the run-end outputs and track zref by adjusting u[t, tf] via run-to-run control (see Section 5.2).

Another less ambitious but very eﬃcient control approach has also been proposed, whereby, if the prediction

of zfalls outside a predeﬁned no-control region, a mid-course correction is implemented39. Note that the

approach can easily be modiﬁed to accommodate more than one mid-course correction.

Still another variant of the use of PLS models involves controlling the process in the reduced space (scores)

of a latent variable model rather than in the space of the discretized inputs40 .

Use of online tracking. An alternative for meeting run-end references using online measurements consists

in tracking some feasible trajectory, zref (t), whose main purpose is to enforce zref at ﬁnal time, that is,

zref (tf) = zref 41 . This necessitates to be able to measure or estimate z(t) during the batch.

5 Run-to-run Control

Run-to-run control takes advantage of the repetition of batches that is characteristic of batch processing.

One uses information from previous batches to improve the performance of the next batch. This can be

achieved with respect to both run-time and run-end objectives, as discussed in the next two subsections.

5.1 Iterative Learning Control of Run-time Proﬁles (Strategy 3)

The time proﬁles of the manipulated variables can be generated using Iterative Learning Control (ILC),

which exploits information from previous runs to improve the performance of the current run 42 . This

strategy exhibits the limitations of open-loop control with respect to the current run, in particular the

fact that there is no feedback correction for run-time disturbances. Nevertheless, this scheme is useful for

generating time-varying feedforward input terms. The ILC controller has the formal structure

uk+1[0, tf] = Iyk[0, tf], yr ef [0, tf],(14)

19

where Iis the ILC law for run-time outputs. ILC uses the entire proﬁles of the previous run to generate the

input proﬁles for the next run as illustrated in Figure 5. ILC has been successfully applied in robotics43,44

and in batch chemical processing 45,46.

ILC

Controller I

Batch Process

uk [0, tf]

k+1 --> k

uk+1[0, tf] y

k

[0, t

f

]

y

ref

[0, t

f

]

Figure 5: ILC of the run-time proﬁles yk[0, tf] in a batch process. The run update is indicated by k+ 1 →k.

5.1.1 ILC problem formulation

A repetitive batch process is considered in operator notation:

yk[0, tf] = Guk[0, tf], xk(0),(15)

where uk[0, tf] and xk(0) represent the input proﬁles and the initial conditions in run k,Gis the operator

representing the system. This relationship can be obtained from Model (1)-(2) as follows:

•With the inputs uk[0, tf] and the initial conditions xk(0) speciﬁed, integrate Eq. (1) to obtain xk[0, tf].

•Eq. (2) allows computing yk[0, tf] from xk[0, tf] and uk[0, tf].

For the case where the system operator is linear and the signals are in discrete time, Eq. (15) can

be expressed in matrix form as follows:

yk=G uk+y0,k,(16)

with yk∈ ℜNp,uk∈ ℜN m and G∈ ℜN p×N m, where Nis the ﬁnite number of time samples, pthe number

of outputs, and mthe number of inputs. y0,k ∈ ℜN p represents the response to the initial conditions xk(0).

ILC tries to improve trajectory following by utilizing the previous-cycle tracking errors. The ILC

update law for the inputs uk+1 is given as

uk+1 =Auk+Bek, ek=yr ef −yk,(17)

20

where yref ∈ ℜN p are the references to be tracked, and A∈ ℜN m×N m and B∈ ℜN m×Np are operators

applied to the previous-cycle inputs and tracking errors, respectively. In the remainder of this section

dealing with ILC, the signals without an explicit time dependency are expressed as vectors containing the

Nﬁnite time samples, that is, uk∈ ℜN m ,yk∈ ℜNp and ek∈ ℜNp .

5.1.2 ILC convergence and residual errors

An ILC law is convergent if the following limits exist:

lim

k→∞ yk=y∞and lim

k→∞ uk=u∞.(18)

One way to ensure convergence is that the relation between ukand uk+1 be a contraction mapping in some

appropriate norm:

kuk+1k ≤ ρkukk,0≤ρ < 1.(19)

Using (16) in Eq. (17) gives:

uk+1 = (A−BG)uk+B(yref −y0,k ).(20)

Note that (yref −y0,k ) does not represent errors as in Eq. (17), but the diﬀerences between the reference

trajectories and the responses to the initial conditions. Also, in contrast to Eq. (17), where ekdepends on

ukthrough yk, (yref −y0,k ) is independent of uk. Hence, the convergence of the algorithm depends on the

homogenous part of Eq. (20) and the condition for convergence is:

kA−BGk<1.(21)

If the iterative scheme converges, with uk=uk+1 =u∞and y0,k =y0,k+1 =y0, Eq. (20) gives:

u∞= (I−A+BG)−1B(yref −y0).(22)

Since the converged outputs y∞are not necessarily equal to the desired outputs yref , the ﬁnal tracking

errors can be diﬀerent from zero. The ﬁnal tracking errors are given by

e∞=yref −(Gu∞+y0) = I−G(I−A+BG)−1B(yref −y0).(23)

21

If BG is invertible, this equation can be rewritten as

e∞=G(BG)−1(I−A) (I−A+BG)−1B(yref −y0).(24)

Note that the residual errors are zero when A=Iand non zero otherwise. Hence, the choice A=Iprovides

integral action along the run index k. Zero errors imply that perfect system inversion has been achieved.

5.1.3 ILC with current-cycle feedback

In conventional ILC schemes, only the tracking errors of the previous cycle are used to adjust the inputs in

the current cycle. To be able to reject within-run perturbations, the errors occurring during the current cycle

can be used as well, which leads to modiﬁed update laws. However, from the points of view of convergence

and error analysis, these modiﬁed schemes simply correspond to diﬀerent choices of the operators Aand B

in Eq. (17) as shown next. Let

uk+1 =uff

k+1 +ufb

k+1,(25)

with uff

k+1 =¯

Auf f

k+¯

Bek, uf b

k+1 =Cek+1 ,(26)

where the superscripts (·)ff and (·)f b are used to represent the feedforward and feedback parts of the inputs,

and ¯

A,¯

Band Care operators.

Using Eq. (26) in Eq. (25) and combining it with Eqns (16)-(17) gives:

uk+1 = (I+CG)−1¯

A(I+CG)−¯

BGuk+ (I+CG)−1¯

B+C−¯

AC(yr ef −y0)

=A−BGuk+B(yref −y0),(27)

where A= (I+CG)−1(¯

A+CG) and B= (I+C G)−1¯

B+ (I−¯

A)C. The convergence condition and

residual errors can be analyzed using the operators Aand B, similarly to what was done in Section 5.1.2.

5.1.4 ILC with improved performance

The standard technique for approximate inversion is to introduce a forgetting factor in the input update 47 .

This causes the residual tracking errors to be non zero over the entire interval. Note, however, that the

22

main diﬃculty with the feasibility of inversion arises during the ﬁrst part of the trajectory due to unmatched

initial conditions48. After a certain catch-up time, trajectory following is relatively easy. Hence, the idea is

to allow non-zero tracking errors early in the run and have the errors decrease progressively with run time t,

which can be achieved with an input shift. This is documented next as part of an ILC implementation that

includes (i) input shift, (ii) shift of the previous-cycle errors, and (iii) current-cycle feedback49.

The iterative update law is written as follows:

uk+1(t) = uf f

k+1(t) + Kf b ek+1(t),(28)

where Kf b ∈ ℜNm×N p are the proportional gains of the online feedback controller. The feedforward part of

the current inputs, uf f

k+1(t), consists of shifted versions of the feedforward part of the previous inputs and

the previous-cycle tracking errors:

uff

k+1(t) = uf f

k(t+δu) + Kf f ek(t+δe),(29)

where Kf f ∈ ℜNm×N p are the proportional gains of the feedforward controller, δuthe time shift of the

feedforward input trajectories and δethe time shift of the previous-run error trajectory. The remaining parts

of the inputs and errors are kept constant, that is, uf f

k[tf−δu, tf] = uff

k(tf−δu) and ek[tf−δe, tf] = ek(tf−δe),

respectively.

5.2 Run-to-run Control of Run-end Outputs (Strategy 4)

In this case, the interest is to steer the run-end outputs ztowards zref using the input parameters π. For

this, one uses the static model (7) that results from the input parameterization (6). The run-to-run feedback

control law can be written as

πk+1 =Rzk, zref ,(30)

uk[0, tf] = Uπk, t,(31)

23

where Urepresents the input parameterization and Ris the run-to-run feedback control law for run-end

outputs as illustrated in Figure 6. For example, one can use the discrete integral control law

πk+1 =πk+K(zref −zk),(32)

where Kis an appropriate gain matrix. To be able to meet exactly all the references, it is necessary to have

at least as many inputs parameters as there are run-end outputs.

Run-to-run

Controller R

Batch

Process

zk

uk[0, tf]

Input

Generator

Output

Selector

yk[0, tf]

πk

k+1 --> k

Static Map

zref

πk+1

Figure 6: Run-to-run control of the run-end outputs zin a batch process. The input generator constructs

uk[0, tf] from πkaccording to Eq. (31), while the output selector selects the run-end outputs.

5.2.1 Control algorithm

Run-to-run control proceeds as follows:

1. Initialization. Parameterize the input proﬁles u(t) as Uπ , tin such a way that the input parameter

vector πhas the same dimension as the vector of run-end outputs zin order to generate a square

control problem. Set k= 1 and select an initial guess for πk.

2. Iteration. Construct the inputs uk[0, tf] using Eq. (31) and apply them to the process. Complete the

run and determine zkfrom the measurements.

3. Update the input parameters using the control law (30). Set k:= k+ 1 and repeat Steps 2-3 until

convergence.

24

5.2.2 Convergence analysis

The convergence of the run-to-run algorithm can be determined by analyzing the closed-loop error dynamics,

where the errors are ek:= zref −zk.

We discuss next a convergence analysis that considers a linearized version of System (7) and the

linear integral control law (32). With the linear model,

zk=S πk,(33)

where S=∂M

∂π

¯πis the sensitivity matrix computed at the nominal operating point corresponding to ¯π, the

linearized error dynamics become:

ek+1 =ek−S Kek= (I−S K)ek.(34)

The eigenvalues of the matrix (I−S K) should be within the unit circle for the algorithm to converge.

Convergence is determined by both the sensitivity matrix Sand the controller gains K. Convergence analysis

of run-to-run control algorithms is also possible when the relationship between πkand zkis nonlinear, but

it typically requires additional assumptions regarding the nature of the nonlinearities50 .

5.2.3 Run-to-run stability

The interest in studying stability in the run index karises from the necessity to guarantee convergence of

run-to-run control schemes. Here, the standard notion of stability applies as the independent variable kgoes

to inﬁnity. The main conceptual diﬀerence with the stability of continuous processes is that “equilibrium”

refers to entire trajectories. Hence, the norms have to be deﬁned in the space of functions Lsuch as the

integral squared norm l2of the signal x(t):

kx[0, tf]kl2=Ztf

0x(t)2dt1/2.(35)

For studying stability with respect to the run index k, System (1) is considered under closed-loop

operation, that is, with all possible online and run-to run feedback loops. Hence, we no longer restrict our

25

attention to the input-output run-to-run π-zbehavior, but we consider the entire state vector. For simplicity,

let us assume that tfis constant for all runs. When dealing with the kth run, the tra jectories of the (k−1)st

run are known, which ﬁxes uk[0, tf] according to the ILC control law (14). These input proﬁles, along with

the online feedback law (8) or (13), are applied to System (1) to obtain xk(t) for all tand thus xk[0, tf]. All

these operations can be represented formally as:

xk[0, tf] = Fxk−1[0, tf], x0[0, tf] = xinit[0, tf],(36)

where xinit[0, tf] are the initial state tra jectories obtained by integration of Eq. (1) for k= 0, x0,0=xinit

and u0[0, tf] = uinit[0, tf]. Eq. (36) describes the run-to-run dynamics associated with all control activities.

Run-to-run stability is considered around the equilibrium trajectory computed from (36), that is ¯x[0, tf] =

F¯x[0, tf], and is investigated using a Lyapunov-function method51.

6 Batch Automation

Although batch processing was dominant until the 30’s, batch process control received signiﬁcant interest

only 50 years later 52 . This is mainly due to the fact that, when control theory started to emerge in the 40’s,

there was a shift from the dominance of batch processing to continuous operations. Hence, the dynamics and

control of batch processes became a subject of investigation in the 80’s. Engineers then tried to carry over

to batch processes the experience gained over the years with the control of continuous processes. However,

the speciﬁcities of batch processes make their control quite challenging.

This is also true at the implementation level. If continuous processes typically exhibit four main

regimes, namely, start-up, continuous operation, possibly grade transition and shut down, batch processes

are much more versatile since, in the absence of a steady state, the system evolves freely from a set of initial

conditions to a ﬁnal state. During this transient operation, the recipe can be rather varied, ranging from

basic isothermal operation to a succession of complex operations that can involve discrete decisions. Also,

batch processes are often repeated over time, with or without changes between batches, which makes the

task of designing a control strategy much more involved than for continuous operations.

26

The control of batch operations relies on various types of controllers that include stand-alone con-

trollers, programmable logic controllers (PLCs), distributed control systems (DCS), and personal computers

(PCs). These four platforms are discussed next.

6.1 Stand-alone Controllers

The basic stand-alone controllers are single-loop controllers (SLCs)53. These devices generally embed a

microprocessor with ﬁxed functionalities such as PID control. The popularity of SLCs arose from their

simplicity, low cost and small size. Originally, stand-alone controllers were capable of integrating dual-

loop control, that is, they can handle two control loops with basic on/oﬀ and PID control. More recently,

these devices have incorporated self-tuning algorithms and, more importantly for batch processes, time

scheduling and sequencing functionalities. Still, the main advantage of SLCs is their low cost per control

loop. Furthermore, they are often used in parallel or in combination with more advanced control structures

such as DCS, mainly because they can achieve very acceptable control performance for minimal investment

regarding cost, maintenance and required knowledge.

6.2 Programmable Logic Controllers

PLCs were introduced in the process industries in the early 70’s, following the development in the automotive

industry, as computing systems “that had the ﬂexibility of a computer, yet could be programmed and main-

tained by plant engineers and technicians”53. Before PLCs, relays, counters and timers were mainly used,

but they oﬀered much less ﬂexibility. PLCs can perform complex computations as they have a computational

power that is comparable to a small PC 54 .

Considerable progress was made with the introduction of micro-PLCs, which can be installed close to

the process at a much lower cost. The development of Supervisory Control and Data Acquisition (SCADA)

constitutes another breakthrough that has increased the scope of applications. It is possible to handle

numerous operations distributed over a large distance, thereby opening up the applicability of PLCs to cases

27

where batch process operations are fully integrated in the plant-wide operation of a production site.

6.3 Distributed Control Systems

DCS are control systems with control elements spread around the plant. They generally propose a hierarchical

approach to control, with several interconnected layers operating at diﬀerent time scales. At the top of the

hierarchy is the production/scheduling layer, where the major decisions are taken on the basis of market

considerations and measurements collected from the plant. These decisions are sent to an intermediate layer,

where there are used, together with plant measurements and estimated ﬂuctuations on price and raw-material

quality, to optimize the plant performance. These decisions can take the form of setpoint trajectories for

low-level controllers. This intermediate layer, which is often referred to as the real-time optimization layer, is

implemented via DCS55. Because of their distributed nature, DCS are helpful to integrate batch operations

into the continuous operation of a plant, which has considerable impact on the planning, scheduling and

real-time optimization tasks 56 . The control and optimization hardware has improved continuously over the

past decades, thereby keeping pace with the development of advanced control and optimization methods.

Although the diﬀerences between DCS and SCADA can appear to be subtle, one could state the

following56: DCS (i) are mainly process driven, (ii) are focused on small geographic areas, (iii) are suited

to large integrated chemical plants, (iv) rely on good data quality, and (v) incorporate powerful closed-loop

control hardware. On the other hand, SCADA is generally more “event driven”, which makes it more suited

to the supervision of multiple independent systems.

6.4 Personal Computers

PCs can also be used for the control and optimization of batch processes. It has been reported that 25% of

batch process control involve personal computers. Of course, this ﬁgure does not mean that 25% of batch

process control involves only PCs, as PCs are often used in combination with other controllers such as SLCs

or PLCs. In industry, PCs are not only used to program and run control algorithms, but they include

28

many other elements such as interfaces, communication protocols and networking53. The development of

user-friendly software interfaces, of simulation tools, and of fast and reliable optimization software has been

key to the development of batch processing. With the relatively low cost of PCs, it has become economically

viable to use them routinely for control. While a PC gives a lot of ﬂexibility, especially for taking educated

decisions in the context of a multi-layered control structure, it can also be used as a stand-alone advanced

controller or even as an online optimization tool. In practice, however, PC-based control is often limited to

supervisory control, where the PC sends recommendations to low-level controllers. This can take the form of

input proﬁles to implement or of output tra jectories to track during batch operation. These recommendations

can then be implemented directly by means of low-level controllers, or manually by an operator who can

take the ﬁnal decision of implementing or not the recommendations53.

7 Control Applications

7.1 Control of Temperature and Final Concentrations in a Semi-batch Reactor

7.1.1 Reaction system

Consider the reaction of pyrrole Awith diketene Bto produce 2-acetoacetyl pyrrole C. Diketene is a very

aggressive compound that reacts with itself and other species to produce the undesired products D,E, and

F57. The desired and side reactions are

A+B→C,

2B→D,

C+B→E,

B→F.

The reactions are highly exothermic, that is, they produce heat. The reactor is operated in semi-batch mode

with Apresent initially in the reactor and Badded with the feedrate proﬁle u(t). Moreover, the reaction

29

system is kept isothermal by removing the heat produced by the chemical reactions through a cooling jacket

surrounding the reactor. The ﬂowrate of cooling ﬂuid is adjusted to keep the reactor temperature at its

desired setpoint.

7.1.2 Model of the reactor

The nonlinear dynamic model of the reaction system is obtained from a material balance for each species

and a total mass balance 58 :

dcA

dt =−u

VcA−k1cAcB,(37)

dcB

dt =u

V(cin

B−cB)−k1cAcB−2k2c2

B−k3cCcB−k4cB,(38)

dcC

dt =−u

VcC+k1cAcB−k3cCcB,(39)

dcD

dt =−u

VcD+k2c2

B,(40)

dcE

dt =−u

VcE+k3cCcB,(41)

dcF

dt =−u

VcF+k4cB,(42)

dV

dt =u, (43)

where ciis the concentration of the ith species, Vis the reactor volume, kjis the rate constant of the jth

reaction, cin

Bis the inlet concentration of B, and uis the volumetric feedrate of B.

7.1.3 Control objective

One would like to operate the reactor safely and eﬃciently. If the heat produced cannot be removed by

the cooling jacket, the reactions accelerate and produce even more heat. This positive feedback eﬀect,

known as thermal runaway, can result in high temperature and pressure and thus lead to an explosion;

this is precisely what caused the Bhopal disaster59. The performance of the reactor is evaluated by its

productivity, namely, the amount of desired product available at ﬁnal time, as well as its selectivity, namely,

the portion of converted reactant Bthat forms the desired product C. Note that the performance criteria

30

(productivity and selectivity) are evaluated at the ﬁnal time, whereas the manipulated variable (feedrate

of B) is a run-time proﬁle. The control objective is to operate isothermally at 50 ◦C and match the ﬁnal

concentrations cB,ref and cD,ref obtained in the laboratory

7.1.4 Online control, ILC and run-to run control

The control strategies, illustrated in Figure 7, include (i) a cascade scheme for online feedback control of

Tr(t) and ILC update of Tf f

j,ref [0, tf], and (ii) run-to-run control of the ﬁnal concentrations through update

of the input parameters u1and u2.

cooling fluid

Slave

B

Master

TT

TT

R2R

c

B (

t

f

) c

D

( t

f

)

u1 u2

ILC Tr

T

ff [0,tf]

j,ref

[0,tf]

T

fb

j,ref

T

r,ref

T

j,ref

T

r

T

j

Figure 7: Online and run-to-run strategies for controlling the reactor temperature Tr(t) and the ﬁnal con-

centrations cB(tf) and cD(tf).

Control of the reactor temperature Tris implemented by means of a PI/P-cascade scheme that

adjusts the ﬂowrate of cooling ﬂuid. The master loop contributes a feedback term to the reference value for

the jacket temperature Tj. This reference value, which also includes a feedforward term, reads

Tj,ref,k (t) = Tf f

j,ref,k (t) + Tf b

j,ref,k (t),(44)

Tfb

j,ref,k (t) = KRek(t) + 1

τIZt

0

ek(τ)dτ,(45)

Tff

j,ref,k+1 [0, tf] = Tf f

j,ref,k [0, tf] + KILC ek[0, tf],(46)

31

where ek(t) := Tr,ref (t)−Tr,k(t), KRis the proportional gain, and τIis the integral time constant of the PI

master controller. KI LC is the gain of the ILC controller. Eq. (45) computes the feedback term Tf b

j,ref,k (t)

using a PI feedback law, while Eq. (46) implements run-to-run adaptation of the feedforward term Tf f

j,ref,k (t)

based on ILC.

The second manipulated variable is u, the feedrate of reactant B, through which the two reactions

can be steered and brought to the desired ﬁnal concentrations. Since these ﬁnal concentrations are measured

only at the end of the batch, the feedrate of Bis adjusted on a run-to-run basis. Two input parameters are

needed to control these two concentrations. Hence, the feedrate proﬁle u[0, tf] is parameterized using the

two feedrate levels u1and u2, each valid over half the batch time. The sensitivities of the ﬁnal concentrations

z= (cB(tf), cD(tf))Twith respect to π= (u1, u2)Tis evaluated experimentally, which gives S=∂z

∂π

k. With

this notation, the discrete integral control law has the form

πk+1 =πk+KRtR S−1[zref −zk],(47)

where KRtR is the 2 ×2 diagonal gain matrix of the run-to-run controller.

The contributions of ILC and run-to-run control to the control of reactor temperature and ﬁnal

concentrations are illustrated in Figures 8 and 9, respectively. One sees that adjustment of Tf f

j,ref using ILC

reduces the maximal temperature excursion from 57.1◦C to 52.7◦C.

7.2 Scale-up via Feedback Control

Short times to market are required in the specialty chemicals industry. One way to reduce this time to

market is by skipping the pilot-plant investigations. However, due to scale-related diﬀerences in operating

conditions, direct extrapolation of conditions obtained in the laboratory is often impossible, especially when

terminal objectives must be met and path constraints respected. In fact, ensuring feasibility at the industrial

scale is of paramount importance. This section presents an example for which the combination of online and

run-to-run control allows meeting production requirements over a few batches.

32

01234

-10

10

30

50

70

T [°C]

t [h]

tf

01234

-10

10

30

50

70

T [°C]

t [h]

tf

Tj,ref

ff

Tr

Tj,ref

Tj,ref

Tr

Tj,ref

ff

Tj

Tj

Figure 8: Temperature control around the constant setpoint Tr,ref = 50 ◦C, without ILC (left), and after

three ILC iterations (right). The red and grey solid lines represents the reactor temperature Trand the

jacket temperature Tj, respectively. The blue and green dashed lines represent the feedforward term of the

jacket temperature setpoint Tf f

j,ref and the resulting jacket temperature setpoint Tj,r ef, respectively.

0 1 2 3 4

10

20

30

40

50

u [l/h]

t [h]

tf

01234

0

0.1

0..2

0..3

t [h]

tf

cB,ref

u1

u2

0 1 2 3 4

0

0.1

0..2

0..3

t [h]

tf

cD,ref

cD

cB

Figure 9: Run-to-run control for meeting the run-end concentrations cB,ref and cD,ref . The inlet feedrate is

parameterized using the two levels u1and u2(left). The concentration proﬁles for cB(t) and cD(t) are shown

in the center and at right, respectively. The dashed lines show the initial proﬁles, while the solid lines show

the corresponding proﬁles after two iterations. The references cB,ref and cD,ref are reached via run-to-run

adjustment of u1and u2.

33

7.2.1 Problem formulation

Consider the following parallel reaction scheme60:

A+B−→ C, 2B−→ D .

The desired product is C, while Dis undesired. The reactions are exothermic. A 1-L reactor is used in the

laboratory, while a jacketed reactor of 5 m3is used in production. The manipulated inputs are the feedrate

F(t) and the coolant ﬂowrate through the jacket Fj(t). The operational requirements can be formulated as:

Tj(t)≥10◦C,(48)

yD(tf) = 2nD(tf)

nC(tf) + 2 nD(tf)≤0.18,(49)

where Tj(t) is the jacket temperature, and nC(tf) and nD(tf) denote the numbers of moles of Cand Dat

ﬁnal time.

7.2.2 Laboratory recipe

The recipe obtained in the laboratory proposes to initially ﬁll the reactor with A, and then feed B at some

constant feedrate ¯

F, while maintaining the reactor isothermal at Tr= 40◦C. As cooling is not an issue for

the laboratory reactor equipped with an eﬃcient jacket, experiments were carried out using a scale-down

approach, that is, the cooling rate was artiﬁcially limited so as to anticipate the limited cooling capacity of

the industrial reactor. Scaling down is performed by the introduction of an operating constraint that limits

the cooling capacity; in this case, the maximal cooling capacity of the industrial reactor is simply divided

by the scale-up factor:

[qc,max]lab =[(Tr−Tj,min )UA]prod

r,(50)

where r= 5000 is the scale-up factor and U A = 3.7×104J/mol ◦C is the estimated heat-transfer capacity

of the production reactor. With Tr−Tj,min = 30◦C, the maximal cooling rate is 222 J/min. Table 1

summarizes the key parameters of the laboratory recipe and the (simulated) laboratory results.

34

Table 1: Laboratory recipe and results for the scale-up problem.

Recipe parameters Laboratory results

Tr= 40◦CcB,in = 5 mol/L nC(tf) = 0.346 mol

cA,o = 0.5 mol/L cB,o = 0 mol/L yD(tf) = 0.170

V0= 1 L tf= 240 min max

tqc(t) = 182.6 J/min

¯

F= 4 ×10−4L/min

7.2.3 Scale-up via online control, ILC and run-to run control

The goal of scale-up is to reproduce in production the productivity and selectivity that are obtained in the

laboratory, while enforcing the desired reactor temperature. With selectivity and productivity as run-end

outputs, the feedrate proﬁle F[0, tf] is parameterized using two adjustable parameters, namely, the constant

feedrate levels F1and F2, each one valid over half the batch time. Temperature control is done via a combined

feedforward/feedback scheme as illustrated in Section 7.1.

A control problem can be formulated with the following manipulated variables (MVs), controlled variables

(CVs) and setpoints (SPs):

•MVs: u(t) = Tj,ref (t), π=F1, F2T

•CVs: y(t) = Tr(t), z=nC(tf), yD(tf)T

•SPs: yref = 40◦C, zref =1630 mol,0.17 T

.

The setpoints are chosen as follows:

•Tr,ref = 40◦C is the value used in the laboratory.

•nC,ref is the value obtained in the laboratory multiplied by the scale-up factor rminus a 100-mol backoﬀ

to account for scale-related uncertainties and run-time disturbances. The introduction of backoﬀ makes

the control problem more ﬂexible.

•yD,ref is the value obtained in the lab, which respects the upper bound of 0.18.

35

The control scheme is shown in Figure 10. The input proﬁles are updated using (i) the cascade feedback

controller Kto control the reactor temperature Tr(t) online, (ii) the ILC controller Ito improve the reactor

temperature by adjusting Tf f

j,ref (t) , and (iii) the run-to-run controller Rto control z= [ nC(tf), yD(tf)]T

by adjusting π= [ F1, F2]T. Details regarding the implementation of the diﬀerent control elements can

be found elsewhere 60 .

Online

Measurements

Online

R

Run-to-run

K

I

Run-end

Measurements

Batch

Process

Trajectory

Generation

Run

Update

Tj,ref,k+1[0,tf]

ff

Tj,ref,k(t)

ff

Tj,ref,k(t)

fb

Tj,ref,k(t)

Tj,ref,k(t)

F1,k+1, F2,k+1

F1,k, F2,k

Fk(t)

Tr,ref

Tr,k(t)

nC,ref yD,ref

Tr,ref

xk(t)

xk[0,tf]

ek(t)

ek[0,tf]

+

+

nC,k(tf)

yD,k(tf)

Figure 10: Control scheme for scale-up implementation. Notice the distinction between online and run-to-

run activities. The symbol ∇represents the concentration/expansion of information between a proﬁle (e.g.

xk[0, tf]) and an instantaneous value (e.g. xk(t)).

7.2.4 Simulation results

The recipe presented in Table 1 is applied to the 5-m3industrial reactor, equipped with a 2.5-m3jacket.

In this simulated case study, uncertainty is introduced artiﬁcially by modifying the two kinetic parame-

ters, which are reduced by 25% and 20%, respectively. Also, Gaussian noise with standard deviations of

0.001 mol/L and 0.1 ◦C is considered for the measurement of the ﬁnal concentrations and for the reactor

temperature, respectively. It turns out that, for the ﬁrst run, application of the laboratory recipe with

36

π=r¯

F , r ¯

FTviolates of the ﬁnal selectivity of D in the ﬁrst batch. Upon adapting the MVs with the

proposed scale-up algorithm, the free parts of the recipe are modiﬁed iteratively to achieve the production

targets for the industrial reactor, as illustrated in Figure 11.

yD(tf)

5 10 15

0.165

0.17

0.175

0.18

0.185

Batch index

nC(tf)

1520

1560

1600

1640

1480

Figure 11: Evolution of the production of C,nC(tf), and the yield of D,yD(tf), for the large-scale industrial

reactor. Most of the improvement is achieved in the ﬁrst 7 runs (the dashed lines represent the target values).

7.3 Control of a Batch Distillation Column

A binary batch distillation column is used to illustrate in simulation the application of iterative learning

control. This example is described in detail elsewhere61.

7.3.1 Model of the column

A binary batch distillation column with pequilibrium stages is considered (see Figure 12). Using standard

assumptions62 and writing molar balance equations for the holdup in the reboiler and for the liquid on the

37

Stage p

Reboiler

Stage 2

Condenser

Distillate

Measurement

Input r = L/V

M2

Mp

Mp+1

V, y1

V, yp-1

V, y2

V, yp

M1, x

1

L, xp

L, x3

L, x2

L, x

p+1

D, xp+1

fD(t)

xD(t)

Figure 12: Binary batch distillation column.

various stages and in the condenser, the following model of order (p+ 2) is obtained:

˙

M1= (r−1) V, (51)

˙x1=V

M1

(x1−y1+rx2),(52)

˙xi=V

Miyi−1−yi+r(xi+1 −xi), i = 2,...,p, (53)

˙xp+1 =V

Mp+1

(yp−xp+1),(54)

where xiis the molar liquid fraction, yithe molar vapor fraction, Mithe molar holdup on Stage i,Vthe

vapor ﬂowrate, and Dthe distillate ﬂowrate. Stage 1 refers to the reboiler, Stage pto the top of the column,

and Stage p+ 1 to the condenser. The internal reﬂux ratio r=L

V=V−D

Vis considered as the manipulated

variable. The vapor-liquid equilibrium relationship is:

yi=α xi

1 + (α−1)xi

, i = 1,··· , p , (55)

38

where αis the relative volatility. The composition of the accumulated distillate, xD, which is measured with

the sampling time h= 0.1h, is given by:

xD(t) = Pp

i=1 xi(t)Mi(t)−xi(0)Mi(0)

M1(t)−M1(0) .(56)

The model parameters, the initial conditions and the control objective xD,ref are given in Table 2.

Table 2: Model parameters and initial conditions, i= 2,··· , p.

p10 M1(0) 100 kmol

α1.6x1(0) 0.5

Mi0.2 kmol xi(0) 0.5

Mp+1 2 kmol xp+1(0) 0.5

V15 kmol/h xD,ref 0.9

tf10 h h0.1 h

7.3.2 Operational objective

A batch is divided into 2 operational phases:

1. Start-up phase with full reﬂux up to time ts:r= 1, t= [0, ts], ts= 1.415 h.

2. Distillation phase: r∈[0,1], t= (ts, tf], tf= 10 h.

The objective is to adjust the reﬂux ratio to get the distillate purity xD(tf) = 0.9 at the end of the

batch. This will be obtained by tracking an appropriate reference trajectory for the distillate purity xD(t)

in the second phase. This trajectory, which has the property of ending at the desired distillate purity at

ﬁnal time, is chosen to be linear, with xD,r ef (ts) = 0.925 and xD,ref (tf) = 0.9. The accumulated distillate

purity xD(t) is measured in Phase 2.

In order to obtain a realistic test scenario, the following uncertainty is considered:

39

- Perturbation: The vapor rate ﬂuctuates every 0.5hfollowing an uniform distribution in the range

V= [13,17] kmol/h.

- Measurement noise: 5% multiplicative Gaussian noise is added to the product composition xD(t).

The values of the squared tracking error Ptf

tse2(t) and the ﬁnal tracking error e(tf) upon convergence

are averaged over 20 realizations of the perturbation and measurement noise. Also, the variance ve(tf) of

the ﬁnal tracking error is calculated from 20 realizations.

7.3.3 Trajectory tracking via ILC

Trajectory tracking involving a single output, xD(t), and a single input, r(t), is implemented on a run-to-run

basis via ILC. The initial input trajectory used for ILC is linear with r(ts) = 0.898 and r(tf) = 0.877. The

residual tracking error cannot be reduced to zero for all times because of the non-zero tracking error at

time tsarising from uncertainties in the start-up phase, that is, xD(ts)6=xD,r ef (ts). As a consequence, the

ILC schemes that try to enforce zero residual tracking error do not converge in this case. Instead, non-zero

tracking error has to be tolerated to enforce convergence. This can be accomplished by applying a forgetting

factor or a time shift to the feedforward trajectory. Three ILC schemes without current-cycle feedback are

considered, with N= 100 time samples:

•β= 0.999. A forgetting factor βis applied to the feedforward input trajectory in Eq. (26) with ¯

A=β I

and ¯

B=K.

•δu= 0.25 h. A small shift is applied to the feedforward input trajectory in Eq. (29).

•δu= 1 h, δe= 1 h. The same large shift is applied to the feedforward input and the error trajectory in

Eq. (29).

The proportional gain K= 0.1Iis determined as a compromise between robustness and perfor-

mance. ILC with forgetting factor converges after 30 runs, while the schemes with time shift of the trajec-

tories converge after 25 runs as illustrated in Figure 13. The ﬁnal tracking error is slightly smaller with the

40

latter methods, especially when the time shift of the feedforward trajectory is reduced to δu= 0.25 has

shown in Table 3. Figure 14 shows the tracking performance of ILC with input shift after 20 runs. Note

that, since it is necessary to have distillate in the product tank to be able to take measurements, tracking

starts one sampling time after the start of Phase 2, that is, at at ts+h.

1 5 10 15 20 25 30

0.85

0.9

0.95

1

1.05

1.1

1.15

1.2

Batch index k

β = 0.999

δu = 0.25 h

δu = δ

e = 1 h

Squared tracking error

Figure 13: Evolution of the squared tracking error for three ILC schemes.

0 2 4 6 8 10

0.8

0.85

0.9

0.95

1

Time t [h]

xD

Reference

δu = 0.25 h

tsts+h tf

Figure 14: Tracking performance using ILC with input shift.

41

Table 3: Comparison of the three tracking schemes after 30 runs in terms of the squared tracking error

Ptf

tse2(t), the tracking error at ﬁnal time e(tf), and its variance ve(tf).

Strategy Ptf

tse2(t)|e(tf)| × 103ve(tf)×105

β= 0.999 0.8738 6.5 4.1

δu=0.25h0.8745 0.5 3.0

δu=1h, δe= 1h0.8833 2.2 3.5

42

PART B. OPTIMIZATION

Process optimization is the method of choice for improving the performance of chemical processes while

enforcing operational constraints63. Long considered as an appealing tool but only applicable to academic

problems, optimization has now become a viable technology64,65. Still, one of the strengths of optimization,

namely, its inherent mathematical rigor, can also be perceived as a weakness, since engineers might some-

times ﬁnd it diﬃcult to obtain an appropriate mathematical formulation to solve their practical problems.

Furthermore, even when process models are available, the presence of plant-model mismatch and process

disturbances makes the direct use of model-based optimal inputs hazardous.

In the last 30 years, the ﬁeld of real-time optimization (RTO) has emerged to help overcome the

aforementioned modeling diﬃculties. RTO integrates process measurements into the optimization framework.

This way, process optimization does not rely exclusively on a (possibly inaccurate) process model but also on

process information stemming from measurements. The ﬁrst widely available RTO approach was the two-

step approach that adapts the model parameters on the basis of diﬀerences between predicted and measured

outputs and uses the updated process model to re-compute the optimal inputs66,67. However, in the presence

of structural plant-model mismatch, this method is very unlikely to drive the plant to optimality68,69 . Hence,

alternatives to the two-step approach have recently been developed. For example, the modiﬁer-adaptation

scheme also proposes to solve a model-based optimization problem, but using a plant model with ﬁxed

model parameters 69 . Correction for uncertainty is made via modiﬁer terms that are added to the cost and

constraint functions of the optimization problem. As the modiﬁers include information on the diﬀerences

between the predicted and the plant necessary conditions of optimality (NCO), this approach is prone to

reach the plant optimum upon convergence. Yet another ﬁeld has emerged, for which numerical optimization

is not used online. With the so-called self-optimizing approaches70–72 , the optimization problem is recast

as a control problem that uses measurements to enforce certain optimality features for the plant. However,

note that self-optimizing control70 and extremum-seeking control71, which have been developed to optimize

continuous plants, are not suited to the optimization of batch processes.

43

The optimization of a batch process falls naturally in the category of dynamic optimization, that is,

with inﬁnite-dimensional time proﬁles as decision variables. However, the introduction of input parameteri-

zation can transform the dynamic optimization problem into a static optimization problem with a countable

number of decision variables, a so-called nonlinear program (NLP).

This optimization part is organized as follows. Various aspects of numerical optimization for both

static and dynamic optimizations are presented ﬁrst, followed by a discussion of two RTO classes, namely, re-

peated numerical optimization and optimizing control. The theoretical developments will then be illustrated

by three case studies.

8 Numerical Optimization

Apart from very speciﬁc cases, the standard way of solving an optimization problem is via numerical opti-

mization, for which a model of the process is required. As already mentioned, the optimization of a batch

process can be formulated as a dynamic optimization problem. In this chapter, we will consider the following

constrained dynamic optimization problem:

min

u[0,tf]J:= φx(tf)(57)

s.t. ˙x(t) = Fx(t), u(t), x(0) = x0,(58)

Sx(t), u(t)≤0,(59)

Tx(tf)≤0,(60)

where φis the terminal-time cost functional to be minimized, x(t) the n-dimensional vector of state proﬁles

with the known initial conditions x0,u(t) the m-dimensional vector of input proﬁles, Sthe nS-dimensional

vector of path constraints that hold during the interval [0, tf], Tthe nT-dimensional vector of terminal

constraints that hold at tf, and tfthe ﬁnal time, which can be either free or ﬁxed†. The optimization

†If tfis free, it becomes a decision variable. Note that it is also possible to have additional constant decision variables. In this

case, the decision variables would include the vector of input proﬁles u[0, tf] and the vector of constant quantities, ρ. This

general case will not be considered here for simplicity of exposition, see72 for details.

44

problem (57)-(60) is said to be in the Mayer form, that is, Jis a terminal-time cost. When an integral cost

is added to φ, the corresponding problem is said to be in the Bolza form, while when it only incorporates

the integral cost, it is referred to as being in the Lagrange form. Note that these three formulations can be

made equivalent by the introduction of additional states 73.

The nonlinear dynamic optimization problem (57)-(60) is diﬃcult to solve due to the time depen-

dency of the various variables. In particular, the inputs u(t) represent an inﬁnite-dimensional decision

vector. This time dependency can be removed via time discretization using, for example, the method of

orthogonal collocation74, which however results in a large vector of decision variables. Another approach

consists in viewing the batch process as a static map between a small number of input parameters πand a

small number of run-end outputs zas discussed in Section 3.3. These approximations reduce the dynamic

optimization problem into a static problem. Next, the following problems will be addressed successively:

dynamic optimization, reformulation of a dynamic optimization problem as a static optimization problem,

static optimization, and eﬀect of uncertainty.

8.1 Dynamic Optimization

Consider the optimization problem (57)-(60), for which we will present the necessary conditions of optimality.

8.1.1 Pontryagin’s minimum principle

The NCO for a dynamic optimization problem are given by Pontryagin’s minimum principle (PMP). The

application of PMP provides insight in the optimal solution and generates explicit conditions for meeting

active path and terminal constraints and forcing certain sensitivities to zero.

Let us deﬁne the Hamiltonian function H(t),

H(t) = λT(t)Fx(t), u(t)+µT(t)Sx(t), u(t),(61)

45

and the augmented terminal cost

Φ(tf) = φx(tf)+νTTx(tf),(62)

where λT(t) are the adjoint variables such that

˙

λT(t) = −∂H

∂x (t), λT(tf) = ∂Φ

∂x(tf),(63)

µ(t)≥0 are the Lagrange multipliers associated with the path constraints, and ν≥0 are the Lagrange

multipliers associated with the terminal constraints. In dynamic optimization problems, there are both

path and terminal objectives to satisfy, with these objectives being either constraints or sensitivities. PMP

and the introduction of the Hamiltonian function H(t) allow (i) assigning the NCO separately to path and

terminal objectives, and (ii) expressing them in terms of meeting constraints and sensitivities, as illustrated

in Table 4. The solution is generally discontinuous and consists of several intervals or arcs73,75. Each arc is

characterized by a diﬀerent set of active path constraints, that is, this set changes between successive arcs.

Table 4: NCO for a dynamic optimization problem 73 .

Path Terminal

Constraints µT(t)Sx(t), u(t)= 0, µ(t)≥0νTTx(tf)= 0, ν ≥0

Sensitivities ∂H

∂u (t) = 0 – ‡

8.1.2 Solution methods

Solving the dynamic optimization problem (57)-(60) corresponds to ﬁnding the best input proﬁles u[0, tf]

such that the cost functional is minimized, while meeting both the path and terminal constraints. Since the

decision variables u[0, tf] are inﬁnite dimensional, the inputs need to be parameterized using a ﬁnite set of

parameters to be able to use numerical techniques. These techniques can be classiﬁed in two main categories

‡If the decision variables include the constant vector ρ, there will be terminal sensitivity elements for the determination of ρ,

see72 for details.

46

according to the underlying formulation73, namely, direct optimization methods that solve the optimization

problem (57)-(60) directly, and PMP-based methods that attempt to satisfy the NCO given in Table 4.

Direct optimization methods are distinguished further depending on whether the system equations

are integrated explicitly or not 73 . In the sequential approach, the system equations are integrated explicitly,

and the optimization is carried out in the space of the decision variables only. This corresponds to a

“feasible-path” approach since the diﬀerential equations are satisﬁed at each step of the optimization. A

piecewise-constant or piecewise-polynomial approximation of the inputs is often used. The computationally

expensive part of the sequential approach is the accurate integration of the system equations, which needs

to be performed even when the decision variables are far from the optimal solution!

In the simultaneous approach, an approximation of the system equations is introduced to avoid explicit

integration for each candidate set of input proﬁles, thereby reducing the computational burden. Since

the optimization is carried out in the full space of discretized inputs and states, the diﬀerential equations

are satisﬁed only at the solution 74,76 . This is therefore called an “infeasible-path” approach. The direct

approaches are by far the most commonly used. However, note that the input parameterization is often

chosen arbitrarily by the user, which can aﬀect both the eﬃciency and the accuracy of the approach77 .

On the other hand, PMP-based methods try to satisfy the ﬁrst-order NCO given in Table 4. The

NCO involve the state and adjoint variables, which need to be computed via integration. The diﬀerential

equation system is a two-point boundary-value problem since initial conditions are available for the states

and terminal conditions for the adjoints. The optimal inputs can be expressed analytically from the NCO

in terms of the states and the adjoints, that is, u∗[0, tf] = Ux[0, tf], λ[0, tf]. The resulting diﬀerential-

algebraic system of equations can be solved using a shooting approach78, that is, the decision variables

include the initial conditions λ(0) that are chosen in order to satisfy λ(tf).

47

8.2 Reformulation of a dynamic optimization problem as a static optimization

problem

Consider the optimization problem (57)-(60). Numerically, this problem can be solved by control vector

parameterization (CVP) 76 , whereby the inﬁnite-dimensional inputs u(t) are parameterized using a ﬁnite

number of parameters π. Typically, the inputs are divided into small intervals, and a polynomial approxima-

tion of the inputs is used within each interval. The input parameter vector πcan also include the switching

times between the intervals. It follows that the inputs can be expressed as u(t) = U(π, t) and the optimization

performed with respect to π, where πis the nπ-dimensional vector of constant decision variables.

The diﬃculty with such a parameterization is that the path constraints S(x, u) have to be discretized

as well, resulting in as many constraints as there are discretization points (typically a large number). However,

if structural information regarding the shape of the optimal solution is available, the path constraints can

be handled diﬀerently using an alternative parameterization as explained next.

The solution to a constrained terminal-time dynamic optimization problem is typically discontinuous

and consists of various arcs. For a given input, the ith arc ηi(t) can be of two types:

1. ηi,path : Arc ηi(t) is determined by an active path constraint (constraint-seeking arc),

2. ηi,sens : Arc ηi(t) is not governed by an active path constraint, it is sensitivity seeking.

Let us assume that the path constraints are met using either (i) assignment of input values in

the case of active input bounds, or (ii) online feedback control via the measurement or estimation of the

constrained quantities and their regulation to track the constraint values. The other parts of the inputs are

parameterized as in CVP as u(t) = U(π, t). Since the path constraints are handled directly, they do not

involve the decision variables π. Note that the knowledge of the active path constraints in various intervals

is crucial for this parameterization. The ﬁnal states, which aﬀect the objective function and the terminal

constraints, can be expressed in terms of the decision variables πas follows:

x(tf) = x0+Ztf

0

Fx(t),U(π, t)dt := F(π),(64)

48

as tfis an element of π. Hence, the dynamic optimization problem (57)-(60) can be reformulated as:

min

πφF(π)(65)

s.t. TF(π)≤0,(66)

since the system dynamics are included in Fand the path constraints are taken care of as mentioned above.

It follows that the dynamic optimization problem can be reformulated as the following static NLP:

min

πJ:= Φ(π) (67)

s.t. G(π)≤0.(68)

The scalar cost function to be minimized is written Φ(π) := φF(π), and the ng-dimensional vector of

constraints is written generically G(π) := TF(π).

Some remarks are necessary at this point:

•This reformulation is a mathematical way of viewing the dynamic process as a static input-output

“π-(Φ, G)” map. However, the process remains dynamic, with the terminal cost and the terminal

constraints being aﬀected by the process dynamics.

•This reformulation is clearly an approximation since it relies on: (i) a parameterization of the inputs,

and (ii) approximate tracking of the state constraints.

•NLP (67)-(68) is a complex problem if one wants to predict Φ and Gfrom πsince the system dynamics

need to be integrated according to Eq. (64). However, the situation is much simpler in the experimental

context, as one simply apply the inputs πto the plant and measure both Φ and G. In this case, the

model used to predict Φ and Gis replaced by the plant itself. This is a perfect example of the power

of measurement-based optimization compared to numerical optimization, as will be detailed later.

8.3 Static Optimization

The optimization problem is of algebraic nature with a ﬁnite number of constant decision variables.

49

8.3.1 KKT necessary conditions of optimality

With the formulation (67)-(68) and the assumption that the cost function Φ and the constraint functions G

are diﬀerentiable, the Karush-Kuhn-Tucker (KKT) conditions read 79:

G(π∗)≤0,(69)

∇Φ(π∗) + (ν∗)T∇G(π∗) = 0,(70)

ν∗≥0,(71)

(ν∗)TG(π∗) = 0,(72)

where π∗denotes the solution, ν∗the ng-dimensional vector of Lagrange multipliers associated with the

constraints, ∇Φ(π∗) the nπ-dimensional row vector denoting the cost gradient evaluated at π∗, and ∇G(π∗)

the (ng×nπ)-dimensional Jacobian matrix computed at π∗. For these equations to be necessary conditions,

π∗needs to be a regular point for the constraints, which calls for linear independence of the active constraints,

that is, rank{∇Ga(π∗)}=ng,a , where Garepresents the set of active constraints, whose cardinality is ng ,a.

Condition (69) is the primal feasibility condition, Condition (71) the dual feasibility condition, and

Condition (72) the complementarity slackness condition. The stationarity condition (70) indicates that, at

the solution, collinearity between the cost gradient and a linear combination of the gradients of the active

constraints prevents from ﬁnding a search direction that would result in cost reduction while still keeping

the constraints satisﬁed.

8.3.2 Solution methods

Static optimization can be solved by state-of-the-art nonlinear programming techniques. In the presence

of constraints, the three most popular approaches are 80: (i) penalty-function methods, (ii) interior-point

methods, and (iii) sequential quadratic programming. The main idea in penalty-function methods is to

replace the solution to a constrained optimization problem by the solution to a sequence of unconstrained

optimization problems. This is made possible by incorporating the constraints in the ob jective function via

50

a penalty term, which penalizes any violation of the constraints, while guaranteeing that the two problems

share the same solution (by selecting suﬃciently large weights). Interior-point methods also incorporate the

constraints in the objective function 81 . However, the constraints are approached from the feasible region,

and the additive terms increase to become inﬁnitely large at the value of the constraints, thereby acting more

like a barrier than a penalty term. A clear advantage of interior-point methods is that feasible iterates are

generated, while for penalty function methods, feasibility is only guaranteed upon convergence. Note that a

barrier-penalty function that combines the advantages of both approaches has also been proposed82. Another

way of computing the solution to a static optimization problem is to solve the set of NCO, for example

iteratively using sequential quadratic programming (SQP). SQP methods solve a sequence of optimization

sub-problems, each one minimizing a quadratic approximation to the Lagrangian function L:= Φ + νTG

subject to a linear approximation of the constraints. SQP typically uses Newton’s or quasi-Newton methods

to solve the KKT conditions.

8.4 Eﬀect of Uncertainty

8.4.1 Plant-model mismatch

The model used for optimization consists of a set of equations that represent an abstract view, yet always

a simpliﬁcation, of the real process. Such a model is built based on conservations laws (for mass, numbers

of moles, energy) and constitutive relationships that express kinetics, equilibria and transport phenomena.

The simpliﬁcations that are introduced at the modeling stage to obtain a tractable model aﬀect the quality

of the process model in two ways:

(i) some physical or chemical phenomena are assumed to be negligible and are discarded, and

(ii) some dynamic equations are assumed to be at quasi-steady state.

Hence, the structure of the model that is used diﬀers from the “true” model structure. This di-

chotomy gives rise to the so-called structural plant-model mismatch. Furthermore, the model involves a

51

number of physical parameters, whose values are not known accurately. These parameters are identiﬁed

using process measurements and, consequently, are only known to belong to some interval with a certain

probability. For the sake of simplicity, we will consider thereafter that all modeling uncertainties, although

unknown, can be incorporated in the vector of uncertain parameters θ.

8.4.2 Model adequacy

Uncertainty is detrimental to the quality of both model predictions and optimal solutions. If the model

is not able to predict the process outputs accurately, it will most likely not be able to predict the NCO

correctly. On the other hand, even if the model is able to predict the process outputs accurately, it is often

unable to predict the NCO correctly since it has been trained to predict the outputs and not, for instance,

the gradients that are key constituents of the NCO. Hence, numerical optimization is capable of computing

optimal inputs for the model, but it often fails to reach plant optimality.

The property that ensures that a model-based optimization problem will be able to determine the

optimal inputs for the plant is referred to in the literature as “model adequacy”. For a given model-based

RTO scheme, a model is adequate if the RTO scheme is able to predict the correct set of active plant

constraints and the correct alignment of plant gradients. Model adequacy represents a major challenge

in process optimization since, as discussed earlier, models are trained to predict the plant outputs rather

than the plant NCO. In practice, application of model-based optimal inputs leads to suboptimal, and often

infeasible, operation.

9 Real-time Optimization

In the presence of modeling errors and process disturbances, the control trajectories computed oﬄine lose

their optimal character. One way to reject the eﬀect of uncertainty on the overall performance (with respect

to both optimality and feasibility) is by adequately incorporating process measurements in the optimization

52

framework. This is the ﬁeld of real-time optimization. Measurements can be incorporated in two diﬀerent

ways as illustrated in Figure 15:

a. Adapt the process model and repeat the optimization. At each iteration, the model parameters are

updated and the optimization problem solved numerically to generate u∗[0, tf]. This adaptation and

optimization can also be repeated online, in a single run, to estimate at time tthe current states ˆx(t)

and compute the inputs u∗[t, tf] for the remaining part of the batch.

b. Adapt the inputs through optimizing feedback control. Here, the optimization is implicit since opti-

mality and feasibility are enforced via feedback control to satisfy the necessary conditions of optimality.

The control scheme often involves (i) online elements that generate parts of the inputs, u∗

a(t), and (ii)

run-to-run elements that generate the other part of the inputs, u∗

b[0, tf], via the input parameters π∗.

These RTO schemes will be discussed in the following subsections.

9.1 Repeated Numerical Optimization

As additional information about the process becomes available, either during or at the end of the run,

this information can be used to improve future operations. This will be documented next for both online

implementation using economic MPC and run-to-run implementation via the two-step approach.

9.1.1 Economic MPC (Strategy 1)

The approach is similar to MPC discussed in Section 4.2.1, but for the fact that the cost function is no longer

tailored to tracking run-end references, but rather to minimize an economic cost function. The problem can

53

Economic MPC

1

2

3aOnline NCO tracking

3b

Two-step approach

Optimizing feedback control

(track optimality conditions)

Repeated num. optim.

(of an updated model)

Run-to-run

Online

Implementation

aspect

Optimization via

Run-to-run NCO tracking

u

[0,t

f

]

∗

u

(t), u

[t, t

f

]

∗

∗

u[0, y[0,

tf]

tf]

Id. & Opt. for next run

P

∗

ua

( t ) y(t)

track path objectives

P

∗

u

b

[0, z

tf]

P

∗

track terminal objectives

on a run-to run basis

eMPC

u[t,

tf]

P

∗

E

x(t)

^

y(t)

Figure 15: RTO schemes for batch processes. The schemes are classiﬁed according to whether the opti-

mization is implemented via repeated numerical optimization or feedback control (horizontal division) and

whether it is implemented online or on a run-to-run basis (vertical division). eMPC stands for “economic

model predictive control”. Note that NCO tracking can be implemented using the four control approaches

given in Figure 2 to track the path and terminal objectives either online or on a run-to-run basis. Most often,

the path constraints are implemented online, while the terminal objectives are implemented on a run-to-run

manner (see the similarities with Table 4).

54

be formulated as follows:

min

u[ti,tf]J:= φx(tf)(73)

s.t. ˙x(t) = Fx(t), u(t), x(ti) = ˆx(ti),(74)

Sx(t), u(t)≤0,(75)

Tx(tf)≤0,(76)

x(tf)∈ X ,(77)

where tiis the discrete time instant at which the optimization is performed, ˆx(ti) is the state estimate at that

time, and Xis the bounded region of state space where the ﬁnal state should lie. Numerical optimization

yields the control sequence, u∗[ti, tf], of which only the ﬁrst part, u∗[ti, ti+h], is applied in an open-loop

fashion to the plant. Numerical optimization is then repeated at every sampling instant. The scheme is

illustrated in Figure 16.

Economic

MPC

Batch Process y(t)

State Estimation

x(t)

‹

z

u [t, tf]

*

Figure 16: Economic MPC to optimize a batch process repeatedly online.

Note that re-optimization performed during a run requires a valid process model, including estimates

of the current states. Similarly, correction through state feedback requires knowledge of the current states or

their estimates. Hence, the main engineering challenge in the context of batch processes lies in estimating the

states and the parameters from a few noisy measurements. Although the use of extended Kalman ﬁlters83

or moving-horizon estimation 84 has become increasingly common and successful for continuous processes,

these methods are diﬃcult to use for batch processes due to the large variations that the states typically go

through and the limited time of operation. Good accounts of these diﬃculties are available85,86.

55

The weakness of this method is clearly its reliance on the model; if the model parameters are not

updated, model accuracy plays a crucial role. However, if the model is updated, there is a conﬂict between

parameter estimation and optimization, the so-called dual control problem87, since parameter estimation

requires persistency of excitation, that is, the inputs must be suﬃciently varied to uncover the unknown

parameters, a condition that is usually not satisﬁed when near-optimal inputs are applied. Finally, note that

the scheme is called “economic MPC” because MPC typically addresses the tracking of given trajectories.

Here, there are no optimal trajectories to track, but rather some economic cost function to minimize. This

topic has gained a lot of interest in recent years88 .

9.1.2 Two-step approach (Strategy 2)

In the two-step approach, measurements are used to reﬁne the model, which is then used to optimize the

process 67. The two-step approach has gained popularity over the past thirty years mainly because of its

conceptual simplicity. Yet, the two-step approach is characterized by certain intrinsic diﬃculties that are

often overlooked. In its iterative version, the two-step approach involves two optimization problems, namely,

one each for parameter identiﬁcation and process optimization as shown in Figure 17 and described next:

Identiﬁcation ˆ

θk:= arg min

θkyp,k[0, tf]−yk[0, tf]kl2(P1)

s.t. ˙xk(t) = Fxk(t), u∗

k(t), θ, xk(0) = x0,k ,

yk(t) = Hxk(t), u∗

k(t), θ,

θ∈Θ

Optimization u∗

k+1[0, tf] := arg min

u[0,tf]φxk(tf)(P2)

s.t. ˙xk(t) = Fxk(t), u(t),ˆ

θk, x(0) = x0,k,

S(xk(t), u(t),ˆ

θk)≤0,

Txk(tf),ˆ

θk≤0,

where Θ indicates the set in which the uncertain parameters θare assumed to lie. The subscript (·)pis

used to indicate that the outputs are the plant measurements yp,k [0, tf], by opposition to the values pre-

dicted by the model, yk[0, tf].

56

Identification

Optimization

Plant

Updated Model Parameters

Updated Inputs

Process Performance

Uncertainty

Run Delay

and

OK?

yes

no

Stop

uk [0,tf]

*

θk

ˆ

yp,k [0,tf]

yp,k [0,tf]

Figure 17: Basic idea of the two-step approach with its two optimization problems.

The ﬁrst step identiﬁes best values for the uncertain parameters by minimizing the l2-norm of the

output prediction errors. The updated model is then used for computing the optimal inputs for the next

iteration. Algorithmically, the optimization of the performance of a batch process proceeds as follows:

1. Initialization. Set k= 1 and select an initial guess for the input proﬁles u∗

k[0, tf].

2. Iteration. Apply the inputs u∗

k[0, tf] to the plant. Complete the batch while measuring the run-time

outputs yp,k[0, tf].

3. Compute the distance between the predicted and measured outputs and continue if this distance exceeds

the tolerance, otherwise stop.

4. Solve the identiﬁcation problem (P1) and compute ˆ

θk.

5. Solve the optimization problem (P2) and compute u∗

k+1[0, tf]. Set k:= k+ 1 and repeat Steps 2-5.

The two-step approach suﬀers from two main limitations. First, the identiﬁcation problem requires

suﬃcient excitation, which is however rarely the case since the inputs are computed for optimality rather

57

than for the sake of identiﬁcation. Hence, one has to make sure that there is suﬃcient excitation, for example

via a dual optimization approach that adds a constraint to the optimization problem regarding the accuracy

of the estimated parameters 89 . The second limitation is inherent to the philosophy of the method. Since

the adjustable handles are the model parameters, the method assumes that (i) all the uncertainty (including

process disturbances) can be represented by the set of uncertain parameters, which is rarely the case.

9.2 Optimizing Feedback Control

The second class of RTO methods proposes to adapt the process inputs via feedback control, that is, without

repeating the numerical optimization. Since feedback control is used to enforce the necessary conditions of

optimality of the dynamic optimization problem, the resulting scheme is called “NCO tracking”.

9.2.1 NCO tracking (Strategy 3)

NCO tracking is a feedback control scheme, where the controlled variables (CVs) correspond to measurements

or estimates of the plant NCO, and the manipulated variables (MVs) are appropriate elements of the input

proﬁles. Enforcing the plant NCO is indeed an indirect way of solving the optimization problem for the

plant, along the lines of the PMP-based methods discussed in Section 8.1.2.

Necessary conditions of optimality. The NCO of the dynamic optimization problem (57)-(60) encompass four

parts as shown in Table 4: (i) the path constraints, (ii) the path sensitivities, (iii) the terminal constraints,

and (iv) the terminal sensitivities. There are as many NCO as there are degrees of freedom in the optimization

problem, thus making the system of equations perfectly determined. NCO tracking proposes to enforce the

four NCO parts via tailored feedback control. In particular, some objectives are met using online control,

while others are met via run-to-run control, thereby fully exploiting the versatility of control approaches

for batch processes, as discussed in Sections 4 and 5. The NCO-tracking scheme is therefore a two-level

(online and run-to-run) multivariable feedback control problem, as illustrated in Figure 18. The design of

the NCO-tracking controller is supported by the concept of “solution model” 72.

58

Control

Design

Numerical

Optimization

PROCESS zua(t)

model

solution

model

process

NCO-tracking

controller

Real-time (online & run-to-run)

Offline

Characterization solution

optimal

ub[0,tf]y(t) y[0,tf]

NCO-tracking

Controller

*

*

Figure 18: NCO-tracking scheme. The oﬄine activities include the generation of a solution model and the

design of a NCO-tracking controller. The real-time activities involve online and run-to-run control to enforce

the plant NCO; u∗

a(t) and u∗

b[0, tf] are the input parts that are generated online and on a run-to-run basis,

respectively.

Solution model. The solution model is a qualitative model of the optimal solution, which includes (i) the

types and sequence of arcs, thus assuming no change in active constraints, (ii) the degrees of freedom (MVs)

that one would like to use for control, and (iii) the corresponding NCO (CVs). It is important to understand

that the solution model is simply a tool that helps solve the problem at hand in an eﬃcient way. Although

there may be an exact solution model for the process model at hand, this is certainly not the case for the

unknown plant. Hence, the designer will be able to propose alternative (from simple to more complex)

solution models to tackle the design of the multivariable control problem. This aspect, which represents one

of the strengths of NCO tracking, will be detailed in the ﬁrst case study.

The development of a solution model involves three main steps:

1. Characterize the optimal solution in terms of the types and sequence of arcs by performing numerical

optimization using the best available plant model. This plant model need not be very accurate, but it

ought to provide the correct types and sequence of arcs. One typically performs a robustness analysis

59

to ensure that the qualitative solution remains structurally valid in presence of uncertainty.

2. Select a ﬁnite set of input arcs and parameters to represent (or approximate) the input proﬁles, and

formulate the NCO for this choice of degrees of freedom. Note that the NCO will change with the

choice of the degrees of freedom.

3. Pair the MVs and the NCO to form a multivariable control problem.

Steps 2 and 3 might require iterations since they greatly aﬀect the implementation that is discussed next.

Implementation aspects. In its general form, NCO tracking involves both online and run-to-run control,

as determined by the solution model and the resulting MV-CV pairing. Some NCO elements are typically

implemented online, while others are easier on a run-to-run basis. For example, a path constraint is easily

enforced online via constraint control, while both terminal constraints and terminal sensitivities are easier

to meet iteratively over several runs. The decision regarding which NCO elements to implement online and

which on a run-to-run basis depends on the nature of the various arcs. For a given input, the various arcs

can be of two types:

(i) A constraint-seeking input arc is associated with a path constraint being active in a given time interval.

If the path constraint is an input bound, the input is simply set at the bound, while in the case of a

state constraint, feedback control can be used to track the value of the constrained quantity.

(ii) A sensitivity-seeking input arc requires ∂H

∂u (t) = 0, which is diﬃcult to implement as such since H(t)

is a function of the adjoint variables λ(t). Hence, one tries to approximate the sensitivity-seeking

arc using a parsimonious input parameterization with only a few constant parameters. This way, the

number of degrees of freedom are no longer the inﬁnite-dimensional u(t), but rather the few constant

input parameters. Furthermore, with the input parameterization, the batch process is viewed as the

static map z=Mπgiven by Eq. (7). Consequently, the path condition ∂H

∂u (t) = 0 is replaced by

terminal conditions of the type ∂ φ(x(tf))

∂π = 0, which can be enforced on a run-to-run basis. Note that

the eﬀect of the approximations introduced at the solution level can be assessed in terms of optimality

loss.

60

In summary, the ease of implementation and therefore the success of NCO tracking depends on the

quality of the approximations that are introduced to generate the solution model (choice of MVs, correspond-

ing NCO, and pairing MV-NCO). These are true engineering decisions that are made with plant optimality

in mind!

10 Optimization Applications

10.1 Semi-batch Reactor with Safety and Selectivity Constraints

A simple semi-batch reactor with jacket cooling is considered to illustrate the NCO-tracking approach and,

in particular, the generation of alternative solution models72.

10.1.1 Problem formulation

•Reaction system: A+B→C, 2B→D, isothermal, exothermic reactions.

•Objective: Maximize the amount of Cat a given ﬁnal time.

•Manipulated input: Feedrate of B.

•Path constraints: Input bounds; heat-removal constraint expressed as a lower bound on the cooling

jacket temperature.

•Terminal constraint: Upper bound on the amount of Dat ﬁnal time.

61

Model equations. Assuming perfect control of the reactor temperature through adjustment of the cooling