Journal of Process Control 24 (2014) 399414

Contents lists available at ScienceDirect

Journal of Process Control

journal homepage: www.elsevier.com/locate/jprocont

Realization of online optimizing control in an industrial semi-batch

polymerization
Tiago F. Finkler a,b, , Michael Kawohl b , Uwe Piechottka b , Sebastian Engell a
a
b

Process Dynamics and Operations Group, Technical University of Dortmund, Germany

Evonik Industries AG, Germany

a r t i c l e

i n f o

Article history:
Received 16 January 2013
Received in revised form 26 May 2013
Accepted 10 September 2013
Available online 25 December 2013
Keywords:
Model based control
Nonlinear control
Optimizing control
Polymerization control
Semi-batch processes

a b s t r a c t
In this work, the realization of an online optimizing control scheme for an industrial semi-batch polymerization reactor is discussed in detail. The goal of the work is the automatic minimization of the duration
of the batch without violating the tight constraints for the product specication which translate into
stringent temperature control requirements for a highly exothermic reaction. Crucial factors for a successful industrial implementation of the control scheme are the development and the validation of a
process model that is suitable for process optimization purposes and the estimation of unmeasured process states and the online compensation of model uncertainties. Two implementations are proposed, a
direct online optimizing control scheme and a simplied scheme that combines a model-predictive temperature controller and a monomer feed controller that steers the cooling power to a predened value
in a cascaded fashion. We show by simulation results with a validated process model that both schemes
achieve the goals of tight temperature control and reduction of the batch time. The performance of the
NMPC controller is superior, on the other hand the cascaded scheme could be directly implemented into
the DCS of the plant and is in daily operation while the online optimizing scheme requires an additional
computer and is currently in the test phase.
2013 Elsevier Ltd. All rights reserved.

1. Introduction
Linear MPC (Model Predictive Control) was rst proposed and
implemented in the 1970s by Cutler and Richalet. Since then, it
has emerged as the standard solution for multivariate high performance control problems in the chemical industry, especially in the
petro-chemical sector. This technique was successfully applied in
real production units long before it was well understood from the
theoretical point of view. By now linear MPC is quite mature and
it is routinely applied in industry [1]. However, as in the chemical
industry most of the process dynamics are inherently nonlinear,
linear models are often only capable to describe the behavior of
the processes accurately near a xed operating point. Especially
in time-varying processes as batch and semi-batch processes with
tight operating windows, linear MPC may fail to deal provide sufcient control performance. In such cases, the use of nonlinear
models may improve the operation of the process, because the
safety margins due to model inaccuracies can be reduced. Therefore, the use of NMPC (Nonlinear Model Predictive Control) has
been motivated by the growing demands on process economics

Corresponding author at: Evonik Industries AG, Germany. Tel.: +49 6181592087.
E-mail addresses: tiago.nkler@evonik.com, tiago.nkler@bci.tu-dortmund.de
(T.F. Finkler).
0959-1524/$ see front matter 2013 Elsevier Ltd. All rights reserved.
http://dx.doi.org/10.1016/j.jprocont.2013.09.028

under tighter product quality specications and environmental

regulations. While the structure of MPC enables a straightforward
extension by employing nonlinear models in the optimization over
a nite prediction horizon, in contrast to the linear case, the resulting problems are usually non-convex and numerically demanding.
A solid theoretical background has also been developed for NMPC
over the last decades [2,3] and several successful applications in real
systems have been reported [4,5]. But, despite of the effort that has
been made, there are still several obstacles to be overcome for the
implementation of NMPC controllers in real production units, in
particular the need for accurate models or online compensation of
model inaccuracies and the robustness of the online optimization
under signicant uncertainties. Therefore the utilization of NMPC
in industry is still quite limited.
During the last years, it has been realized that the potential
of MPC can go far beyond tracking references and rejecting disturbances. Instead, economic cost functions that are usually only
considered in an upper steady state real time optimization layer
can be optimized within the MPC formulation. This enables the
possibility of updating the operation point or the trajectories to
external factors (like changes in the availability of utilities or in
the cost of the raw materials) online during the process operation,
resulting in what is called DRTO (Dynamic Real Time Optimization) or online optimizing control [6]. Dynamic online optimization
can also be realized by using economics-motivated cost functions

400

T.F. Finkler et al. / Journal of Process Control 24 (2014) 399414

Fig. 1. Scheme of the industrial polymerization process.

or by tracking necessary conditions of optimality [7]. Our NMPC

controller follows this approach by maximizing the monomer feed
online over a nite prediction horizon.
In this work, which is an extended version of [8], the challenge of
implementing an optimizing NMPC controller in a real production
unit is addressed using an industrial semi-batch polymerization
process as a case study. An online optimizing NMPC scheme that
simultaneously optimizes the cooling power and the monomer
dosage is proposed and thoroughly tested in simulations. Because
of a simpler implementation and good robustness properties that
have been shown in [9,10], a simplied batch time minimization
scheme that combines a scalar MPC temperature controller and a
monomer feed rate controller is also investigated. The effectiveness
of both control schemes is illustrated by simulations and by results
taken from the application at the real plant. Crucial factors for a
successful industrial implementation of the NMPC, as the development and validation of a process model that is suitable for online
optimization purposes, the estimation of unmeasured states and
the online compensation of uncertainties, are discussed in detail.
After this introductory section, the remainder of the paper is
organized as follows. In Section 2, the industrial polymerization
process is described, different control schemes that have been proposed in the literature are briey discussed and the need for an
advanced control solution is exposed by means of historical process data. In Sections 3 and 4 issues regarding the development,
calibration and validation of a semi-rigorous process model are
discussed. Aiming at using the model for online optimization of
the batch operation, the issues of state estimation and online compensation of model errors are dealt with in Section 5 and a NMPC
scheme for the online optimization of the process is proposed and
investigated in the Section 6. In Section 7, the improvement of the
currently used control solution is revisited on the basis of the developed process model and the alternative optimizing scheme that
can be easily implemented in the DCS is introduced and real results
from the plant are shown. Conclusions are drawn in Section 8.
2. Process description and motivation
The unit investigated in this work is an industrial semi-batch
polymerization reactor that produces a liquid polymer with a relatively low molecular weight in a catalyzed solution polymerization.
The reactor system, which is schematically shown in Fig. 1, is
composed of the reactor vessel where a solution polymerization
reaction takes place, a mechanical stirrer that keeps the reactant

mixture homogeneous and a cooling circuit that removes the heat

of reaction. The reactor is equipped with four inlet ports through
which solvent, monomer, and two catalyst components can be
independently dosed into the reactor.
As it is usual in the chemical industry, this reactor is used to
produce different commercial polymer grades which, in this case,
differ from each other by the polymer viscosity. Since this property
is mainly determined by the temperature at which the polymerization takes place, a precise temperature control is required to
ensure that the end-product will have an acceptable quality. Currently, the product quality is checked by off-line measurements of
the viscosity of the product. The different products are obtained by
different recipes that can be generally described by the three following steps. Firstly, during the pre-reaction step, the two catalyst
compounds and a certain amount of solvent are inserted into the
reactor. During the feeding or reaction step, monomer and solvent
are continuously fed into the reactor, the polymerization starts and
the reactor temperature is raised to the desired value. Finally, during the holding step, the monomer feed is stopped and the reactant
mixture is kept inside the reactor for a pre-established period such
that a high monomer conversion (low residual monomer content)
is achieved.
The standard operation strategy for this process consists in running the batch with constant monomer feed during the whole
reaction period while a cascade of PID controllers, which is illustrated in Fig. 2, takes care of controlling the reaction temperature to
the desired value. Although this operation strategy has been successfully applied during the last years, several difculties related
to the control of the reaction temperature (especially at the beginning of the reaction period) have been reported by the operators.
In Fig. 3, some real process trends are presented in order to illustrate potential for improving the operation and control of the
process.1
These process trends show that, for both products, a relatively
large temperature peak (which has a negative impact on the product quality) is observed at the beginning of the reaction period.
For Product B, the system can be operated with the maximum
monomer feed rate during the whole batch and the cooling constraint is not active (the opening of the cooling valve is below
50% during almost the whole reaction period). There is a relatively
large temperature peak at the beginning of the reaction period,
but the controller can then bring the reaction temperature back
to the setpoint without intervention. For Product A, the cooling
constraint becomes active at the beginning of the feeding period.
When the temperature peak occurs, the maximum cooling power
is reached and the operators are forced to reduce the monomer
feed to bring the system under control. After the intervention by
the operator, the system is operated close to the limit of the cooling capacity and during most of the reaction period the quality of
the temperature control is not good. An alternative for eliminating the temperature peak is to feed the monomer more slowly into
the reactor. This could improve the product quality but would also
increase the batch duration, which is not desired. The problem of
nding the right compromise between the quality of the temperature control and the batch duration, i.e. to choose the velocity
at which the reactants are fed into the reactor, is a well-known
problem in the operation of semi-batch reactors. Several possible
control schemes to handle this challenge can be found in the literature. For example, [8,11,12] investigate the idea of maximizing
the monomer conversion by tracking optimal trajectories that were
computed ofine by manipulating the reactant feed and the jacket
temperature. In semi-batch emulsion polymerizations, usually a

Note that, for condentiality reasons, all the values are scaled.

T.F. Finkler et al. / Journal of Process Control 24 (2014) 399414

401

Fig. 2. Cascade control structure.

Fig. 3. Temperature trends from the real plant.

decomposition approach is applied, in which numerical optimization based on a rigorous model of the process is performed off-line
to calculate optimal trajectories. Some authors, e.g. [13,14], suggest
the use of feedback controllers to track these trajectories online. In
[15], the overall heat of reaction, the conversion and also the concentrations of the monomers in the particle phase are determined
by calorimetry with help from a process model. Master curves
are calculated off-line for the total-conversion-dependent amounts
of monomer and chain-transfer agent (CTA) that will produce a
desired molecular weight distribution. The principle of computing
the optimal feed trajectories ofine can be used to minimize the
batch duration as well. However, as industrial chemical processes
are subjected to a considerable number of uncertainties, an ofine
optimization of the recipe has to take large safety margins into
account in order to guarantee a robust operation. Therefore, several
other contributions focused on adaptive optimizing schemes that
react online against disturbances so that the process can be operated closer to the limits of its operations. For instance, in [16,17],
strategies for the time-optimal operation of emulsion polymerizations that combine model-based state estimation and PI control
have been proposed. In [18], a temperature control scheme for
semi-batch polymerization reactors has been proposed that uses
atness-based feedforward and PID controllers. This scheme was
also tested successfully at an industrial reactor. The monomer feed
rate is taken into account in the feedforward controller but it is not
optimized nor adapted. Constraints are only taken into account by
the choice of design parameters. A feed-rate maximizing control
scheme for semi-batch processes which only uses PI controllers
was introduced in [9,10] and was shown to be remarkably robust
with only small losses of performance compared to a full modelbased online optimization. This scheme has also employed in this
industrial case study.
Few of the mentioned approaches were tested experimentally,
and even fewer have been successfully implemented in industry.

The development of an optimizing control scheme that is capable

of providing good temperature control and of operating the batch
process with maximum productivity under the inuence of several uncertain factors, from cooling water and outside temperature
to monomer and solvent purity and catalyst efciency variations
and fouling of the reactor is still a challenging task. This challenge
is addressed here by two different optimizing control schemes.
The NMPC scheme simultaneously optimizes the cooling power
and the monomer dosage so that the batch time is minimized by
solving moving horizon optimization problems over a nite horizon. It automatically maximizes the monomer dosage so that the
system is operated as close as possible to the limit of the capacity of the cooling system without violating the tight constraints
for the reaction temperature. For the case where the cooling constraint is not active, i.e. when Product B is produced, the system
is operated with maximal monomer feed rate during the whole
reaction period and the cooling power is used to control the temperature. The model used in the NMPC scheme is adapted online,
the correction terms and the process states are estimated by an
extended Kalman lter. The second scheme combines a PFC MPC
controller [19] for the reactor temperature with a PID controller
that adjusts the monomer feed rate such that a predened cooling power level is commanded by the temperature control loop
[9,10]. This scheme implicitly maximizes the monomer feed without using a process model or state estimation during the online
operation.

3. Process modeling
In this work, two different process models are employed, a full
model and a reduced model. The complete model was developed in
order to represent the real plant in simulation studies. It is based on
a detailed description of the polymerization kinetics and it is used

402

T.F. Finkler et al. / Journal of Process Control 24 (2014) 399414

to check the robustness of the developed NMPC controller against

uncertainties in the model and uncertainties in the operation scenarios and to predict the viscosity of the nal product. The reduced
model was developed for the online optimization of the process
operation. It can be simulated much faster than the complete model
and contains correction terms that are updated online in order to
compensate model uncertainties using the available measurements
in an extended Kalman lter scheme.

dMCATB
dt
dMP

1,0

dt

= FCATB ri

(10)

= ri rp1,0 rr,0 +

dMP

2,0

= rr,0 rp2,0 +

The polymerization investigated here is a live polymerization

reaction where there is no termination reaction and the only chainbreaking reaction is the chain-transfer to monomer. The reaction
mechanism is based on the assumption that two different types of
active complexes, I and II, are involved in the reaction. The active
complex of type I is formed when the catalyst complex is rst combined with monomer. It is highly reactive but not stable and it is
gradually converted into the active complex of type II that is stable but less reactive. Both the active complexes of types I and II
react with monomer to form live polymer molecules that grow by
the propagation reaction. By letting the variables M, CATA and CATB ,
P1,0* and P2,0* , P1,i* and P2,i* and Pi denote the monomer, the catalyst
compounds of types A and B, the active sites of types I and II, the
live polymer of type I and II molecules with chain length equal to
i and the dead polymer molecules with chain length equal to i, the
polymerization mechanism is described by Eqs. (1)(6):
ki

CATA + M + CATB P1,0

i = 1, 2, . . .,

dMP

1,i

dt
dMP

2,i

dt
dMPd,i
dt

kp1

kp2

P2,i
+ M P2,i+1
,

(12)

(13)

= rp2,i1 rp2,i + rr,i rfm2,i ,

i = 1, 2, . . ., N

(14)

N


N


rfm1,i +

KA(TJ TR ) +
dTR
=
dt

i = 1, 2, . . ., N

rfm2,i ,

F Cpi (Ti
i=S,M i

TR ) +

N

dTJin

(4)

(rp1,i + rp2,i ) HR

(16)

(2)

i = 1, 2, . . .,

i=1

CpR MR

KA(TR TJ ) + FJ CpJ (TJin TJout )

dTJout
=
CpJ MJ
dt

(3)

(15)

i=1

(1)

i = 1, 2, . . .,

Fcool ()(Tcool TJout ) + FJ (TJout TJin )

Mmix

(17)

(18)

where MM and MS are the monomer holdup and solvent holdup,

MCATA , and MCATB are the holdups of the two catalyst compounds
A and B, MP and MP are the holdups of the free active sites
1,0

2,0

of types I and II, MP and MP are the holdups of living polymer

kfm

P1,i
+ M Pi + P1,0
,
kfm

P2,i
+ M Pi + P2,0
,

i = 1, 2, . . .,

(5)

i = 1, 2, . . .,

(6)

Based on this reaction mechanism, the complete process model

was developed by setting up appropriate material and energy balances around the reactor and the cooling circuit. In the complete
model, the polymer molecules of different sizes are independently
balanced.2 The model is composed of Eqs. (7)(18):


dMM
= FM ri
(rp1,i + rp2,i + rfm1,i + rfm2,i )
dt
N

(7)

i=1

dMs
= Fs
dt

dt

rfm2,i

i = 1, 2, . . ., N

1,i

dMCATA

N


= rp1,i1 rp1,i rr,i rfm1,i ,

i=1

dt

P1,i
+ M P1,i+1
,

(11)

i=1

3.1. Complete model

kr

rfm1,i

i=1

dt

P1,i
,
P2,i

N


= FCATA ri

(8)

(9)

2
Theoretically this leads to an innitely large number of balance equations. However, for practical purposes, as the average degree of polymerization of this product
is small, the amount polymer molecules with more than one thousand monomer
units can be neglected.

2,i

molecules of types I and II with chain length equal to i, MPd,i is the

holdups of dead polymer molecules with chain length equal to i, TR ,
Tcool , TJin , TJout , TJ , TM and TS are the inner reactor temperature, the
brine temperature at the cooling circuit inlet, the brine temperature
at the jacket inlet, the brine temperature at the jacket outlet, the
average jacket temperature between jacket inlet and outlet, the
monomer inlet temperature and the solvent inlet temperature, ri
is the rate of the initiation reaction, rr,i is the rate of the conversion
of active species of type I into active species of type II for the live
polymer molecules with length equal to i, rp1,i and rp2,i are the rates
of the propagation reaction for the live polymer molecules of types
I and II and length equal to i, rfm1,i and rfm2,i are the rates of the
chain transfer to monomer reaction for the live polymer molecules
of types I and II and length equal to i, N is the maximum degree of
polymerization that is considered, K is the heat transfer coefcient,
A is the heat transfer area, V is the volume of the reactive mixture
inside the reactor, FM , FS , FCATA and FCATB are the inlet ow rates of
monomer, solvent and catalyst components A and B, respectively,
FJ is the constant ow rate of cooling brine that crosses the jacket,
Fcool is the ow rate of cooling brine that enters the cooling system,
which depends on the position of the cooling valve , CpR is the
thermal capacity of the reactant mixture, MR is the overall mass
of the reactant mixture, CpJ is the thermal capacity of the cooling
brine; MJ is the mass of cooling brine inside the jacket, Mmix is a time
constant that describes the dynamics of the jacket back-mixing,
CpM , CpP , and CpW are the thermal capacities of monomer, polymer
and cooling brine, respectively, HR is the heat of reaction.

T.F. Finkler et al. / Journal of Process Control 24 (2014) 399414

3.2. Reduced model

The reduced model has been developed for predicting the future
behavior of the process during the online optimization. As it will
be discussed in Section 6, the objectives of the optimization are the
maximization of the monomer feed and the best possible tracking of
the desired reaction temperature. The description of the complete
molecular weight distribution is not relevant for these purposes.
Therefore, instead of the complete reaction mechanism described
by Eqs. (2)(6), only a lumped polymerization reaction is considered and, instead of solving separate balance equations for the
polymer molecules of different sizes, only the bulk polymer mass
is balanced, what reduces the model size considerably. Nonetheless, aiming at quality monitoring, it is desired that the operators
receive some information about the average molecular during the
online optimization. Thus, the number and the weight averages of
the molecular weight distribution are predicted along the batch
from the full model using the method of moments technique.3 The
reduced model is composed of Eqs. (19)(33):
dMM
= FM rp
dt

(19)

dMs
= Fs
dt

(20)

dMCATA
dt
dMCATB
dt

= FCATA ri

(21)

= FCATB ri

(22)

dMP
= rp
dt

(23)

dMC
= ri
dt

(24)

KA (TJ TR ) +
dTR
=
dt

F Cpi (Ti
i=S,M i

TR ) + rp HR

CpR MR

KA (TR TJ ) + FJ CpJ (TJin TJout )

dTJout
=
CpJ MJ
dt

dt

d(0 V )
= kfm CM 0
dt

(31)

d(1 V )
= kfm CM 1
dt

(32)

d(2 V )
= kfm CM 2
dt

(33)

where CM is the concentration of monomer, CCATA and CCATB are

the concentrations of catalyst compounds of types A and B, MC*
is the overall mass of active species inside the reactor, MP is the
holdup of the lumped polymer mass, i and i are moments of
order i of the molecular weight distribution of live and dead poly , which represent the
mer chains. The variables rp , KA* and Fcool
corrected polymerization rate, the corrected heat transfer coefcient and the corrected ow rate of cooling brine, are computed
according to Eqs. (34)(36):
rp = rp rp

(34)

KA = KA KA

(35)

Fcool
= (KV )Fcool

(36)

where rp , KA and KV are correction factors that are updated
online so that the uncertainties in the model computations for the
reaction rate rp , for the heat transfer coefcient KA and for the ow
rate of cooling brine Fcool are compensated.

4. Model calibration and validation

(25)

dTJin

403

()(T
Fcool
cool TJout ) + FJ (TJout TJin )

Mmix

(26)

In order to solve the mass and energy balances of both models,

several parameters, input variables and algebraic relations have to
be known. All the ow rates and temperatures of the inlet streams
are measured and the values of the constants FJ , MJ , Mmix , HR and
the thermal capacities of the different substances were provided by
the process engineering team of the industrial unit. But no information was available regarding the unknown vales KA, Fcool as well as
the constants of all the reactions that compose the polymerization
mechanism. In this section, algebraic relations that enable the computation of these quantities as a function of the states of the system
and of the inputs that are applied to the system are determined with
the help of historical process data.

(27)

d(0 V )
= k0 CCATA CCATB CM
dt

(28)

d(1 V )
= k0 CCATA CCATB CM + kp CM 0 + kfm CM (0 1 )
dt

(29)

d(2 V )
= k0 CCATA CCATB CM + kp CM (21 0 ) + kfm CM (0 2 )
dt
(30)

3
It is well known that the number-based average molecular weight (Mn) and
the mass-based average molecular weight (Mw) can be directly computed from the
relevant moments of the molecular weight distribution. The ODEs for computing
the variation of these moments along the batch can be easily obtained by combining the denition of each moment with the kinetic expressions that describe the
polymerization mechanism. For brevity this derivation is not reproduced here.

4.1. Computation of the heat transfer coefcient

The energy ow that is established between the reactor and the
jacket depends on the product between the heat transfer coefcient K and the heat transfer area A. At the beginning of the batch,
while the reactant mixture is being prepared, the product KA is
expected to increase continuously due to the addition of material that leads to an increase of the heat transfer area A. In the
later stages of the batch, on the other hand, the product KA is
expected to decrease due to the formation of polymer. In this work,
an empirical correlation for the computation of KA as a function
of total mass in the reactor and the monomer mass that reects
this dependence is proposed. This empirical correlation is given by
(37):
KA = C1 (MM + MS + MP ) + C2 MP .

(37)

404

T.F. Finkler et al. / Journal of Process Control 24 (2014) 399414

Fig. 4. Estimation of the heat transfer coefcient for Product A.

Fig. 5. Estimation of the heat transfer coefcient for Product B.

By combining (37) with (17) it is easy to show that, after

some algebraic manipulation, the linear expression (38) that relates
the constants C1 and C2 with the process measurements can be
obtained.4
CpJ MJ

dTJout
FJ CpJ (TJin TJout ) = (MM + MS + MP )(TR TJ )C1
dt

+MP (TR TJ )C2

(38)

This equation can then be used to estimate these constants from

historical process data by linear regression. In Figs. 4 and 5, two
histograms with the results of independent estimations for C1 and
C2 based on historical process data of around 130 batches are presented for both products.5 The averages of these estimates are then
taken as the nominal values of C1 and C2 for the proposed correlation.
In order to quantitatively check the predictions provided by the
proposed correlation, one can invert the energy balance around the

Note that the term (MM + MS + MP ) that multiplies C1 corresponds to the overall mass of content inside the reactor, which is continuously measured during the
operation. The polymer mass MP that multiplies the C2 is not measured online. But
value from MP during the batch can be well approximated by a straight line that goes
from zero to the total mass of polymer at the end of the batch, which is measured.
As it can be seen from the results that will be presented later in Fig. 10, the error
associated to this approximation is very small.
5
As the viscosity of both products is different, the parameters C1 and C2 were
estimated separately for each of the products. For sake of brevity, only the estimation
results for Product B are shown here.

jacket (17) and obtain an explicit formula so that the product KA

can directly be computed from the process measurements along a
batch. In Figs. 4 and 5 the result from (37) is compared with the
value of KA computed along several batches as well. One can see
that, after the initial period where the inversion of (17) is known to
be inaccurate,6 the predictions of the proposed correlation match
the process measurements quite well.
This correlation is then employed in both the complete and
the reduced model. When the complete model is used to play
the role of the real processes in simulation studies, the uncertainties on the heat transfer efciency can be simulated by letting C1
and C2 vary within the range shown in Figs. 4 and 5. When the
reduced model is used for online optimization purposes, the correction factor KA is updated to compensate such uncertainties,
as it will be discussed later. As it can be seen from the results
in Figs. 4 and 5, one can expect KA to vary within the range
[0.751.25].
4.2. Computation of the ow rate of the cooling brine
In order to solve the energy balance at the jacket inlet (18), the
relationship between the position of the cooling valve and the ow
rate of fresh cooling brine that is fed at the jacket inlet must be

6
The computation of KA is expected to be quite inaccurate at the beginning of
the batch, when the reactant mixture is being prepared and heated up from room
to reaction temperature. During this initial period, the reaction activity is low and
the difference between jacket temperature and reactor temperature is very small,
what makes the inversion of (17) error-prone.

T.F. Finkler et al. / Journal of Process Control 24 (2014) 399414

Fig. 6. Characteristic curve of the cooling valve.

known. Therefore, the characteristic curve of this valve was determined experimentally with help of an external ow meter. In Fig. 6,
the experimental measurements of the ow rate are plotted against
the valve position and this curve is approximated by a forth order
polynomial, which is taken as the nominal characteristic curve of
the cooling valve.
The nominal characteristic curve is used in both, the complete
and the reduced model. When the complete model is used to simulate the real process, the coefcients of the polynomial can be
changed in order to represent the effect of disturbances that affect
the cooling supply, e.g. pressure oscillations at the inlet of the valve.
When the reduced model is used for the online optimization of the
process operation, the correction factor KV is updated to compensate such disturbances. Such compensation is important, especially
because the reactor investigated in this work shares the cooling
supply with other production units. Therefore relatively large variations of the correction factor may occur in order to provide an
accurate calculation of the available cooling power, depending on
what is going on in the other plants. Based on operation experience,
KV is expected to vary within the range [0.251.0].
4.3. Kinetic relations
The rates of the different reactions involved in the polymerization mechanism are given by Eqs. (39)(44):
ri = ki CCATA CCATB CM V

(39)

rr,i = ki CP V,

(40)

i = 1, 2, . . ., N

1,i

rp1,i = kp1 exp

 Ea 
1

RTR

 Ea 

CP

1,i1

CM V,

i = 1, 2, . . ., N

(41)

405

computed based on the results of a calorimetric analysis of the historical process data, which are presented in Fig. 7. This data was
generated from the explicit formula for computing the reaction rate
from the process measurements along a batch that is obtained from
inverting the energy balance (16) of the reactor. The trajectories
of the reaction rate along more than 200 batches are presented in
Fig. 7. The accuracy of the calorimetric computations is also checked
in the histogram where the monomer conversion computed from
the calorimetric analysis is compared to the laboratory analysis that
is performed at the end of each batch. As it can be seen from this
histogram, the cumulative error along one trajectory is in the order
of 5%.
The tting of the propagation constants was divided into two
steps. In the rst step kp2 and Ea2 were estimated using data from
the end of the batches, where the reactor temperature is practically constant and only reactant species of type II are expected to
be present. Note that in this region all the trajectories for the reaction rate are almost identical. Moreover, as data of two different
products which are produced at different temperatures is available,
kp2 and Ea2 can be independently and precisely estimated in this
rst tting step. The second step of the tting procedure was based
on average trajectories for the reaction rate. The parameters kr , kp1 ,
and Ea1 were adjusted so that the kinetic model ts the average trajectories of both products as well as possible. Here it is important
to remark that the estimation of these constants is highly dependent on the data at the startup of the reaction, where the reactant
species of type I are the dominant ones. In this region, the trajectories for the reaction rate are spread over a relatively large region.
Therefore it is must be taken into account that the estimations of
kr , kp1 , and Ea1 are signicantly less accurate than the estimates of
kp2 and Ea2 .
The reaction constants that lead to the trajectories shown in
Fig. 7 are taken as the nominal reaction constants. When the complete model is used to represent the real process, these nominal
values can be modied to test the robustness of the controller with
respect to modeling inaccuracies. The lumped reaction rate of the
reduced model rp , which is given by Eq. (45), depends only on kp2 ,
Ea2 , TR , CM and the overall concentration of active species inside the
reactor CC . When the reduced model is used in the online optimization, the reaction rate computed from Eq. (45) is multiplied by the
correction factor rp that compensates the model errors. Note that,
in the absence of a correction, i.e. for rp equal to one, the reduced
model behaves as if only active species of type II were present. This
is a convenient choice because, as explained before, this reaction
kinetics can be modeled very well. The correction factor rp is then
used to compensate any uncertainties that cause deviations from
this behavior, e.g. the presence of active species of type I. As it can
be seen from the results in Fig. 7, when the rate of polymerization
computed by the reduced model is corrected one may expect rp
to vary up to a value of 5.0.
rp = kp2 exp

 Ea 
2

RTR

CC CM V

(45)

rfm1,i = kfm CP CM V,

i = 1, 2, . . ., N

(43)

Finally, the reaction constant kfm that determines the speed of

the chain transfer to monomer was estimated based on the laboratory measurements of the polymer average molecular weights for
both products.

rfm2,i = kfm CP CM V,

i = 1, 2, . . ., N

(44)

4.4. Model validation

rp2,i = kp2 exp

RTR

1,i

2,i

CP

2,i1

CM V,

i = 1, 2, . . ., N

(42)

where ki , kr , kp1 , kp2 and kfm are unknown reaction constants that
have to be estimated and CP and CP are the concentrations of
1,i

2,i

polymer molecules of types I and II with length equal to i. As the

initiation step is very fast, it can be assumed as being instantaneous,
i.e. ki is set to innity. The reaction constants for the propagation
steps and for the formation of active species of types I and II were

The full model was intensively checked with respect to its ability to predict the dynamic behavior of the process. By comparing
simulations of batch runs using the cascaded temperature control
structure with the same controller tuning parameters as in the real
plant with historical data of more than 200 batches, it was veried
that the model represents the dynamics of the real plant faithfully.

406

T.F. Finkler et al. / Journal of Process Control 24 (2014) 399414

Fig. 7. Calorimetric estimation of the reaction kinetics.

Fig. 8. Comparison of the model predictions (continuous blue line) with real process measurements (dotted blue line) for typical batch runs. (For interpretation of the
references to color in this gure legend, the reader is referred to the web version of the article.)

In order to illustrate the prediction capability of the model, the

operation scenarios of the real batches that were previously presented in Fig. 3 are reproduced in the closed-loop simulations
showed in Fig. 8. In these simulations, the cascade of PID controllers
is tuned using exactly the same parameters as in the real process.
Apart from the ow rate of cooling brine, which is computed from
the closed-loop simulations, all the other model inputs, e.g. ow
rates and temperatures of the inlet feeds, are taken directly from
the process measurements. As it can be seen from the plots, the
model is able to reproduce the dynamics of the plant very well.
In Fig. 9, the molecular weight distribution of the nal product
is presented and the average molecular weight distribution is compared with the target value of the product specication. As it can be
seen from the gure, the model prediction matches the target value
almost exactly. The variation of the average molecular weights and
the variation of the polydispersity along the batch are also shown
in Fig. 9.

5. State estimation and compensation of model

uncertainties
Since not all the states of the reduced process model can be measured online, the unmeasured states have to be computed in order
to initialize the NMPC controller at every sampling period along a
batch. In this work, this issue is addressed by an EKF extended
Kalman lter [25] that uses the temperature measurements TR ,
TJout and TJin and the energy balances ((25)(27)) to estimate rp , KA*

online during a batch.

and Fcool

Following the methodology described in [20], one can invert

the energy balances ((25)(27)) to build the observability map
described by ((46)(48)).
rp =

(dTR /dt)CpR MR KA (TJ TR )

KA =

Fcool
=

F Cpi (Ti
i=S,M i

HR
(dTJout /dt)CpJ MJ FJ CpJ (TJin TJout )
TR TJ
Mmix (dTJin /dt) FJ (TJout TJin )
Tcool TJout

TR )

(46)

(47)

(48)

/ TJ ,
Eqs. (46)(48) are always dened, except for TR =
Tcool =
/ TJout . Moreover, apart from KA*, all the variables on the right
hand side of (46)(48) are either known or measured. Therefore,
/ TJ ,
according to this observability map, KA* is observable if TR =

Fcool
is observable if Tcool =
/ TJout and rp is observable if KA* is known.
By analyzing the historical data of the process, one veries that,
apart from some short periods at the very beginning of the batch
when the difference between TR and TJ is small, these observability
conditions are always fullled. Therefore, it is possible to estimate

the current value of the variables rp , KA* and Fcool

during a batch and
use these estimates to compute the monomer holdup, the overall
polymer holdup and the moments of the molecular weight distribution by integrating the corresponding Eqs. (19)(33), what is
realized by setting the corresponding elements on the diagonal of
the covariance matrix Q to zero.

T.F. Finkler et al. / Journal of Process Control 24 (2014) 399414

407

Fig. 9. Molecular weight distribution and its evolution for a typical batch run.

Considering that the goal here is to use the reduced model to

predict the future behavior of the process in the online optimiza
at a given
tion, besides of estimating the values of rp , KA* and Fcool
time instant, it is important to know how these values evolve over
the consider prediction horizon in the all possible operation scenarios. In order to adapt the predictions of the nominal model to
the different uncertainties (e.g. reactivity, heat transfer efciency
and cooling capacity) the computation of the correction factors is
here mixed within the EKF computations. This is realized by inserting ((34)(36)) into the energy balances ((25)(27)) and extending
the reduced model with the pseudo states as described by Eqs.
(49)(51):
drp

=0

(49)

dKA
=0
dt

(50)

dKV
=0
dt

(51)

dt

In this representation, the outcome of the EKF provides the computations for the correction factors directly. The estimations for rp ,

KA* and Fcool

can then be obtained based on the computed correction factors with help from Eqs. (34)(36).
The EKF was tuned within the simulation environment that uses
the complete model to represent the real process. A simple tuning
strategy was employed to determine the diagonal elements of the
covariance matrix Q. The elements corresponding to TR , TJin and
TJout are set to one or two orders of magnitude smaller than the
variance of the temperature measurements7 and the elements corresponding to the correction coefcients were adjusted so that the
simulated reaction rate, heat transfer coefcient and cooling power
are well tracked without excessive oscillations.8 The state vector
of the EKF is given in Eq. (52) and the diagonal elements of the
covariance matrix are presented in Eq. (53).9
x = [MM , MS , MCAT1 , MCAT2 , MP , TR , TJout , TJin , rp , KA ,
KV, 0 , 1 , 2 , 0 , 1 , 2 ]

(52)

This means that the measurements are considered more reliable than the model.
The oscillations in the EKF predictions are not desired because they are transmitted to the plant when the loop of the NMPC controller is closed.
9
Note that the elements corresponding to the simulated states are set to zero.
8

diag(Q ) = [0, 0, 0, 0, 0, 2.5 105 , 2.5 102 , 2.5 102 ,

104 , 104 , 106 , 0, 0, 0, 0, 0, 0]

(53)

The extended Kalman lter provides very good estimates of

the system state over the full range of uncertainties exposed in
Figs. 4, 5 and 7. In order to illustrate the effectiveness of the EKF,
, as well as the predictions of
the estimates for rp , KA* and Fcool
monomer consumption, polymer formation and average molecular
weight are presented in Fig. 10. The robustness of the EKF can also
be seen from the robustness analysis that is discussed in Section 6.2.
After several simulation tests, the EKF was integrated with the DCS
of the real process so that online inferences of the non-measured
states became available.
Note that the computation of the correction factors depends on
the simulated process states, e.g. MM and MP , and that the simulation error is integrated within their computations. In practice it is
not an issue because, as it is shown in Fig. 10, it is possible to tune
the EKF such that the simulation error is almost negligible. Furthermore, when compared to the usual representation that uses rp
and KA as pseudo-states [20] the approach of computing the correction factors turned out to be advantageous because of its better
prediction qualities.

6. Online optimizing NMPC

In an NMPC scheme, the closed loop control inputs are computed based on the repeated solution of an open-loop optimal
control problem in which the future behavior of the process is predicted based on a process model. At every sampling period, the
controller is initialized based on the current information on the
state of the process and the discrete control movements that minimize an objective function over a given prediction horizon are
determined. The rst control movement from this optimal discrete
sequence is then applied to the plant and the controls are held
constant until the next measurement becomes available. The controller is reinitialized using the newest process information and
the whole procedure is repeated. It is not the target of this paper
to provide a deep description of the MPC principles. The interested reader can nd several good descriptions in the literature
[21,22].
The reduced process model is used to predict the process behavior in the NMPC computations. As it is not possible to directly
measure all the relevant system states (only reactor and jacket
temperature measurements are available), the controller initialization is done at every sampling period based on the information
provided by the EKF. The correction factors in the model are also
updated every sampling period and they are held constant in the

408

T.F. Finkler et al. / Journal of Process Control 24 (2014) 399414

Fig. 10. EKF estimates for a simulation of a typical batch run. (a) The dotted blue lines and the solid red lines denote the simulated trajectories and the EKF estimates for rp ,
KA and Fcool . The solid green lines show the respective correction factors, i.e. rp , KA and KV. (b) The dotted blue lines and the solid red lines denote the simulated values
and the EKF estimates for MM , MP and Mw . (For interpretation of the references to color in this gure legend, the reader is referred to the web version of the article.)

optimal control computations over the prediction horizon. Two

different implementations of the NMPC were tested, a sequential implementation and a simultaneous implementation [23,24].
Both implementations use the solver SNOPT from the Tomlab
optimization suite. In the sequential implementation, the model is
embedded in the objective function and it is repeatedly integrated
over the prediction horizon in every iteration of the optimization by
an explicit rst order method with constant integration step. In the
simultaneous implementation, the dynamic optimization problem
is formulated so that the system states and the control movements
at each sampling period along the prediction horizon are degrees
of freedom for the optimizer, and, in order to guarantee that the
solution of the optimization respects the process model, the model
equations are incorporated to the optimization problem as additional equality constraints, which are integrated by an implicit rst
order method with constant integration step. Both implementations are capable of solving the problem considerably faster than
real time and, for the prediction horizons considered in this work,
the results provided by both methods are very similar. For safety
reasons, the implementation of the controller at the real industrial
process must consider that eventually the optimization might have
to be terminated before convergence. In case of premature termination, there is no guaranty that the suboptimal solution provided
by the simultaneous implementation will make sense physically
(because if the constraints are not fullled the model equations
might not be satised). The sequential approach on the other hand
will always provide a suboptimal solution that respects the process
model. Therefore the sequential implementation was preferred for
the industrial application.
In the optimization setup that is proposed here, the monomer
dosage and the cooling power have to be optimized simultaneously.
This is here realized by minimizing the objective function given by
(54), which is composed of the weighted summation of the temperature tracking residual errors, a dominant term that maximizes
the monomer feed and a penalty term that penalizes control movements.

J=

NP


ref 2

wTi (TRi TR ) +

NP


i=1

max
wFM (FMi FM
)
i

i=1

The optimization is constrained by pre-dened bounds for the

reaction temperature, by the valve opening limits and by the maximal monomer feed rate, as described by Eqs. (55)(57).
TRmin < TRi < TRmax

(55)

0 < i < 100%

(56)

max
0 < FMi < FM

(57)

Note that, in the absence of plant-model mismatch, the satisfaction of the hard constraint (55) sufces to keep the reaction
temperature within the specication bounds even in the absence
of the quadratic cost for the tracking of the reference temperature
in (54). In our implementation, the quadratic cost is kept because
it brings the reaction temperature the center of the specication
range, providing more margin for the controller to react to fast
disturbances.
When the optimal control problem is solved, the control movements over the control horizon as well as the evaluation of the
objective function and constraints over the prediction horizon are
discretized considering a discretization period of one minute, with
NC = 2 and NP = 5. These relatively short prediction horizons were
chosen based upon simulations, picking the shortest horizons that
gave a good performance in all situations. As the computation of
the solution of the open-loop optimal control problem takes only
around ve seconds, the NMPC computation and the EKF estimation
are run every ten seconds. This short sampling period is very helpful
because it provides the controller with more accurate information
about the state of the real process, what is important for updating
the correction factors in order to compensate model uncertainties.
The NMPC cost parameters were tuned by extensive tests within
the simulation environment that uses the complete model to represent the plant. After several simulation trials the weights wTRi , wFM
i

and wi were set to 400, 100 and 400, respectively.10 This tuning
provides a good compromise between the main control objectives,
i.e. temperature tracking and feed maximization, without excessive
oscillation of the position of the cooling valve. The nonlinear controller was checked with respect to robustness and performance
by simulating several scenarios with plant-model mismatch which
were realized by varying key parameters of the complete model as

NC

i=1

wi (i i1 )

(54)
10
Note however that the weights depend directly on the units of the reactor
temperature and the monomer feed, which are here arbitrarily scaled due to condentiality.

T.F. Finkler et al. / Journal of Process Control 24 (2014) 399414

409

Fig. 11. Simulation of the NMPC controller for Product A (nominal scenario). (a) The solid blue lines are the simulated process measurements. The dotted green line is the
setpoint for TR . The red dotted lines are the specication bounds (1 K). (b) The dotted blue lines and the solid red lines denote the simulated trajectories and the EKF
estimates for rp , KA and Fcool . The solid green lines show the respective correction factors, i.e. rp , KA and KV. (For interpretation of the references to color in this gure
legend, the reader is referred to the web version of the article.)

Fig. 12. Simulation of the NMPC controller for Product B (nominal scenario). (a) The solid blue lines are the simulated process measurements. The dotted green line is the
setpoint for TR . The red dotted lines are the specication bounds (1 K). (b) The dotted blue lines and the solid red lines denote the simulated trajectories and the EKF
estimates for rp , KA and Fcool . The solid green lines show the respective correction factors, i.e. rp , KA and KV. (For interpretation of the references to color in this gure
legend, the reader is referred to the web version of the article.)

kr , kp1 , Ea1 , kp2 , Ea2 , C1 , C2 and the characteristic curve of the cooling valve. The controller was demonstrated to be robust toward a
wide range of uncertainties in both the reaction rate and the heat
transfer coefcient, even under strong variations in the cooling supply. Moreover, the simulations have also shown that the NMPC can
shorten the duration of the batch signicantly and reduces the temperature peak at the early stages of the batch to less than 0.5 K. In
the following, the effectiveness of the controller is illustrated by
simulations.
6.1. Simulation of nominal operation scenarios
A simulation run for Product A where the nominal parameters of
the complete model are used to represent the behavior of the real
plant is shown in Fig. 11. The results show that the NMPC controller
is capable of driving the process very close to its operational limits. At the beginning of the reaction period, monomer is inserted
into the reactor with maximum ow rate and the reactor temperature rapidly stabilizes at the desired value. For approximately
two thirds of the feeding period, the cooling power is adjusted so

that the temperature remains around its setpoint, then the cooling
power constraint is reached and the monomer feed is automatically reduced by the NMPC controller. The EKF estimates are also
reported in Fig. 11. The plots show that the observer compensates
the model uncertainties and follows the trajectories of the reaction
rate and of the heat transfer coefcient almost exactly. rp , KA
and KV converge to their true values, i.e. 1.00. As expected, the correction factor for the reaction rate reaches the value of 1.00 only
in the later stages of the batch, when only active species of type
II are present. Before, the larger values of rp indicate the higher
reactivity due to the presence of the active species of type I.
In Fig. 12, simulation results of a batch run for Product B are
presented. Again, the nominal parameters of the complete model
are used for the representation of the real plant. In this case, the
cooling constraint does not become active and the monomer is
inserted into the reactor with maximal ow rate during practically
the whole feeding period. The results show that the temperature
overshoot at the beginning of the reaction period is considerably
reduced and the quality of the temperature control along the batch
is signicantly improved by the NMPC.

410

T.F. Finkler et al. / Journal of Process Control 24 (2014) 399414

Fig. 13. Simulation of the NMPC for Product A (scenario with plant-model mismatch and cooling failure after half of the feeding period). (a) The solid blue lines are the
simulated process measurements. The dotted green line is the setpoint for TR . The red dotted lines are the specication bounds (1 K). (b) The dotted blue lines and the solid
red lines denote the simulated trajectories and the EKF estimates for rp , KA and Fcool . The solid green lines show the respective correction factors, i.e. rp , KA and KV. (For
interpretation of the references to color in this gure legend, the reader is referred to the web version of the article.)

Fig. 14. Simulation of the NMPC for Product B (scenario with plant-model mismatch and cooling failure after half of the feeding period). (a) The solid blue lines are the
simulated process measurements. The dotted green line is the setpoint for TR . The red dotted lines are the specication bounds (1 K). (b) The dotted blue lines and the solid
red lines denote the simulated trajectories and the EKF estimates for rp , KA and Fcool . The solid green lines show the respective correction factors, i.e. rp , KA and KV. (For
interpretation of the references to color in this gure legend, the reader is referred to the web version of the article.)

6.2. Robustness analysis

In order to illustrate the robustness of the NMPC controller, simulations with considerable plant-model mismatch are presented
here. The model errors are simulated by changing the key parameters of the complete model so that the trajectories of the reaction
rate and the heat transfer coefcient vary by 25% when compared
to nominal ones (corresponding to the nominal values of the key
parameters). Moreover, in order to simulate a strong cooling failure, the cooling capacity is reduced by 50% during the second half
of the feeding period. The results show that the NMPC controller is
robust over the whole uncertainty range.
The simulation of the worst case scenario where the reactivity
is 25% bigger and the efciency of the heat transfer is 25% smaller
as in the nominal case are presented in Fig. 13 for a batch where the
Product A is produced. In this scenario, the cooling power constraint
becomes active at the early stages of the feeding period and the
NMPC adjusts the monomer feed so that the system is operated
on the limit of its cooling capacity. When the cooling failure takes
place, the NMPC reacts quickly by adjusting the monomer feed and
the disturbance is rejected and the temperature of the reactor is

kept around the target value during the whole batch. In Fig. 14,
the same worst case scenario is simulated for a batch where the
Product B is produced. In this case, the NMPC starts driving the
system along the maximal feed bound. When the failure event takes
place, the cooling constraint is activated and the controller keeps on
driving the system along the maximal cooling bound by adjusting
the monomer dosage.
7. Alternative optimizing control structure
A disadvantage of the optimizing nonlinear model-predictive
controller is that it cannot be run on the DCS by which the plant is
controlled but requires an additional computer and software environment for its implementation. This is technically possible, but
not very much liked in industry. DCS are robust systems with error
checking mechanisms, redundancy etc., while additional hard- and
software imply an additional effort for installation and in particular
for long-term maintenance. The two systems have to be synchronized and additional visualizations have to be built. Moreover,
running a nonlinear optimization online inevitably raises concerns
about its stability and robustness. Therefore, and also to have a

T.F. Finkler et al. / Journal of Process Control 24 (2014) 399414

411

Fig. 15. Block diagram of the alternative control scheme.

quick win strategy to convince the plant management and plant

operators, an alternative scheme was investigated as well. This
scheme is based on the simple idea that in order to minimize the
batch time, the cooling capacity should be fully used, with some
safety margin to react to fast disturbances. So the cooling valve position is controlled by the monomer feed in an outer loop while the
temperature controller acts as the inner loop. The scheme is illustrated in Fig. 15. This combination results in an optimizing control
scheme that automatically balances the amount of heat produced
by the reaction with the available cooling capacity of the system.
The main motivation for using this alternative control structure is
to take advantage of the fact that, independent of any external disturbances that may affect the system, e.g. uctuations in the cooling
water or in the monomer temperatures, the position of the cooling
valve will always provide a quantitative measure of how much cooling power is still available. As a result, operation conditions which
are close to optimality are tracked by the additional controller and
the system is operated almost on the limit of its cooling capacity,
signicantly reducing the batch duration.
In [9,10] it has been shown for the well-known benchmark
example of the ChyllaHaase polymerization reactor [26] that even
with simple PID controllers this scheme realizes most of the potential of a full NMPC scheme and is even more robust for model errors
that are not compensated by model adaptation. The same scheme
was demonstrated for an emulsion polymerization in [27].
This optimizing scheme is here applied to a real industrial
process. Aiming at improving the control performance by taking
advantage of the information provided by the process model, the
cascade of PID controllers for the control of the reactor temperature is replaced by a cascade of predictive functional control (PFC)
model predictive controllers. The combination of the cascade of PFC
controllers with an additional PID controller that manipulates the
monomer feed is illustrated in Fig. 15.
In this work, the alternative optimizing scheme is implemented
in a discrete-time fashion. At the beginning of every sampling
period, the next position of the cooling valve is computed by the
cascade of PFC controllers and the monomer feed rate is computed
by the PID controller. The inputs are then applied to the process
and held constant during the next sampling period. In the following the cascade of PFC controllers is described rst. Then the
optimizing scheme is investigated in simulations so that it can be
compared with the full NMPC controller. Finally the effectiveness
of the scheme is illustrated by results from the implementation in
the real plant.

model-predictive controller that makes use of a reference trajectory to specify the desired behavior of the closed loop system. In the
PFC algorithm, as in the standard MPC approach, a nonlinear process model is used to simulate the process over a prediction horizon
and discrete control movements are computed every sampling
period. The control action is computed by determining the input
change that makes the process model match the reference trajectory at given coincidence points over the prediction horizon. For
the special case where only one coincidence point is used and a
rst order reference trajectory is specied, the control movement
can be determined by solving an algebraic system of equations,
i.e. one can compute the input movement that makes the model
match the coincidence point at the end of the prediction horizon
by inverting the model equations and there is no need to solve an
optimization problem. This feature makes it easy to implement PFC
controllers directly within most of the commercial digital control
systems. Therefore PFC has become popular way of realizing MPC
regulators in industry [19,28].
In this work, the internal model of the master PFC controller is
given by Eq. (58), which represents a modied energy balance for
the reactor where the terms accounting for the heat transfer coefcient and for the reaction rate are replaced by the approximations
KAPFC and rPPFC . The value of the reaction rate within the computations of the master PFC controller is approximated as being equal to
the value of the current monomer feed rate that enters the reactor,
i.e. rPPFC is assumed to be equal to FM . Note that this corresponds
to a pseudo-stationary assumption with respect to the amount of
monomer inside the reactor and, as the additional PID controller
in fact uses the monomer feed to control the reactivity of the system, FM provides the PFC controller with valuable information with
respect to the variations of the reaction rate. The heat transfer coefcient KAPFC is approximated by a constant value taken from the
plateaus that can be seen in Figs. 4 and 5. At every sampling period,
the master PFC controller tracks the desired reaction temperature
by applying a step change in the setpoint of the jacket temperature
while the slave PFC controller tracks the desired jacket temperature by applying a step change in the position of the cooling valve.
The step change in the jacket temperature is computed so that the
model prediction for the reactor temperature, which is obtained by
integrating (58), matches a reference trajectory at the end of the
prediction horizon. All the terms on the right hand side of Eq. (58)
are kept constant during a sampling period.
KAPFC (TJ TR ) +
dTR
=
dt

F Cpi (Ti
i=S,M i

TR ) + rPPFC HR

CpR MR
(58)

7.1. Cascade of PFC controllers

The predictive functional control (PFC) method has been
pioneered by Richalet [19]. The PFC algorithm is an effective

The internal model of the slave PFC controller is described by Eq.

(59), which represents a modied energy balance for the jacket.
In the convective term of the modied energy balance (59), the

412

T.F. Finkler et al. / Journal of Process Control 24 (2014) 399414

Fig. 16. Simulations of the alternative control structure for Product A. (a) Nominal scenario: the solid blue lines are the simulated process measurements. The dotted green
lines are the setpoints for TR and  and the red dotted lines are the specication bounds (1 K). (b) Worst case scenario: The solid blue lines are the simulated process
measurements. The dotted green lines are the setpoints for TR and  and the red dotted lines are the specication bounds (1 K). (For interpretation of the references to color
in this gure legend, the reader is referred to the web version of the article.)

temperature of the cooling brine entering the jacket is computed

from a static energy balance performed at the back mixing point
where the fresh coolant entering the system is mixed with the hot
coolant stream that leaves the jacket. Within the PFC computations,
the characteristic curve of the cooling valve is represented by the
function (), which is a linear t to the experimental data presented in Fig. 6. At every sampling period, the slave PFC controller
computes the step change in the position of the cooling valve so that
the model prediction for the jacket temperature, which is obtained
by integrating (59), matches a reference trajectory at the end the
prediction horizon. All the terms on the right hand side of (59) are
kept constant in the computations within a sampling period.
KAPFC (TR TJ ) + CpJ ()(Tcool TJout )
dTJout
=
CpJ MJ
dt

(59)

7.2. Simulation and robustness analysis

The alternative optimizing scheme was tuned based on simulations where the complete model represented the real process. The
desired tracking speed of the PFC loops was regulated by choosing
a proper time constant for the rst-order reference trajectories that
are used in the computation of the control movements (as described
previously). The reference trajectories of the master PFC controller
and of the slave PFC controller have time constants in the order of
a few minutes. The PID controller that manipulates the monomer

feed was tuned in such a way that the variations in the monomer
ow rate enter as a slow disturbance into the cascaded PFC loops.
The setpoint for the valve position set is a tuning parameter that
determines the desired cooling power. In order to provide some
safety margin for the cascaded PFC controllers to reject fast disturbances, e.g. sharp decreases in the cooling supply, set was set to
75%.
In Fig. 16, the alternative control scheme is simulated for the
Product A considering the same nominal and worst case scenarios
that were used for evaluating the NMPC controller in Section 6. The
feeding period is started with maximal monomer feed and afterwards the monomer feed controller is activated. By comparing the
simulations shown in Fig. 16 with the simulation in Figs. 11 and 12
it can be concluded that the quality of the simplied temperature
control scheme is not as good as that of the NMPC controller but still
satisfactory. The batch time with the simplied scheme is longer
which is due to the use of only 75% of the cooling water valve
opening, on the other hand, this provides a safety margin and in a
real implementation the bound for the NMPC controller might also
be reduced using a soft-constraint for this reason. In the nominal
scenario, the system is driven with maximal monomer feed until
the cooling power reaches the desired level, which is specied by
the setpoint for the position of the cooling valve. Then the monomer
feed is adjusted during the remainder of the feeding period so
that the position of the cooling valve remains ne around its setpoint. Note that, as rPPFC is approximated by the measurement FM , a

Fig. 17. Simulations of the alternative control structure for Product B. (a) Nominal scenario: the solid blue lines are the simulated process measurements. The dotted green
lines are the setpoints for TR and  and the red dotted lines are the specication bounds (1 K). (b) Worst case scenario: the solid blue lines are the simulated process
measurements. The dotted green lines are the setpoints for TR and  and the red dotted lines are the specication bounds (1 K). (For interpretation of the references to color
in this gure legend, the reader is referred to the web version of the article.)

T.F. Finkler et al. / Journal of Process Control 24 (2014) 399414

413

Fig. 18. Tests of the alternative optimizing scheme at the real plant rate. (a) Test for Product A: the solid purple line, the solid gray line and the dashed gray line denote the
reactor temperature, the position of the cooling valve and the monomer feed. (b) Test for Product B: the solid purple line, the solid gray line and the dashed gray line denote
the reactor temperature, the position of the cooling valve and the monomer feed. (For interpretation of the references to color in this gure legend, the reader is referred to
the web version of the article.)

relatively large error is introduced in the energy balance (58) when

the monomer feed drops abruptly to zero at the end of the feeding
period. As a result, a temperature peak is observed at this region. For
the worst case scenario, the maximum cooling power is reached at
the beginning of the feeding period. When the cooling constraint is
reached, the feed controller reduced the monomer feed and brings
the system under control. A slight temperature peak is observed
at the beginning of the feeding period, but the reaction temperature is always kept within the 1 K bounds. When the cooling
failure event takes place (at t = 0.5), the monomer feed is automatically adjusted so that the position of the cooling valve remains
at 75%. The differences in the reaction to the disturbance at the
middle of the batch in the worst-case scenario are surprisingly
small.
In Fig. 17, the alternative control scheme is simulated for the
Product B. As before, the simulations consider the same nominal
and worst case scenarios that were used for evaluating the NMPC
controller in Section 6 and the feeding period is started with maximal monomer feed. By comparing the simulations from Fig. 17 with
the ones from Figs. 13 and 14, similar conclusions can be drawn as
for Product A. The control performance is not as good as for the
NMPC controller, but still satisfactory.
7.3. Implementation at the real plant
The alternative optimizing structure was implemented within
the DCS of the real process. In Fig. 18a, the performance of the
alternative optimizing scheme is illustrated by trend lines of a real
batch where the Product A is produced. When compared to the original cascade control structure, the alternative optimizing scheme
improves the quality of the temperature control and reduces the
duration of the feeding period considerably, by around 25%. In
Fig. 18b, results of a real batch where the Product B is produced
are shown. In this case, the cooling constraint does not become
active and the monomer is inserted into the reactor with maximal
ow rate during the whole feeding period. As it can be seen from
the plot, the temperature peak at the beginning of the batch was
considerably reduced (from 3 K to round 0.5 K).
The simplied control scheme achieves a signicant reduction in
the batch time by reducing the feeding period by 25% for Product A,
and for both products it signicantly improves temperature control.

Manual reductions of the feed rate in order to bring the process

back to the operating window are no longer required. The scheme
could be easily implemented in the DCS of the plant and is in daily
operation. A further advantage is that the monomer feed controller
can be replaced by manual operation while leaving the temperature
control scheme unaltered. It is planned to compare the performance
of the two schemes, the simplied one and the full NMPC scheme
also at the real plant, but implementation issues so far delayed the
test of the NMPC controller.

8. Conclusion
In this work, the operation of an industrial polymerization reactor is optimized online using two different model-based control
schemes. A process model that describes the dynamics of the relevant system states and the MWD of the polymer along the batch was
developed and validated using historical process data. After intensive investigations in a simulation environment, important issues
related to the practical implementation of the NMPC scheme, e.g.
state estimation and real-time solution of the optimization problem were addressed. In addition, a simplied control scheme with
model-predictive temperature control and a master controller for
the monomer feed was developed that could be integrated into the
DCS of the real process and is in daily operation at the real plant.
This scheme improves the quality of the temperature control and
reduces the duration as well as the variation of the batches already
signicantly. For the case where the cooling constraint is not active
and the monomer can be inserted into the reactor with maximal
ow rate during the whole feeding period, the results provided by
the PFC temperature controller are as good as the ones provided by
the generic NMPC scheme. For the case where the monomer feed
rate and the cooling power have to be simultaneously optimized,
the NMPC scheme provides a superior performance in simulations.
This is due to the fact that the NMPC controller takes the monomer
dosage and the cooling power simultaneously into account in the
optimal control computations while the alternative scheme optimizes the cooling power in an inner loop and the monomer dosage
in an outer loop. While better results can be expected from the use
of the full NMPC scheme at the real plant, it is an open issue whether
the gains over the simplied scheme will justify the additional

414

T.F. Finkler et al. / Journal of Process Control 24 (2014) 399414

effort for the maintenance of a dedicated solution implemented

on special soft- and hardware.
The main engineering effort for the development of the solutions presented here went into the setup, parameter estimation
and validation of the model that is used in the controller and in the
development of a robust and reliable state and parameter estimator. What made the situation favorable for this endeavor was the
fact that the structure of a rst-principles model of the polymerization was known and that from large amounts of historical data, the
missing parameters could be determined and the reliability of the
model could be judged. This model also predicts the inuence of
the operating parameters on the product quality sufciently well
which is a prerequisite for an optimization of recipe parameters.
It can be concluded that online optimizing NMPC provides a
versatile tool to simultaneously achieve very good control and
performance of polymerization processes. The available methods
and optimization algorithms are sufcient for the implementation at real processes of medium complexity as the one considered
here. However, the solutions must be engineered carefully which
involves in particular a signicant effort for process modeling and
parameter tuning and simulation studies. Simpler schemes which
approximately give the same performance are attractive because
of easier implementation, easier understanding by the operators,
and easier switching between manual and automatic operation.
The cascaded scheme proposed here has the structural advantage
to rely only on measured information and not on a plant model,
reducing the dependency on model quality and the computational
effort at the price of a reduced performance. Both schemes handle
the trade-off between accurate temperature control and maximizing the feed ow rate automatically and also improve the safety
of process operation, as the efciency of the cooling system (the
available cooling power) is automatically considered during operation whereas in a manual operation of the feed the operators have
to observe the temperature control loop and to reduce the feed if
they observe that the control valve saturates and the temperature
cannot be kept at the set-point any more. Of course there may be
other situations where the benet of model-based control schemes
is smaller. The engineering of simple near-optimal schemes is an
open eld for research, as well as the systematic robustication
of model-based schemes for which we propose the approach of
scenario-based multi-stage optimization [29,30].
References
[1] S.J. Qin, T.A. Badgwell, An Overview of NMPC Applications, Birkhuser, Basel,
2000.
[2] D.Q. Mayne, J. Rawlings, C. Rao, P. Scokaert, Constrained model predictive control: stability and optimality, Automatica 36 (6) (2000) 798814.
[3] H. Chen, F. Allgwer, A quasi-innite horizon nonlinear model predictive control scheme with guaranteed stability, Automatica 34 (2000)
12051217.
[4] Z.K. Nagy, B. Mahn, R. Franke, F. Allgwer, Nonlinear model predictive control of
batch processes: an industrial case study, in: 16th IFAC World Congress, Prague,
July 48, 2005.
[5] A. Kpper, S. Engell, Optimizing control of the Hashimoto SMB process: experimental application, in: Preprints of the 8th International IFAC Symposium on
Dynamics and Control of Process Systems, 2007.

[6] S. Engell, Feedback control for optimal process operation, J. Process Control 17
(2007) 203219.
[7] D. Bonvin, B. Srinivasan, Optimal operation of batch processes via the tracking
of active constraints, ISA Trans. 42 (2003) 123134.
[8] T.F. Finkler, M. Kawohl, U. Piechottka, S. Engell, Realization of optimizing
control in an industrial polymerization reactor, in: Preprints of the 8th IFAC
Symposium on Advanced Control of Chemical Processes, Singapore, July 1013,
2012.
[9] T.F. Finkler, M.B. Dogru, S. Engell, Optimizing control of semi-batch polymerization reactors, in: Preprints of the 2nd Conference on Advances on Control
and Optimization of Dynamics Systems, Bangalore, India, February 1618,
2012.
[10] T.F. Finkler, S. Lucia, M.B. Dogru, S. Engell, A simple scheme for the timeoptimal operation of semi-batch polymerizations, Ind. Eng. Chem. Res. 52
(2013) 59065920.
[11] J.S. Chang, W.Y. Hsieih, Optimization and control of semibatch reactors, Ind.
Eng. Chem. Res. 34 (1995) 545556.
[12] B. Srinivasan, D. Bonvin, E. Visser, S. Palanki, Dynamic optimization of batch
processes: Part 2, Comput. Chem. Eng. 27 (1) (2003) 2744.
[13] M. Vicente, S.B. Amor, L.M. Gugliotta, J.R. Leiza, J.M. Asua, Control of molecular weight distribution in emulsion polymerization using on-line reaction
calorimetry, Ind. Eng. Chem. Res. 40 (2001) 218227.
[14] C. Kravaris, R.A. Wright, J.F. Carrier, Nonlinear controllers for trajectory tracing
in batch processes, Comput. Chem. Eng. 13 (1989) 7382.
[15] C. Gentric, F. Pla, M.A. Lati, J.P. Corriou, Optimization and nonlinear control of a batch emulsion polymerization reactor, Chem. Eng. J. 75 (1999)
3146.
[16] R. Gesthuisen, S. Krmer, S. Engell, Hierarchical control scheme for timeoptimal operation of semibatch polymerizations, Ind. Eng. Chem. Res. 43 (2004)
74107427.
[17] W. Mauntz, A contribution to observation and time-optimal control of emulsion co-polymerization reactions, Dortmund Technical University, 2010 (Ph.D.
Thesis).
[18] V. Hagenmeyer, M. Nohr, Flatness-based two-degree-of-freedom control of
industrial semi-batch reactors using a new observation model for an extended
Kalman lter approach, Int. J. Control 81 (2008) 428438.
[19] J. Richalet, D. ODonovan, K.E. strm, Predictive Functional Control: Principles
and Industrial Applications (Advances in Industrial Control), Springer, London,
2009.
[20] S. Krmer, Heat balance calorimetry and multirate state estimation applied to
semi-batch emulsion copolymerization to achieve optimal coltrol, Dortmund
Technical University, 2006 (Ph.D. Thesis).
[21] L. Magni, D.M. Raimondo, F. Allgwer, Nonlinear Model Predictive Control:
Towards New Challenging Applications, Springer, Berlin, 2009.
[22] P.S. Agachi, Z.K. Nagy, M.V. Cristea, A. Imre-Lucaci, Model Based Control, WileyVCH, Weinheim (Germany), 2006.
[23] M. Diehl, S. Findeisen, S. Schwarzkopf, I. Uslu, F. Allgwer, H.G. Bock, E.D. Gilles,
J.P. Schlder, An efcient algorithm for nonlinear predictive control of largescale systems, Part I: Description of the method, Automatisierungstechnik 50
(2002) 557567.
[24] C. Kirches, L. Wirsching, H.G. Bock, J.P. Schlder, Efcient direct multiple shooting for nonlinear model predictive control on long horizons, J. Process Control
22 (2012) 18321843.
[25] R.E. Kalman, R. Bucy, New results in linear ltering and prediction, J. Basic Eng.
(ASME) 83D (1961) 98108.
[26] R. Chylla, D. Haase, Temperature control of semi-batch polymerization reactors,
Comput. Chem. Eng. 17 (1993) 257264.
[27] K. Pelz, H. Brandt, T.F. Finkler, S. Engell, Scheme for time-optimal operation of
semi-batch emulsion polymerization reactors, in: Preprints of the International
Symposium on Advanced on Control of Chemical Processes, Singapore, July
1013, 2012.
[28] H. Bouhenchir, M. Cabassud, M.V. Le Lann, Predictive functional control for the
temperature control of a chemical batch reactor, Comput. Chem. Eng. 30 (6/7)
(2006) 11411154.
[29] S. Lucia, T.F. Finkler, X. Basak, S. Engell, A new robust NMPC scheme and its
application to a semi-batch reactor example, in: Preprints of the 8th IFAC Symposium on Advanced Control of Chemical Processes, Singapore, July 1013,
2012.
[30] S. Lucia, T.F. Finkler, S. Engell, Multi-stage nonlinear model predictive control
applied to a semi-batch polymerization reactor under uncertainty, J. Process
Control 23 (9) (2013) 13061319.