ISA Transactions 51 (2012) 2229

Contents lists available at SciVerse ScienceDirect

ISA Transactions
journal homepage: www.elsevier.com/locate/isatrans

Design of a self-tuning regulator for temperature control of a polymerization

reactor
D. Vasanthi , B. Pranavamoorthy, N. Pappa
Department of Instrumentation Engineering, MIT campus, Anna University, Chrompet, Chennai-600 044, India

article

info

Article history:
Received 20 April 2011
Received in revised form
14 July 2011
Accepted 29 July 2011
Available online 20 August 2011
Keywords:
Online estimation
Unscented Kalman filter
Self-tuning control

abstract
The temperature control of a polymerization reactor described by Chylla and Haase, a control engineering
benchmark problem, is used to illustrate the potential of adaptive control design by employing a selftuning regulator concept. In the benchmark scenario, the operation of the reactor must be guaranteed
under various disturbing influences, e.g., changing ambient temperatures or impurity of the monomer.
The conventional cascade control provides a robust operation, but often lacks in control performance
concerning the required strict temperature tolerances. The self-tuning control concept presented in this
contribution solves the problem. This design calculates a trajectory for the cooling jacket temperature in
order to follow a predefined trajectory of the reactor temperature. The reaction heat and the heat transfer
coefficient in the energy balance are estimated online by using an unscented Kalman filter (UKF). Two
simple physically motivated relations are employed, which allow the non-delayed estimation of both
quantities. Simulation results under model uncertainties show the effectiveness of the self-tuning control
concept.
2011 ISA. Published by Elsevier Ltd. All rights reserved.

1. Introduction
Polymerization is the process of reacting monomer molecules
together in chemical reactions to form three-dimensional networks of polymer chains. The widely used reactors in the chemical
industry for the production of fine pigments, chemicals, polymers
and pharmaceuticals are batch and semi-batch reactors. Measurement and control of polymerization reactors is very challenging
due to the complexity of the polymerization kinetics. In the literature, hierarchical approaches to the control system design, and
reviews of traditional regulatory techniques and advance control
strategies for batch and semi-batch process are presented [1,2].
Often the reactions show exothermic behavior, and tight specifications have to be met. Therefore, an exact temperature control is
required. In many cases, conventional linear control algorithms are
reported not to fulfil this requirement. Industrial polymerization
reactors are usually controlled by a cascade structure consisting of
a master controller for the reactor temperature and an underlying slave controller for the cooling circuit. The cascade controller
gives a robust operation but fails to meet the strict temperature
tolerances [3]. Hence there is a need to employ advanced control
techniques such as the concepts of adaptive control [4] to satisfy
the control requirements.

Corresponding author.
E-mail address: vasanthi@annauniv.edu (D. Vasanthi).

Therefore, various control concepts have been used in the

literature to deal with the benchmark problem. A model predictive
controller is used in [5] combined with an extended Kalman filter
for the estimation of the reaction heat and heat transfer coefficient.
A nonlinear adaptive controller is designed in [6] to adjust the
cooling jacket temperature, which serves as the set point for
an underlying proportionalintegral (PI) controller of the cooling
jacket. A further approach [7] illustrates the potential of feedforward control by extending the conventional cascade control
concept in the framework of a two degree of freedom control
concept.
This paper solves the problem by employing an unscented
Kalman filter (UKF) for the non-delayed online estimation of
the reaction heat and heat transfer coefficient which is used to
calculate the trajectory for the cooling jacket temperature, which
in turn enables the reactor temperature to follow the predefined
trajectory for the case of polymer A.
2. The ChyllaHaase bench mark reactor
2.1. Polymerization reactor
The industrial polymerization process described by Chylla and
Haase [7] consists of a stirred tank reactor with a cooling jacket
and a coolant recirculation; see Fig. 1. The reactor temperature is
controlled by manipulating the temperature of the coolant, which
is recirculated through the cooling jacket of the reactor. The slave

0019-0578/$ see front matter 2011 ISA. Published by Elsevier Ltd. All rights reserved.
doi:10.1016/j.isatra.2011.07.009

D. Vasanthi et al. / ISA Transactions 51 (2012) 2229

23

Fig. 1. Schematic diagram of a polymerization reactor.

controller can act in two modes: in cooling mode, cold water is

inserted into the recirculation loop, whereas, in heating mode,
steam is injected into the recirculating water stream.
The polymerization process is simulated for a product which
comprises a specific recipe [3] which is given below.
The recipe for each batch of a specific polymer consists of a
heating phase from 0 to 1800 s, a feed phase from 1800 to 6000 s,
and a holding phase from 6000 to 9600 s.
1. Initial charges of solids (prepolymer, monomer) and water are
placed into the reactor at ambient temperature, Tamb .
2. The temperature of the initial charge is raised to the reaction
temperature set point T set at 1800 s.
3. After 1800 s, pure monomer is fed into the reactor at 7.560
103 kg/s until 6000 s.
4. After the feed addition has stopped, the temperature of the
reaction is held at its set point value T set until 9600 s.

Table 1
Variables and parameters of the reactor model.
Variable

Variable name

min
M

Unit

(t )

Monomer feed rate

kg/s

Qrea = H Rp

Reaction heat

kW

Rp

Rate of polymerization

kg/s

Reaction enthalpy

kJ kg1

Overall heat transfer coefficient

kW m2 K1

Jacket heat transfer area

m2

(UA)loss
Cp,M , Cp,P , Cp,C
1 , 2

Heat loss coefficient

kW/K

Tj =
(Tjin + Tjout )/2
Kp (c )

Specific heat at constant pressure

kJ kg1 K1

Transport delay in jacket and recirculation

loop

Average cooling jacket temperature

Heating/cooling function

Heating/cooling time constant

The above recipe for a specific product comprises two consecutive

control tasks.

Heating up of the reactor to a constant set point before the

monomer feed starts.
Keeping the reactor temperature T (t ) within a tolerance
interval of 0.6 K during the monomer feed and after a
subsequent specified holding period.
Thereby, the monomer feed starts and ends abruptly at specified
time points. The polymer is produced in five subsequent batches,
and is removed between the batches. However, the reactor is only
cleaned after the fifth batch. The control tasks are complicated by
several features of the polymerization process.

of the cooling jacket and the recirculation loop with the outlet
and inlet temperatures of the coolant C . Further variables and the
parameters of the reactor model are listed in Table 1.
dmM
dt
dmP
dt
dT
dt

in
=m
M (t ) +
=

2.2. Polymerization reactor model

A dynamic model of a polymerization reactor has been derived
by Chylla and Haase based on simplified kinetic relations [7].
The temperature dynamics are captured by considering energy
balances for the reactor and the cooling jacket. The reactor model
comprises the material balances (1) and (2) for the monomer mass
mM (t ) and the polymer mass mP (t ), the energy balance (3) with
the reactor temperature T (t ), and the energy balances (4) and (5)

(1)

Qrea

(2)

1
mi Cp,i

in
[m
M (t ) Cp,M (Tamb T ) UA(T Tj )

Changing ambient temperatures in winter and summer.

The heat transfer coefficient decreases significantly during a
batch due to an increasing batch viscosity, and from batch to
batch due to surface fouling.
The reaction kinetics is nonlinear.
Impurity of the monomer.

Qrea

(UA)loss (T Tamb ) + Qrea ] (i = M , P , W )

dTjout
dt
dTjin
dt

1
mc Cp,C

c Cp,C (Tjin (t 1 ) Tjout ) + UA(T Tj )]

[m

dTjout (t 2 )
dt

(3)

Tjout (t 2 ) Tjin

Kp (c )

(4)

(5)

The available measurements of the process are the temperatures

of the reactor and cooling circuitry. The heating/cooling function
Kp (c ) is influenced by an equal-percentage valve with valve
position c (t ) and the following split-range valve characteristic:

0.8 30c /50 Tinlet Tjin (t ) ,

Kp (c ) = 0,

0.15 30c /(502) Tstream Tjin (t ) ,

c < 50%,
c = 50%,
c > 50%.

(6)

24

D. Vasanthi et al. / ISA Transactions 51 (2012) 2229

Table 2
Empirical relations for the polymerization rate Rp , the jacket heat transfer area A, and the overall heat transfer coefficient U.
Variable

Variable

Variable name

Unit

Rp = ikmM

Impurity factor

E
(k1 )k2
k = k0 exp
RT

First-order kinetic constant

s1

= c0 exp(c1 f ) 10
f = mP /(mM + mP + mC )
k0 , k1 , E , R, a0 , c0 , c1 , c2 , c3

Batch viscosity

kg m1 s1

Auxiliary variable

Natural gas constant

M , P , C

Densities

kg m3

B1

Reactor bottom area

Jacket perimeter

B2

Jacket bottom area

h = d0 exp (d1 wall )

Film heat transfer coefficient

wall = c0 exp (c1 f ) 10c2 (a0 /Twall c3 )

Twall = T + Tj /2

Wall viscosity

kg m1 s1

Wall temperature

1
h
f

Fouling factor depending on batch no.

m2 K/kW

d0 , d1

Constants

c2 (a0 /T c3 )

A=

U =

mM

mP

mW

P
B1

+ B2

1
1
h 1 + hf

Table 3
Data for reactor parameters.

Constants
kJ kmol1 K1

kW m2 K1

Table 4
Data for polymer A.

Symbol

Unit

Value

Symbol

Unit

Value for polymer

kJ kmol1 K1

8.314

mM ,0

kg

0.0

(UA)loss
p
1
2

kW K1

0.00567567

mP ,0

kg

11.227

40.2

mW

kg

42.75

22.8

kg m3

900.0

15.0

kg m3

1040.0

kg m3

1000.0

1/hf

m2 K kW1

[0.8, 1.2]
[0.000, 0.176, 0.352, 0.528, 0.704]

M
P
W
Cp,M

kJ kg1 K1

1.675

Tamb

280.382 (winter), 305.382 (summer)

Cp,P

kJ kg1 K1

3.140

Tinlet

294.26

Cp,W

kJ kg1 K1

4.187

Tsteam

449.82

MWM

kg kmol1

104.0

mc

kg

42.750

c
m

kg/s

0.9412

Cp,C

kJ kg1 K1

4.187

k0

s1

55

k1

m s kg1

1000

For c < 50%, ice water with inlet temperature Tinlet is inserted in
the cooling jacket, whereas a valve position c > 50% leads to a
heating of the coolant by injecting steam with temperature Tsteam
into the recirculating water stream [7,5].
Various disturbances and uncertainties are specified in order
to model the following practical problems with the control of
polymerization reactors.

The impurity factor i [0.8, 1.2] in the polymerization rate

RP is random and constant during one batch; it tries to model

fluctuations in monomer kinetics caused by batch-to-batch
variations in reactive impurities. The fouling factor 1/hf in the
overall heat transfer coefficient U increases with each batch and
accounts for the fact that during successive batches a polymer
film builds up on the wall, resulting in a decrease of U [7,6].
The delay times 1 and 2 of the cooling jacket and the
recirculation loop may vary by 25% compared to the nominal
values in Table 3.
The ambient temperature Tamb is different during summer and
winter. This affects the temperature of the monomer feed
in
out
in
m
(0)
M (t ), as well as the initial conditions T (0), Tj (0), Tj
given by Tamb .
Measurement noise is added to the temperature measurements
with standard deviation (y) = 0.5.
Process noise is added with standard deviation = 0.3162.

Table 2 summarizes the empirical relations for Rp , A, and U taken

from [8,3,7]. A detailed description of these relations is given in [3].
The parameter values of the model and of polymer A are listed in
Tables 3 and 4.

k2

0.4

kJ kmol1

29,560.89

c0

kg m1 s1

5.2 105

c1

16.4

c2

2.3

c3

1.563

a0

555.556

Hp

kJ kmol1

70,152.16

d0

kW m2 K1

0.814

d1

m s kg1

5.13

kg/s

7.560 10

in,max

M
m

in
in
tM
,0 , tM ,1

T set

min

[30, 100]

355.382

In order to heat up the reactor to the specified set point T set

in
before the monomer feed m
M (t ) starts, a trajectory is planned for
the desired reactor temperature T (t ) by the polynomial set-up as
given by Eq. (7):
T (t )

Tamb + (T set Tamb )

T

set

i =3

ai

t
t heat

if , t t heat
if t > t

heat

(7)

D. Vasanthi et al. / ISA Transactions 51 (2012) 2229

25

Fig. 2. Conventional cascade control scheme.

Fig. 3. Self-tuning control scheme.

Fig. 4. Response with conventional cascade controller for nominal parameters and summer season: (i) results for first batch, (ii) results for fifth batch.

Here, a3 = 10, a4 = 15, and a5 = 6. In this first interval t t heat ,

the reactor is heated up to the set point T set . For t > t heat , the
temperature T (t ) is kept constant at T set . In the following, the
heat-up time is set to t heat = 30 min, corresponding to the time
in
instant when the monomer feed m
M (t ) starts.
3. Self-tuning cascade control design
A very tight temperature control is necessary in order to
produce polymer of a desired quality. The controller should be
able to maintain the reactor temperature T within an interval of
0.6 K around the desired set point under all operation conditions
and disturbances. Commonly used for chemical reactors is a PI
cascade control structure, which provides a robust operation but
often lacks in control performance. The cascade control structure

is shown in Fig. 2. The master controller regulates the reactor

temperature T by manipulating the set point Tjset of the mean
cooling jacket temperature Tj . The slave controller adjusts the valve
position c in order to control the mean jacket temperature Tj set by
the master controller.
Since the nonlinear function is not approximated and the
Jacobean is not required, an unscented Kalman filter (UKF) is used
in preference to other algorithms.
In order to get an accurate response, it is desirable to calculate
the self-tuning parameters online. However, this requires the
estimation of the time-dependent quantities, the reaction heat
(Qrea ) and the heat transfer coefficient (U), in successive batches.
Fig. 3 shows the block diagram of the considered adaptive control
scheme with a UKF [9,10], which uses the available temperature
measurements to estimate Qrea and U. These estimates are used in
the online design of the self-tuning cascade control.

26

D. Vasanthi et al. / ISA Transactions 51 (2012) 2229

Fig. 5. Response with self-tuning controller for nominal parameters and summer season: (i) results for first batch, (ii) results for fifth batch.

When compared with other forms of Kalman filter algorithms,

a UKF has the following advantages.

It is a deterministic sampling approach.

There is no approximation of the nonlinear function.
It captures the true mean and covariance more accurately.
Calculation of the Jacobean is not required.

In a UKF the nonlinear function is not approximated and the

reactor model is used as given by Eqs. (1)(5); hence the estimates
obtained are accurate.
The uncertainties of the system are modeled as the system
noise in the state error covariance matrix, and the uncertainties
in the measurement are modeled as the measurement noise in the

D. Vasanthi et al. / ISA Transactions 51 (2012) 2229

27

Fig. 6. Response with conventional cascade controller for nominal parameters and winter season: (i) results for first batch, (ii) results for fifth batch.

measurement error covariance matrix. The estimated output not

only estimates the Qrea and U which cannot be measured online,
but also gives a filtered output of the various parameters required
for taking the control action.
Tj = T +

in
mi Cp,i T m
M (t )Cp,M (Tamb T )

Qrea
+ (UA)loss (T (t ) Tamb )

Control scheme

Cascade control
Self-tuning control

U .A

Table 5
Comparison of ISE and IAE values of both control schemes during summer.

(8)

Thus Tj is found with the help of Eq. (8), and this Tj is compared
with the current jacket temperature, Tj , and control action is taken
by manipulating the control valve, which in turn maintains the
reactor temperature at its set point, and the error is kept within
the tolerance limit.
4. Simulation results and discussion
The process is simulated with the conventional PI cascade
control structure as well as the self-tuning control structure and
the performance of both the control schemes are compared.
Fig. 4(i)(a)(c) and (ii)(a)(c) show the simulation results for the
process with conventional cascade control for batches 1 and 5 for
the summer season, respectively. The sampling time for the selftuning control is set to 1 s, which is the same as the sampling time
of the cascade controller. Fig. 5(i)(a)(h) and (ii)(a)(e) give the
simulation results for the process with self-tuning control for the
summer season for batches 1 and 5, respectively.
As per the recipe given in Section 2.1, at 1800 s the monomer,
which is at ambient temperature, is added in the reactor, but the
reactor is at an operating temperature of T set = 355.382 K at that
instant. Therefore the reactor is initially cooled by the monomer
feed before the reaction, and hence there is a need to increase
temperature by means of increasing the jacket temperature. The
process exhibits an exothermic reaction; thus the heat inside the
reactor will be high, and so to compensate this there has to be
reduction in the jacket temperature. At the end of 6000 s, the
monomer feed is stopped and the reaction will come to an end,

ISE

IAE

First batch

Fifth batch

First batch

Fifth batch

764.5785
54.9608

994.4308
85.6100

1455.5
449.0182

1719.6
579.0522

but the polymer produced in the reactor needs to be maintained at

T set = 355.382 K. To maintain this T set , again, there is an increase
in the jacket temperature by means of steam inflow. Due to this
abrupt change in monomer feed there are major changes in the
temperature error and control valve stem position at 1800 and
6000 s.
The deviation of the reactor temperature T and desired reactor
temperature T is due to the time delay in the control system
structure chosen for the slave controller. The respective set point
for the slave controller is adjusted by the master controller in
the case of conventional cascade control, whereas the respective
set point is generated by a parameter adjustment mechanism
in the case of this proposed self-tuning control scheme. The
simulations with the cascade controlled reactor together with the
self-tuning control illustrate that the temperature error can be
reduced significantly compared to that of a conventional cascade
control scheme. The temperature error stays within the tolerance
interval of 0.6 K during the heating stage as well as at the critical
time points when the monomer feed starts and stops. There is
also significant improvement in the performance of the control
valve.
As a measure of assessing control system performance, for both
the control schemes integral squared error and integral absolute
error values are calculated for batches 1 and 5 respectively for the
summer season and are shown in Table 5. It is observed that both
ISE and IAE are much lower for the self-tuning control scheme
compared to the cascade control structure.
Furthermore, a robustness analysis is carried out in order to
compare the performance of the self-tuning control structure with
respect to the conventional PI cascade control structure. It is
assumed that the reactor is operated in the winter season where
the low ambient temperature (Table 3) poses high demand on

28

D. Vasanthi et al. / ISA Transactions 51 (2012) 2229

Fig. 7. Response with self-tuning controller for nominal parameters and winter season: (i) results for first batch, (ii) results for fifth batch.

the cascade control during the heating up of the reactor. The UKF
provides robust estimates of the reaction heat and heat transfer
coefficient, and the reactor temperature stays within the specified
tolerance interval of 0.6 K with the self-tuning control scheme,
as shown in Fig. 6(i)(a)(c) and (ii)(a)(c) and Fig. 7(i)(a)(h) and
(ii)(a)(f).

As a measure of assessing control system performance, for both

control schemes ISE and IAE values are calculated for batches 1
and 5, respectively, for the winter season and are shown in the
Table 6. It is observed that both ISE and IAE are much lower for
the self-tuning control scheme compared to the cascade control
structure.

D. Vasanthi et al. / ISA Transactions 51 (2012) 2229

Table 6
Comparison of ISE and IAE values of both control schemes during winter.
Control scheme

Cascade control
Self-tuning control

ISE

IAE

First batch

Fifth batch

First batch

Fifth batch

848.0387
58.9357

1175.6
75.5651

1413.8
458.6496

1717.3
556.1799

5. Conclusion
The Chylla and Haase polymerization process is a nonlinear
and multi-batch process, for which temperature control is very
difficult. A self-tuning controller has been designed to obtain the
desired control performance. The performance of the self-tuning
cascade control has been studied for batches 1 and 5 for the given
polymer. Also estimates for the heat transfer coefficient U and
reaction heat Qrea are obtained using an unscented Kalman filter for
these batches. On comparison with a conventional cascade control,
the performance of the self-tuning cascade control is found to be
superior, and it meets the control objective of maintaining the
reactor temperature within the tolerance interval of 0.6 K from
the set point.

29

References
[1] Richards John R, Congalidis John P. Measurement and control of polymerization reactors. Computers and Chemical Engineering 2006;30:144763.
[2] Friedrich M, Perne R. Design and control of batch reactorsan industrial
viewpoint. Computers and Chemical Engineering 1995;19:35768.
[3] Chylla RW, Randall Haase D. Temperature control of semi batch polymerization reactors. Computers Chemical Engineering 1993;17:25763.
[4] Astrom Karl J, Wittermark Bjorn. Adaptive control. second ed. Pearson
Education Singapore; 1995. p. 1924.
[5] Helbig A, Abel O, Mhamdi A, Marquardt W. Analysis and nonlinear model
predictive control of the Chylla Haase benchmark problem. In: Proceedings
of the UKACC international conference on control. 1996.
[6] Clarke-Pringle Tracy, MacGregor John F. Non linear adaptive temperature
control of multi product semibatch polymerization reactors. Computers
Chemical Engineering 1997;21:1395409.
[7] Graichen Knut, Hagenmeyer Veit, Zeitz Michael. Feed forward control with
online parameter estimation applied to the ChyllaHaase reactor benchmark.
Journal of Process Control 2006;16:73345.
[8] Hinsberger H, Miesbach S, Pesch HJ. Optimal temperature control of semibatch
polymerization reactors. In: Scientific computing in chemical engineering.
Springer; 1996. p. 7583.
[9] Wan Eric A, van der Merwe Rudolph. The unscented Kalman filter for nonlinear
estimation. In: Proceedings of symposium on adaptive systems for signal
processing, communication and control. IEEE; 2000. p. 1538.
[10] Simon Dan. Non linear filters. In: Optimal state estimation. first ed. John Wiley
& Sons Publication; 2006. p. 395483.