Journal of Process Control 22 (2012) 947958

Contents lists available at SciVerse ScienceDirect

Journal of Process Control

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

Control of industrial gas phase propylene polymerization in uidized bed reactors

Yong Kuen Ho a , Ahmad Shamiri a , Farouq S. Mjalli b, , M.A. Hussain a
a
b

Chemical Engineering Department, Faculty of Engineering, University of Malaya, 50603 Kuala Lumpur, Malaysia
Petroleum & Chemical Engineering Department, Sultan Qaboos University, Muscat 123, Oman

a r t i c l e

i n f o

Article history:
Received 26 July 2011
Received in revised form 3 April 2012
Accepted 3 April 2012
Available online 3 May 2012
Keywords:
Adaptive Predictive Model-Based Control
Fluidized bed reactor
Propylene polymerization
ZieglerNatta catalyst

a b s t r a c t
The control of a gas phase propylene polymerization model in a uidized bed reactor was studied, where
the rigorous two phase dynamic model takes into account the polymerization reactions occurring in the
bubble and emulsion phases. Due to the nonlinearity of the process, the employment of an advanced
control scheme for efcient regulation of the process variables is justied. In this case, the Adaptive Predictive Model-Based Control (APMBC) strategy (an integration of the Recursive Least Squares algorithm,
RLS and the Generalized Predictive Control algorithm, GPC) was employed to control the polypropylene
production rate and emulsion phase temperature by manipulating the catalyst feed rate and reactor cooling water ow, respectively. Closed loop simulations revealed the superiority of the APMBC in setpoint
tracking as compared to the conventional PI controllers tuned using the Internal Model Control (IMC)
method and the standard ZieglerNichols (ZN) method. Moreover, the APMBC was able to efciently
arrest the effects of supercial gas velocity, hydrogen concentration and monomer concentration on the
process variables, thus exhibiting excellent regulatory control properties.
2012 Elsevier Ltd. All rights reserved.

Contents
1.
2.
3.
4.
5.
6.

Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Mathematical modelling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Outline of the adaptive predictive model-based control scheme . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Control system design of the adaptive predictive model-based control scheme on the polymerization reactor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Closed loop performance of the adaptive predictive model-based control scheme on the polymerization reactor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

1. Introduction
The production of polypropylene in uidized bed reactors is one
of the most widely used technologies in the polymerization industry. However, the complicated reaction, heat and mass transfer
mechanisms as well as the complex gas and solid ow characteristics in the reactor introduce extreme nonlinearities in the dynamics
of the reactor. As such, the modelling and control of such a process
is a huge challenge. The fundamental control problem in the propylene polymerization uidized bed reactor is further complicated
by the existence of strong interaction between reactor variables;
of which conventional process control strategies are incapable
of coping. Although there are studies reported in the academic

Corresponding author. Tel.: +968 93284366.

E-mail address: farouqsm@yahoo.com (F.S. Mjalli).
0959-1524/$ see front matter 2012 Elsevier Ltd. All rights reserved.
http://dx.doi.org/10.1016/j.jprocont.2012.04.003

947
951
952
953
955
957
957

literature on the modelling and control of polymerization process in

uidized bed reactors [15], little work was done on the modelling
and control of specically the gas phase uidized bed polypropylene reactor.
The simplied schematic of an industrial gas phase uidized
bed polypropylene reactor is shown in Fig. 1. In the polymerization reactor, growing polymer particles are uidized by a recycle
gas stream containing propylene monomer, hydrogen and nitrogen.
The feed gas ow, which provides the monomer for polymerization
in the presence of ZieglerNatta catalyst and triethyl aluminum
co-catalyst, serves to agitate the bed and uidization through the
distributor while at the same time removing the heat of polymerization from the reactor. The recycle gas, which exits from the top
of the reactor is then compressed and cooled before being fed into
the bottom of the uidized bed.
Due to the highly exothermic nature of the propylene polymerization reactions, the cooling of the recycle gas which exits from the

948

Y.K. Ho et al. / Journal of Process Control 22 (2012) 947958

Nomenclature
a(z1 )

polynomial associated with y(k) in the CARIMA

model
A
cross sectional area of the reactor (m2 )
Archimedes number
Ar
b(z1 )
polynomial associated with u(k) in the CARIMA
model
VFF-RLS design constant
C
CH
hydrogen concentration (kmol/m3 )
CM
Propylene (monomer) concentration (kmol/m3 )
Cpi
specic heat capacity of component i (J/kg K)
Cpg
specic heat capacity of gaseous stream (J/kg K)
Cp,pol
specic heat capacity of solid product (J/kg K)
bubble diameter (m)
db
dp
particle diameter (m)
D
diagonal matrix
gas diffusion coefcient (m2 /s)
Dg
Dt
reactor diameter (m)
catalyst feed rate (g/s)
Fcat
Fcw
cooling water ow rate (g/s)
H
matrix associated with the future slew rates in the
GPC prediction equation
bubble to emulsion heat transfer coefcient
Hbe
(W/m3 K)
Hbc
bubble to cloud heat transfer coefcient (W/m3 K)
Hce
cloud to emulsion heat transfer coefcient (W/m3 K)
active site type
j
k
positive integer denoting the sampling instance
kfh (j)
transfer rate constant for a site of type j with terminal monomer M reacting with hydrogen
kg
gas thermal conductivity (W/m K)
kp (j)
propagation rate constant for a site of type j with
terminal monomer M reacting with monomer M
K
matrix associated with the past slew rates in the GPC
prediction equation
Kbc
bubble to cloud mass transfer coefcient (s1 )
bubble to emulsion mass transfer coefcient (s1 )
Kbe
Kc
proportional gain
Kce
cloud to emulsion mass transfer coefcient (s1 )
[Mi ]
concentration of component i in the reactor
(kmol/m3 )
[Mi ]in
concentration of component i in the inlet gaseous
stream
Mi
ith component
monomer molecular weight (kg/kmol)
Mw
N1
minimum prediction horizon
N2
maximum prediction horizon
NINT[] nearest integer to []
number of active site types
NS
M
control horizon
p
covariance matrix
pressure (Pa)
P
Q
matrix associated with the past outputs in the GPC
prediction equation
move suppression weight for the slew rate
R
Ri
instantaneous rate of reaction for monomer i
(kmol/s)
R(j)
rate at which monomer M is consumed by propagation reactions at sites of type j
Remf
Reynolds number of particles at minimum uidization condition
R
positive denite diagonal weighting matrices containing R

vector of future setpoints

Rp
Rpb
Rpe
Rv

y(k)
y

production rate (kg/s)

bubble phase production rate (kg/s)
emulsion phase production rate (kg/s)
volumetric polymer phase outow rate from the
reactor (m3 /s)
time (s)
sampling time
temperature (K)
discrete dead time
temperature of the inlet gaseous stream (K)
design polynomial in the CARIMA model
upper triangular matrix
supercial gas velocity (m/s)
bubble velocity (m/s)
emulsion gas velocity (m/s)
process input
minimum uidization velocity (m/s)
stochastic noise variable with random distribution
and zero mean
reactor volume (m3 )
volume of the bubble phase (m3 )
volume of the emulsion phase (m3 )
volume of polymer phase in the reactor (m3 )
volume of polymer phase in the bubble phase (m3 )
volume of polymer phase in the emulsion phase
(m3 )
volume of plug ow reactor (PFR) (m3 )
weight for the output residual
positive denite diagonal weighting matrices containing W
process output
vector of future outputs (not including current

value)
vector of past outputs (including current value)

t
ts
T
Td
Tin
T(z1 )
U
U0
Ub
Ue
u(k)
Umf
v(k)
V
Vb
Ve
Vp
Vpb
Vpe
VPFR
W
W

Y(n,j)

nth moment of chain length distribution for living

polymer produced at a site of type j

Greek letters
HR
heat of reaction (J/kg)
u
change in the input variable (or slew rate)
vector of future slew rates (not including current
u
k1

u

k1


d

b
e
mf


g
pol


value)
vector of past slew rates (including current value)
order of the a(z1 ) polynomial
order of the b(z1 ) polynomial
volume fraction of bubbles in the bed
Kalman gain
forgetting factor
dead time of the process
prediction error
void fraction of bubble for Geldart B particles
void fraction of emulsion for Geldart B particles
void fraction of the bed at minimum uidization
gas viscosity (Pa s)
density (kg/m3 )
gas density (kg/m3 )
polymer density (kg/m3 )
VFF-RLS design constant
regressor vector
vector of the estimated process model parameters

Y.K. Ho et al. / Journal of Process Control 22 (2012) 947958

D
I

derivative time constant

integral time constant

Subscripts and superscripts

1
propylene
2
hydrogen
bubble phase
b
e
emulsion phase
g
gas mixture property
component type number
i
in
inlet
j
active site type number
mf
minimum uidization
pol
polymer
ref
reference condition

top of the reactor (as described above) is crucial to maintaining the

polypropylene production rate. The heat removal from the recycle
gas through the cooler in this system is governed by the heat capacity, dew point, and the temperature of the recycle gas. To maintain
acceptable polymer production rate (which is an important goal for
the industry), it is necessary to keep the reactor bed temperature
above the dew point of the reactants to avoid condensation of gas
within the reactor. At the same time, the reactor bed temperature
must also be below the melting point of the polymer to prevent particle melting, agglomeration and consequently reactor shut down.
In view of the various stringent operating conditions required, stabilization of propylene polymerization in uidized bed reactors is
a challenging task which can only be addressed through efcient
control system design.
For the uidized bed polymerization reactor, most of the reactor design and control problems are associated with achieving
adequate production rate and heat removal from the reactor. The
steady state and dynamic behavior of the reactor is inuenced by
many process variables such as the supercial gas velocity, feed gas
temperature, monomer concentration, catalyst activity, and catalyst feed rate, etc. The rst attempt in describing the dynamics of
propylene polymerization was done by Choi and Ray [1]. In their
work, two phases, viz. the bubble and emulsion phases were considered in the proposed dynamic model. Furthermore, it was shown
that the employment of a PI feedback control scheme was capable of
controlling the process transients, albeit still limited by the recycle
gas cooling capacity. Ibrehem et al. [4] proposed that the uidized
bed comprise four phases, namely the bubble, cloud, emulsion and
solid phases, with polymerization reactions occurring only in the
emulsion and solid phases. In the same work, the temperature of
the polyethylene system was controlled by using a neural network
based predictive controller. Moreover, it was shown that the performance of the neural network based predictive controller was
superior to the conventional PID controller in terms of setpoint
tracking. In another study on industrial gas phase polyethylene
reactors, Dadebo et al. [2] showed the nonlinear Error Trajectory
Controller (ETC) exhibited superior responses in terms of speed,
damping and robustness as compared to an optimally tuned PID
controller for the control of temperature over a wide range of operating conditions.
In the present study, the control of a newly proposed two
phase dynamic model for homo-polymerization of propylene in uidized bed reactors will be investigated. In the proposed dynamic
model, the presence of solid particles in the bubble phase (which
has been validated experimentally [68]) as well as the excess
gas in the emulsion phase [79] and polymerization reaction
in both phases are considered to realistically mimic the actual

949

polymerization process which occurs in the uidized bed reactor

[7,8]. In short, the present model takes into account the polymerization reactions which occur in two phases. In this system,
the reactor temperature is controlled by manipulating the cooling
water ow rate of an upstream heat exchanger, which consequently
removes the heat produced by the exothermic polymerization reactions from the recycle gas stream [2]. Concurrently, changes to
the catalyst feed rate and the bed level can be used to control
both the polymer production rate and the solid phase residence
time [10]. The catalyst injection and product withdrawal rates are
adjusted in such a way that maintains a constant bed level inside the
reactor.
In this work, adaptive control strategy was employed to deal
with the control issue of the polypropylene uidized bed reactor. Due to the high nonlinearities involved in the dynamics of the
polymerization reactor, it is therefore beyond the capability of the
conventional PID controller with xed controller settings to achieve
excellent control of the reactor variables. In order to achieve good
control of the reactor variables, a more intelligent and efcient
process control scheme is needed, where the controller is able to
automatically re-design itself in real time according to the changing process dynamics. Such a controller which can readjust itself
(i.e. the controller settings are updated in real time according to
the most recent process dynamics) is referred to as adaptive controllers. Where nonlinearities are concerned, adaptive controllers
are known to be adept in handling such difcult process control
scenarios [11,12].
Among the multitudes of adaptive control techniques available
to date, the self-tuning approach has received the most attention
in the past decades [1315]. In this approach, the controller is
designed in real time to adapt to the most recent dynamics of the
process based on the output of a recursive system identication
procedure (i.e. the estimated process model parameters). Ljung and
Sderstm [16] presented a number of mathematical algorithms
which can be used for identifying a system recursively. Among
these, Seborg et al. [13] reported that the Recursive Least Squares
(RLS) algorithm and the Extended Least Squares (ELS) algorithm
are the two most commonly used recursive system identication
algorithms in adaptive control. However, the use of the RLS algorithm (i.e. the choice of recursive system identication technique
in this study) in adaptive control is comparatively more popular
than the ELS algorithm due to its simplicity and rapid convergence
when properly applied. Given the real time data of the inputs and
the outputs of a process, the RLS algorithm produces sequential
updates of the process model parameters, which can be used subsequently in controller design. Although the success of RLS-based
adaptive controllers in controlling polymerization reactors (both
the lab scale as well as the industrial scale) is evident in the literature [1721], there is virtually no published work to date as
pertaining to the control of a gas phase propylene polymerization
uidized bed reactor.
The choice of the control algorithm to be coupled to the RLS
algorithm is of equal importance to the recursive system identication component in the design of an adaptive controller. In this
study, the Model Predictive Control (MPC) technology, specically
the linear Generalized Predictive Control (GPC) strategy is selected
as the choice of control algorithm. Although the rst principle
polymerization model can be used with the nonlinear MPC to
control the process, this is not done in this work as the mechanistic
polypropylene polymerization process model is overly complicated (e.g. with excessive number of states, etc.) to be used within
a model based framework. In our case, the internal model used by
the GPC for predictive calculations is regenerated at every control
interval by the RLS algorithm. In essence, the nonlinear process
is continuously being identied by the RLS algorithm in the form
of an instantaneous linear model which can be used by the GPC

950

Y.K. Ho et al. / Journal of Process Control 22 (2012) 947958

Fig. 1. Schematic diagram of an industrial gas phase uidized bed polypropylene production reactor.

for prediction as well as computation of optimized control moves.

Such a synergistic combination is referred to as the Adaptive
Predictive Model-Based Control (APMBC). Fig. 2 shows the design
of the APMBC controller on the propylene polymerization uidized
bed reactor. As illustrated in the gure, decentralized APMBC
strategy was adopted in this study. As alluded to previously, the

polypropylene production rate and the emulsion phase temperature were controlled by manipulating the catalyst ow rate
and the heat exchanger coolant ow rate, respectively. Although
the decentralized control strategy is known for its inability to
account for the full multivariable dynamics of the process in the
controller design, it is nonetheless simpler and more practical to

Fig. 2. Simplied schematic of the APMBC design on the gas phase propylene polymerization uidized bed reactor. Fcw is the coolant ow rate, Fcat is the catalyst ow rate,
is the matrix of estimated process model parameters.
and

Y.K. Ho et al. / Journal of Process Control 22 (2012) 947958

951

Table 1
Correlations and equations used in the two-phase model.
Parameter

Formula

Minimum uidization velocity

Bubble velocity
Bubble rise velocity

Remf = [(29.5) + 0.357Ar]

Ub = U0 Ue + ubr
1/2
ubr = 0.711(gdb )

Reference

Emulsion velocity

Ue =

Bubble diameter

db = db0 [1 + 27(U0 Ue )] (1 + 6.84H)

db0 = 0.0085 (for Geldart B)

Kbe =

U0 Ub
1

1
Kbc

Kbc = 4.5

Mass transfer coefcient

1/2

29.5

[41]
[42]
[42]
[43]

1/3

1

1
K

 Ue ce

+ 5.85

 Dg e ubr 
 1 d1b 1
Hbe = H + Hce
bc
 Ue g Cpg 
db

Dg 1/2 g 1/4
db

[44]


[42]

5/4

Kce = 6.77
Heat transfer coefcient

Hbc = 4.5

db

Hce = 6.77(g Cpg kg )

[42]

+ 5.85

1/2

d 5/4

e ubr
db 3

= 0.534 1 exp

Bubble phase fraction

(kg g Cpg )1/2 g 1/4

U0 Umf

1/2b


0.413

Emulsion phase porosity

e = mf + 0.2 0.059 exp

Bubble phase porosity

b = 1 0.146 exp

Volume of polymer phase in the emulsion phase

Volume of polymer phase in the bubble phase
Volume of the emulsion phase
Volume of the bubble phase

VPe = Ah(1 e )(1 )

VPb = Ah(1 b )
Ve = A(1 )H
Vb = AH

implement as controller design and necessary troubleshooting are

done on a loop by loop basis. Furthermore, in the APMBC framework, implementing the decentralized control strategy presents
a simpler and smaller parameter estimation problem (i.e. lesser
parameters to estimate compared to the centralized equivalent)
for the RLS algorithm, which is favourable when deploying such
computationally expensive advanced control algorithm.
In the following sections, the theoretical framework of the
proposed mathematical model for the gas phase propylene polymerization uidized bed reactor will be elucidated followed by a
brief outline on the APMBC scheme. Following these, the control
system design for this study will be discussed. The control outcome
for both the servo and regulatory scenarios are tested with performance comparisons made between the APMBC and conventional PI
controllers tuned using the ZieglerNichols (ZN) and the Internal
Model Control (IMC) methods.

 0.429

[7]

[7]
[7]

4.439

[23]
[23]
[23]
[23]

estimating the volume fraction of bubbles in the bed, the voidage of

the bubble and emulsion phases, the gas velocities in the bubble and
emulsion phases, as well as the mass and heat transfer coefcients
for the improved two phase model are summarized in Table 1.
Assuming that the only signicant consumption of propylene
monomer is by propagation reaction, whereas the consumption of
hydrogen gas occurs via transfer to hydrogen, the following expression for the consumption rate of components (i.e. for monomer and
hydrogen) can be obtained:
For monomer:
Ri =

NS


[Mi ]Y (0, j)kp (j),

i=1

(1)

i=2

(2)

j=1

For hydrogen:
Ri =

2. Mathematical modelling

U0 Umf

U0 Umf

NS


[Mi ]Y (0, j)kfh (j),

j=1

In the present study, the kinetic model of propylene homopolymerization over a ZieglerNatta catalyst developed by Shamiri
et al. [22,23] was combined with the dynamic two phase ow
structure proposed by Cui et al. [7,8] to provide a more realistic
understanding of the phenomenon occurring in the bed hydrodynamics. For ease of reference, the correlations required for

The total polymer production rate for each phase can be calculated from:
Rp =

2


Mwi Ri

(3)

i=1

Table 2
Reaction rate constants obtained at 69 C.
Reaction

Rate constant

Unit

Site Type 1

Site Type 2

Reference

Formation
Initiation

kf (j)
ki (j)
kh (j)
kh r
kp (j)

s1
m3 kmol1 s1
m3 kmol1 s1
m3 kmol1 s1
m3 kmol1 s1
kcal kmol1
m3 kmol1 s1
m3 kmol1 s1
m3 kmol1 s1
m3 kmol1 s1
s1

1
22.88
0.1
20
208.6
7200
0.0462
7.54
0.024
0.0001
0.00034

1
54.93
0.1
20
22.8849
7200
0.2535
7.54
0.12
0.0001
0.00034

[23]
[45]
[23]
[23]
[45]

Propagation
Activation energy
Transfer

Deactivation

kfm (j)
kfh (j)
kfr (j)
kfs (j)
kds (j)

[45]
[45]
[23]
[23]
[45]

952

Y.K. Ho et al. / Journal of Process Control 22 (2012) 947958

where Ri is the instantaneous rate of reaction. The reaction rate

constants in Eqs. (1) and (2) were taken from the literature and are
given in Table 2.
In developing the equations for the improved model, the following assumptions were taken into consideration:

For emulsion:
m


Ue Ae (Te,(in) Tref )

m


[Mi ]e,(in) Cpi Ue Ae (Te Tref )

Rv (Te Tref )

[Mi ]e Cpi

i=1
m

i=1

e Cpi [Mi ]e + (1 e )pol Cp,pol

i=1

(1) The polymerization reactions occur in both the bubble and

emulsion phases.
(2) The emulsion phase is considered to be perfectly mixed, and
does not remain at the minimum uidization condition.
(3) The gas in excess of that required for maintaining the
minimum uidization condition passes through the bed as
bubbles.
(4) The bubbles are assumed to be spherical with uniform size
and travel up through the bed in plug ow at constant
velocity.
(5) Mass and heat transfer resistances between gas and solids in the
emulsion and bubble phases are negligible (i.e. low to moderate
catalyst activity) [24].
(6) Radial concentration and temperature gradients in the reactor
are negligible due to the agitation produced by the up-owing
gas.
(7) Elutriation of solids at the top of the bed is neglected.
(8) Constant particle size is assumed throughout the bed.

Based on these assumptions, the following dynamic material

balances were written for all of the components in the bed:
For bubbles:
[Mi ]b,(in) Ub Ab [Mi ]b Ub Ab Rv b [Mi ]b Kbe ([Mi ]b [Mi ]e )Vb
A
(1 b ) b
VPFR

Rib dz =

d
(V [M ] )
dt b b i b

(4)

Hbe Ve

Ve

(1 e )Rie =

d
(Ve e [Mi ]e )
dt

m


(Te Tb ) Ve e (Te Tref )

m


i=1

Cpi ([Mi ]e ) + (1 e )pol Cp,pol

(7)
d
Cpi ([Mi ]e )
dt

d
(Te Tref )
dt

i=1

The initial conditions for solution of the model equations are as

follows:
[Mi ]b,t=0 = [Mi ]in

(8)

Tb (t = 0) = Tin

(9)

[Mi ]e,t=0 = [Mi ]in

(10)

Te (t = 0) = Tin

(11)

3. Outline of the adaptive predictive model-based control

scheme
In this section, a brief outline on the implementation of the
APMBC scheme will be given as this information is readily available
in standard system identication [16,25] and Model Predictive Control [2628] literature. As alluded to previously, the GPC algorithm
devised by Clarke et al. [29,30] employs an internal model for necessary prediction and control calculations. For a Single Input Single
Output (SISO) process, the Controlled Auto-Regressive Integrated
Moving Average (CARIMA) model is the internal process model
used by the GPC algorithm:
T (z 1 )v(k)
1 z 1

(12)

where k = 0, 1, 2, . . ., while T(z1 ) [2931] is selected as unity here

for simplicity. Dening and as known positive integers, a(z 1 )
and b(z 1 ) are polynomials in the z-domain given by:

[Mi ]e,(in) Ue Ae [Mi ]e Ue Ae Rv e [Mi ]e + Kbe ([Mi ]b [Mi ]e )Ve

a(z 1 )y(k) = b(z 1 )u(k Td ) +

For emulsion:

(1 e )Rpe HR

(5)

a(z 1 ) = 1 +

ai z i

(13)

i=1

The direction of mass transfer was assumed to be from bubble to emulsion phase. Furthermore, the energy balances can be
expressed as:
For bubbles:
m


Ub Ab (Tb,(in) Tref )
Rv (Tb Tref )

 mi=1


Ab HR
VPFR

(6)

Rpb dz


m

+Hbe (Te Tb )Vb Vb b (Tb Tref )

Vb

[Mi ]b Cpi

i=1

b Cpi [Mi ]b + (1 b )pol Cp,pol

i=1

+(1 b )

m


[Mi ]b,(in) Cpi Ub Ab (Tb Tref )

m

i=1

Cpi

d
([Mi ]b )
dt

i=1

Cpi [Mi ]b + (1 b )pol Cp,pol

d
(T Tref )
dt b

b(z 1 ) =

bi z i

(14)

i=1

Eq. (12) is used to predict the future response of the process

and this prediction is used to compute the optimal sequence of the
future input trajectory by minimizing a certain form of the GPC cost
function with respect to the control changes, where only the rst
optimized input move is implemented in the process. This entire
sequence of calculation is repeated at every sampling time (ts ) to
continuously produce the optimized controller moves.
It is to be noted that when T(z1 ) = 1 and removing the integrator 1/ on the white noise term, the CARIMA model in Eq. (12)
corresponds to an Auto-Regressive eXogenous (ARX) model. The
existence of the ARX model within the CARIMA model structure
allows the deployment of the RLS algorithm for online estimation
of the polynomial coefcients in Eqs. (13) and (14), i.e. a1 , a2 , . . .,
a and b1 , b2 , . . . b . In this study, the RLS algorithm proposed by
Fortescue et al. [32] with the modication as proposed by Cordero
and Mayne [33] is used for online estimation of the process model

Y.K. Ho et al. / Journal of Process Control 22 (2012) 947958

parameters at a sampling time identical to that used in the controller, i.e. ts . Hence, the internal model of the GPC algorithm is
updated at the same rate by which online estimation is performed.
The form of the RLS algorithm used in this work is referred to as
the Variable Forgetting Factor Recursive Least Squares (VFF-RLS)
algorithm and is given here:
T

(k 1)(k)
(k) = y(k)
(k) =

(15)

p(k 1)(k)

(16)

1 + (k)p(k 1) (k)

(k) = 1

T (k)(k)

 1 + (k)p(k 1)(k)
T

w(k) = p(k 1) (k) (k)p(k 1)

p(k) =

w(k)/(k)
w(k)

where the details of H, K , and Q matrices can be found in Refs.

[2628,35]. Sufce to note here that H (N2 N1 )M , where N1 is
the minimum prediction horizon, N2 is the maximum prediction
horizon, and M N2 N1 is the control horizon.
In GPC, the control move at every control interval is computed
based on the minimization of a quadratic cost function with respect
. Dening W (N2 N1 )(N2 N1 ) and R MM as positive
to u
k1

denite diagonal weighting matrices [2628], the standard GPC

cost function is:

J=

if trace [w(k)/(k)] C
otherwise

1) + (k)(k)

= (k
(k)

(17)

953

r y

r y

+ u T

k1

R u

(26)

k1

In Eq. (26), the weighting matrices W and R are dened as:

(18)
(19)

0
..

by:
T

= [y(k 1), . . . , y(k ), u(k 1 Td ), . . . , u(k Td )]

(21)

0
W
..
.

W =

.
0

(20)

where is used to indicate the accuracy of the identied model (i.e.

modeling error) and  (0, 1]. The regressor vector, and the vec are represented
tor of the estimated process model parameters

..
.
..

.
0

0
..
.

..
.

0
..
.

0 R
, R = . .

. .. ...

0
.
0
W

(27)

The weights W and R, together with N1 , N2 , and M mentioned

previously are tuning parameters for the GPC controller.
In practical process control implementations, the need for
imposing constraints on the controller arises due to valve span limitation as well as the need to ensure the longevity of the lifetime of
the control valve. The GPC algorithm handles process constraints
by minimizing the GPC cost function J with respect to u
while
k1

satisfying the constraints such as those in Eq. (28). The constrained

GPC problem can be solved using Quadratic Programming (QP) [28].
T

= [a1 , . . . , a , b1 , . . . , b ]

(22)

In implementing the VFF-RLS algorithm, it is important to ensure

that the covariance matrix p is always positive denite. To do this,
Bierman [34] stated that if a matrix is positive denite, it can be
decomposed by using the UDUT factorization algorithm, where U
is an upper triangular matrix while D is a diagonal matrix. In this
case, instead of updating p directly, the factorized components of
p (i.e. U and D) are updated recursively. In this way, the covariance
matrix p is assured of its positive deniteness. Hence, the UDUT
factorization is used in all VFF-RLS implementations in this study.
by the VFF-RLS algorithm
To use the estimated parameters
for prediction, rst the CARIMA model for information at the kth
instance is recast in the following form:
A(z 1 )y(k) = b(z 1 )u(k Td ) + v(k)

(23)

where A(z 1 ) = a(z 1 ) = 1 + A1 z 1 + + A z + A+1 z 1 .

The new polynomial coefcients are dened as:
A1 = a1 1
A2 = a2 a1
..
.
A = a a1
A+1 = a

(24)

To enable prediction of the system behavior N2 steps ahead, Eq.

(23) for multiple instances is then assembled into an augmented
form. White noise v(k) in Eq. (23) has zero mean and can be assumed
zero in the future [28]. Adopting the conventions given by [28] and
dening the notation of arrows pointing right for strictly future (not
including current value) vectors while the notation of arrows pointing left for past (including current value) vectors, the augmented
form of Eq. (23) can be written as follows:
y = H u

k1

+ Ku

k1

+Q y

(25)

umin

k1

umax

umin u

k1

umax

(28)

The solution to the QP problem as described above produces

the vector u
where only the rst move from the optimal
k1

sequence is implemented in the process. As mentioned previously,

the entire sequence of calculations as described in this section is
then repeated at every time step.
4. Control system design of the adaptive predictive
model-based control scheme on the polymerization reactor
In the propylene polymerization uidized bed reactor, ve variables were identied as inputs to the process, viz. the supercial gas
velocity (U0 ), catalyst feed rate (Fcat ), hydrogen concentration (CH ),
propylene concentration (CM ), and coolant ow rate (Fcw ), while the
polypropylene production rate (Rp ) and emulsion phase temperature (Te ) were identied as the outputs of the process. Although the
various inputs described had signicant effect on the output variables, not all are suitable to be selected as manipulated variables.
Based on process experience, Te is usually controlled by manipulating Fcw of an upstream heat exchanger (which is used to remove
the heat of polymerization reaction from the recycle gas stream) [2],
whereas the variation in Fcat can be used to control Rp [10]. Hence,
the pairings of Fcat Rp and Fcw Te (hereafter referred to as the Rp
loop and Te loop, respectively) were used to design the decentralized APMBC in this work. Sufce to also note here that Rp and Te
correspond to y in Eq. (12) where as Fcat and Fcw correspond to u
in the same equation. Additionally, U0 , CM , and CH are classied as
disturbance variables in this study. Besides control loops pairing, an
appropriate model order and discrete dead time must be selected.
In this work, the First Order Plus Dead Time (FOPDT) CARIMA model
structure was selected. Although a FOPDT model may not be able
to capture the complete characteristics of higher order processes, it

Y.K. Ho et al. / Journal of Process Control 22 (2012) 947958

344

1.7

Emulsion Phase Temperature (K)

Polypropylene Production Rate (kg/s)

954

1.6

1.5

1.4
APMBC
Setpoint
1.3

APMBC
Setpoint
342

340

338

336

6000

9000

12000

15000

18000

9000

21000

12000

18000

21000

Time (s)

Time (s)
344

1.7

Emulsion Phase Temperature (K)

Polypropylene Production Rate (kg/s)

15000

1.6

1.5

1.4
Setpoint
IMC-Based PI Controller
1.3

Setpoint
IMC-Based PI Controller
342

340

338

336

6000

9000

12000

15000

18000

9000

21000

Fig. 3. Comparison of the performance between the APMBC and IMC-Based PI controller (Kc = 2330%/kg/s, I = 1057 s, D = 0 s, c = 5.4 s) in tracking series of setpoint
changes in the production rate (Rp ).

often serves as a reasonable approximation for the purpose of control system design [36]. In this study, the order of the polynomials
a(z 1 ) and b(z 1 ) in Eqs. (13) and (14) for both loops was selected
to be = = 1. Discrete dead times (Td ) of 2 and 21 were adopted
for the Rp loop and Te loop, respectively. These values of discrete
dead times were calculated based on Eq. (29) given below [37]:
Td = NINT

d
+1
ts

15000

18000

21000

Time (s)

Time (s)

12000

(29)

where the process dead time  d can be found from analysis of the
open loop reaction curve of the process (of which the procedure
is simple and documented in standard control literature [38]). The
sampling time ts , and the design constants  and C for both loops
were chosen to be 5 s, 5 and 6000 respectively based on process
experience. In addition to these, normalized values of the process
inputs and outputs were fed to the VFF-RLS algorithm in this work.
As with all implementations of recursive parameter estimation
techniques, it is important to only begin the online parameter estimation when the inputs and the outputs of the process achieve
steady state. Such a state when the VFF-RLS algorithm is ready to
deploy is referred to as the Design Level of Operation (DLO). As both
the responses of the Rp and the Te loops were sluggish during startup, considerable duration of time (5000 s) was needed for both
process variables to attain a reasonable DLO. The DLO in this study
for both the Rp and the Te loop was therefore selected to be 5000 s.
To allow the performance of the VFF-RLS algorithm to stabilize,

Fig. 4. Comparison of the performance between the APMBC and IMC-Based PI controller (Kc = 18.53%/K, I = 858.08 s, D = 0 s, c = 73.5 s) in tracking series of setpoint
changes in the emulsion phase temperature (Te ).

experience suggests that model adaptation in the GPC controller

for both schemes should be initiated slightly later at 5300 s.
To ensure good performance of the APMBC controller, the minimum prediction horizon (N1 ), maximum prediction horizon (N2 ),
control horizon (M), weight for the output residual (W), and the
move suppression weight (R) must be appropriately tuned. As a
good start, N1 can be selected to be equal to Td + 1. In addition to
this, Shridhar and Cooper [37] suggested that N2 can be selected as
the settling time of the system while M can vary between 1 and 6.
The value of W is usually set to unity, whereas R is usually lesser
than unity. Although these guidelines were good starting values for
tuning the controller, the exact values used in this work were the
result of further ne tuning based on actual control performance. In
addition to the selection of controller tuning parameters, the constraints as given by Eq. (28) were imposed on both loops. The values
of the tuning parameters as well as the upper and lower limits for
the constraints are given here:
Fcat Rp loop
N1 = 3, N2 = 20, M = 4, W = 1, and R = 0.1
12% Fcat 88%
5% Fcat 5%

(30)

Fcw Te loop:
N1 = 22, N2 = 43, M = 2, W = 1, and R = 0.5
0% Fcw 100%
5% Fcw 5%

(31)

Y.K. Ho et al. / Journal of Process Control 22 (2012) 947958

100

100

APMBC

80

80
60
40

60

20
0
12000

12100

12200

12300

40

Controller Moves (%)

100

APMBC

Controller Moves (%)

955

100
80

80

60
40

60

20
0
13500

14500

20

20

0
6000

9000

12000

15000

18000

9000

21000

12000

15000

18000

21000

Time (s)

Time (s)
100

100
100

80

60
40
20

60

0
12000

12100

12200

12300

40

Controller Moves (%)

IMC-Based PI Controller

80

Controller Moves (%)

14000

40

80

60

40

20

20
IMC-Based PI Controller

0
6000

9000

12000

15000

18000

21000

9000

12000

Time (s)
Fig. 5. Comparison of the corresponding controller moves, i.e. the catalyst feed rate
(Fcat ), between the APMBC and IMC-Based PI controller for the production rate (Rp )
loop.

The constraints imposed on the inputs in Eqs. (30) and (31)

were not arbitrary values; but rather, the values were chosen
based on practical considerations acquired through real operational
experience of the author. The valve limits given in Eq. (30) for
manipulating Fcat correspond to a minimum and maximum catalyst
ow rate of 0.0001 kg/s and 0.0007 kg/s, respectively. As for Fcw , the
limits imposed in Eq. (31) were the entire operating range of the
valve. The slew rates of 5% (as shown in Eqs. (30) and (31)) were
adopted for both inputs in this study to ensure smooth actuator
proles.
5. Closed loop performance of the adaptive predictive
model-based control scheme on the polymerization reactor
In this section, the closed loop performance of the APMBC
scheme in tracking series of setpoint changes for Rp and Te in the
polymerization reactor is evaluated at the operating conditions
shown in Table 3. For this purpose, series of setpoint changes in
opposite directions were introduced for both loops. The magnitude of the setpoints introduced for both loops were typical of the
respective nominal operating ranges. The gain for the Rp loop varied
approximately 22% from the lowest setpoint value to the highest
setpoint value, whereas for the Te loop, the gain varied approximately 2.5% from the lowest setpoint value to the highest setpoint
value. For comparison purposes, conventional PI controllers tuned
using the IMC [39] method and the standard ZN [40] method were
included in this simulation. Both the IMC and ZN PI controller

15000

18000

21000

Time (s)
Fig. 6. Comparison of the corresponding controller moves, i.e. the cooling water
ow rate (Fcw ), between the APMBC and IMC-Based PI controller for the emulsion
phase temperature (Te ) loop.

tuning parameters were calculated based on the analysis of the

open loop process reaction curve obtained from a 10% positive step
perturbation from the nominal operating regions of the process
(shown in Table 3), where the open loop response was approximated by a FOPDT model. Particularly for the IMC, the values of the
design parameter c for both loops were selected as the dead time
of the FOPDT model [36]. The FOPDT models used are given here:
Rp (s) [kg/s]
0.042 e5.4
=
1057s + 1
Fcat (s) [%]

(32)

Te (s) [K]
0.315 e73.5
=
858s + 1
Fcw (s) [%]

(33)

Table 3
Operating conditions and physical parameters considered in this work for modeling
uidized bed polypropylene reactors.
Operating conditions

Physical parameters

V (m3 ) = 50
Tref (K) = 353.15
Tin (K) = 328.15
P (bar) = 25
Propylene concentration (kmol/m3 ) = 0.9
Hydrogen concentration (kmol/m3 ) = 0.015
Catalyst feed rate (kg/s) = 0.0003

 (Pa s) = 1.14 104

g (kg/m3 ) = 23.45
s (kg/m3 ) = 910
dp (m) = 500 106
mf = 0.45

Y.K. Ho et al. / Journal of Process Control 22 (2012) 947958

0.090

0.045

0.000

0.000
-0.002
Rp loop

-0.045

Te loop

-0.004
6000

9000

12000

15000

18000

Emulsion Phase Temperature

Prediction Error (K)

Production Rate
Prediction Error (kg/s)

0.002

21000

Time (s)

Polypropylene Production Rate (kg/s)

956

1.440
IMC-Based PI Controller
APMBC

1.435

1.430

1.425

1.420
23000

24000

25000

Fig. 7. Prediction errors for the production rate (Rp ) and the emulsion phase temperature (Te ) loops during closed loop simulations.

Rp (s) [kg/s]
0.0024 e100
=
1819s + 1
Fcw (s) [%]

(34)

Te (s) [K]
0.45 e95.5
=
1755s + 1
Fcat (s) [%]

(35)

From the simulation results, the ZN PI controller produced

chaotic closed loop transients and non-realizable controller moves
for both the Rp and Te loops. Hence, it shall be more meaningful to restrict the following discussion to the APMBC and IMC PI

Emulsion Phase Temperature (K)

Also given here are the Fcw Rp and Fcat Te relationships for the
sake of completeness:

342.4

Polypropylene Production Rate (kg/s)

29000

IMC-Based PI Controller
APMBC

342.0

341.8

341.6
24000

25000

26000

27000

28000

29000

Fig. 9. Comparison of the performance between the APMBC and the IMC-Based PI
controller in rejecting the effect of hydrogen concentration (CH ) on the production rate (Rp ) and the emulsion phase temperature (Te ) loops. A 10% increment in
hydrogen ow rate was introduced at time = 25,000 s.

IMC-Based PI Controller
APMBC
1.435

1.430

1.425

24000

25000

26000

27000

28000

29000

Time (s)
Emulsion Phase Temperature (K)

28000

Time (s)

1.440

342.4

IMC-Based PI Controller
APMBC

342.2

342.0

341.8

341.6
23000

27000

342.2

23000

1.420
23000

26000

Time (s)

24000

25000

26000

27000

28000

29000

Time (s)
Fig. 8. Comparison of the performance between the APMBC and the IMC-Based PI
controller in rejecting the effect of supercial gas velocity (U0 ) on the production
rate (Rp ) and the emulsion phase temperature (Te ) loops. A 10% increment in the
supercial gas velocity was introduced at time = 25,000 s.

controllers only. Figs. 36 show the performance of the APMBC as

compared to the IMC PI controller in tracking series of setpoint
changes for the Rp and the Te loops as well as the corresponding controller moves. The shortcomings exhibited by the ZN PI
controller (not shown), were not observed for the APMBC and the
IMC-Based PI controller. In general, both the APMBC and the IMCBased PI controller were able to track the changes in the setpoint
for both loops. However, upon further scrutiny, the quality of the
setpoint tracking as demonstrated by the APMBC was superior to
the IMC-Based PI controller in terms of the ability to attain minimal overshoot and minimize the effect of loop interactions. As
opposed to the APMBC where the responses for both loops showed
negligible overshoot and minimal effect of loop interactions, the
IMC-Based PI controller produced overshoots for both the Rp and
the Te loops, in addition to the obvious effect of loop interactions
for the Te loop. Moreover, as depicted in Fig. 5, the IMC-Based PI
controller exhibited sudden spikes in the controller moves prole
for the Rp loop. Such sudden aggressive controller moves (i.e. large
slew rates), although infrequent, should be avoided where possible
during practical implementations. The APMBC, on the other hand,
was able to not only produce controller moves which were well
within the specied input constraints for both loops, but also the
controller moves produced were non-aggressive and sufciently
smooth for practical implementations. These were attributed to
the ability of the APMBC to handle constraints in the inputs and
the slew rates, of which conventional PI controllers are unable to
achieve. To summarize, Table 4 shows the integral absolute error
(IAE) for all three controllers in tracking series of setpoint changes
for both loops. The values of the IAE calculated for both loops were

Polypropylene Production Rate (kg/s)

Y.K. Ho et al. / Journal of Process Control 22 (2012) 947958

1.440
IMC-Based PI Controller
APMBC

1.435

1.430

1.425

1.420
23000

24000

25000

26000

27000

28000

29000

Time (s)
100
IMC-Based PI Controller
APMBC

Controller Moves (%)

80

60

40

20

0
23000

24000

25000

26000

27000

28000

29000

957

The results thus far demonstrated the excellent performance

of the APMBC in handling servo process control problems for the
polymerization reactor as compared to conventional controllers.
However, for a particular controller to be practically useful in the
process industry, the controller must be able to handle regulatory
problems effectively. In this case, three variables were identied as
possible disturbance variables, viz. U0 , CH , and CM . A 10% increment
in the nominal value of each disturbance variable was introduced
at time = 25,000 s, following the setpoint test. The disturbance variables were introduced one at a time to observe their individual
effect on the Rp and Te loops. Figs. 810 show the performance of
the APMBC in rejecting the effects of U0 , CH , and CM on the Rp and
the Te loops. To demonstrate the superiority of the APMBC in regulatory control over a conventional controller, comparison with the
IMC-Based PI controller was also included. From the gures, generally the APMBC was able to reject disturbances in a more efcient
manner than the IMC-Based PI controller. For instance, as depicted
in Fig. 8, the APMBC was able to reject the effect of U0 on both the
Rp and the Te loops in a short time span (i.e. 300 s and 600 s, respectively). The IMC-Based PI controller, on the other hand, paled in
comparison as it took 3000 s and 3500 s respectively to bring the Rp
and the Te back to the setpoints. The same trends were observed for
the effects of CH and CM , where the speed of disturbance rejection
for the APMBC was faster than that of the IMC-Based PI controller
for both the Rp and the Te loops. As a general observation, U0 had
the largest effect on both the Rp and the Te loops in addition to the
signicant effect observed in the Te loop when CM was increased.
On the contrary, CH had negligible effect on the Te loop. Other than
the ones mentioned, the rest had marginal effects on the Rp and the
Te loops.

Time (s)
6. Conclusions
Fig. 10. Comparison of the performance between the APMBC and the IMC-Based PI
controller in rejecting the effect of propylene concentration (CM ) on the production
rate (Rp ) and the emulsion phase temperature (Te ) loops. A 10% increment in the
propylene concentration was introduced at time = 25,000 s.

consistent with the previous discussions, where the performance

of the APMBC was superior to the conventional PI controllers. For
interested readers, the poor performance of the ZN PI controller,
which was not shown in the gures, can also be inferred from
Table 4.
Fig. 7 shows the prediction errors (viz. the modelling errors) for
the VFF-RLS algorithm in modelling the Fcat Rp and the Fcw Te relationships. In general, the modelling errors involved were small. The
prediction error for the Rp loop was of the order of one, while the
prediction error for the Te loop was of the order of 103 . Such magnitudes of modelling errors were not sufcient to impair the capacity
of the APMBC to track changes in setpoints efciently. From the
gure, it was shown that the VFF-RLS algorithm was able to track
changes in the process model parameters (i.e. when the process
was perturbed during a setpoint change) effectively. Hence, the
prediction errors moved towards zero after every setpoint change.
In short, the VFF-RLS algorithm was able to reasonably model the
process with time. Moreover, the GPC controller was sufciently
robust to handle slight modelling errors.

Table 4
Integral absolute error (IAE) for the APMBC, IMC-Based PI controller, and the ZN PI
controller in tracking series of setpoint changes for the production rate (Rp ) and the
emulsion phase temperature (Te ) loops.
Controller

IAE for Rp loop (kg/s)

IAE for Te loop (K)

APMBC
IMC-Based PI controller
ZN PI controller

3.84 104
4.49 104
1.32 106

2.46 103
2.63 103
5.56 103

In this study, due to process nonlinearities, proper selection of

the control scheme for implementation is crucial to addressing
the servo and regulatory concerns of the process. In this work, a
decentralized Two Inputs Two Outputs (TITO) APMBC scheme was
implemented, where the VFF-RLS algorithm was used to capture
the dynamics of the process online and subsequently the controller
was designed based on the most recent process dynamics. The
results show that the APMBC scheme demonstrated better setpoint tracking properties as compared to conventional linear PI
controllers tuned using the high performance IMC method and the
standard ZN method. Furthermore, the APMBC was shown to be
able to regulate the effect of process disturbances efciently. In
spite of all the reported benets of the APMBC scheme, the good
performance was obtained at the expense of increased computational load. However, this is nonetheless justiable for such a
critical and dynamically challenging process.
References
[1] K.Y. Choi, W.H. Ray, The dynamic behaviour of uidized bed reactors for solid
catalysed gas phase olen polymerization, Chemical Engineering Science 40
(1985) 22612279.
[2] S.A. Dadebo, M.L. Bell, P.J. McLellan, K.B. McAuley, Temperature control of
industrial gas phase polyethylene reactors, Journal of Process Control 7 (1997)
8395.
[3] A. Sarvaramini, N. Mostou, R. Sotudeh-Gharebagh, Inuence of hydrodynamic
models on dynamic response of the uidized bed polyethylene reactor, International Journal of Chemical Reactor Engineering 6 (2008).
[4] A.S. Ibrehem, M.A. Hussain, N.M. Ghasem, Mathematical model and advanced
control for gas-phase olen polymerization in uidized-bed catalytic reactors,
Chinese Journal of Chemical Engineering 16 (2008) 8489.
[5] A. Shamiri, M.A. Hussain, F.S. Mjalli, N. Mostou, Improved single phase modeling of propylene polymerization in a uidized bed reactor, Computers &
Chemical Engineering 36 (2012) 3547.
[6] M. Aoyagi, D. Kunii, Importance of dispersed solids in bubbles for exothermic
reactions in uidized beds, Chemical Engineering Communications 1 (1974)
191197.

958

Y.K. Ho et al. / Journal of Process Control 22 (2012) 947958

[7] H. Cui, N. Mostou, J. Chaouki, Characterization of dynamic gassolid distribution in uidized beds, Chemical Engineering Journal 79 (2000) 133143.
[8] H. Cui, N. Mostou, J. Chaouki, Gas and solids between dynamic bubble and
emulsion in gas-uidized beds, Powder Technology 120 (2001) 1220.
[9] A.R. Abrahamsen, D. Geldart, Behaviour of gas-uidized beds of ne powders
part: II. Voidage of the dense phase in bubbling beds, Powder Technology 26
(1980) 4755.
[10] K.B. McAuley, J.F. Macgregor, Nonlinear product property control in industrial
gas-phase polyethylene reactors, AIChE Journal 39 (1993) 855866.
[11] B.W. Bequette, Nonlinear control of chemical processes: a review, Industrial &
Engineering Chemistry Research 30 (1991) 13911413.
[12] R. Di Marco, D. Semino, A. Brambilla, From linear to nonlinear model predictive control: comparison of different algorithms, Industrial & Engineering
Chemistry Research 36 (1997) 17081716.
[13] D.E. Seborg, T.F. Edgar, S.L. Shah, Adaptive control strategies for process control:
a survey, AIChE Journal 32 (1986) 881913.
[14] S.L. Shah, W.R. Cluett, Recursive least squares based estimation schemes for
self-tuning control, Canadian Journal of Chemical Engineering 69 (1991) 8996.
[15] K.J. strm, Theory and applications of adaptive control a survey, Automatica
19 (1983) 471486.
[16] L. Ljung, T. Sderstm, Theory and Practice of Recursive Identication, MIT
Press, Massachusettes, 1983.
[17] D.T. Ahlberg, I. Cheyne, Adaptive control of a polymerization reactor AIChE
Symposium Series, 1976, p. 221.
[18] K.M. Kwalik, F.J. Schork, Adaptive control of a continuous polymerization reactor, in: Proceedings of the American Control Conference, Boston, 1985, pp.
872877.
[19] H.-J. Rho, Y.-J. Huh, H.-K. Rhee, Application of adaptive model-predictive control to a batch MMA polymerization reactor, Chemical Engineering Science 53
(1998) 37293739.
[20] S.A. Mendoza-Bustos, A. Penlidis, W.R. Cluett, Adaptive control of conversion in
a simulated solution polymerization continuous stirred tank reactor, Industrial
& Engineering Chemistry Research 29 (1990) 8289.
[21] G.G. Elicabe, G.R. Meira, Estimation, Control in polymerization reactors. A
review, Polymer Engineering and Science 28 (1988) 121135.
[22] A. Shamiri, M.A. Hussain, F.S. Mjalli, N. Mostou, Kinetic modeling of propylene
homopolymerization in a gas-phase uidized-bed reactor, Chemical Engineering Journal 161 (2010) 240249.
[23] A. Shamiri, M. Azlan Hussain, F. Sabri Mjalli, N. Mostou, M. Saleh Shafeeyan,
Dynamic modeling of gas phase propylene homopolymerization in uidized
bed reactors, Chemical Engineering Science 66 (2011) 11891199.
[24] S. Floyd, K.Y. Choi, T.W. Taylor, W.H. Ray, Polymerization of olens through
heterogeneous catalysis: III. Polymer particle modelling with an analysis of
intraparticle heat and mass transfer effects, Journal of Applied Polymer Science
32 (1986) 29352960.

[25] L. Ljung, System Identication: Theory for the User, Prentice Hall, New Jersey,
1987.
[26] E.F. Camacho, C. Bordons, Model Predictive Control, Springer-Verlag, Berlin,
1999.
[27] J.M. Maciejowski, Predictive Control with Constraints, Prentice Hall, Englewood
Cliffs, New Jersey, 2002.
[28] J.A. Rossiter, Model-Based Predictive Control, CRC Press, Boca Raton, FL, 2004.
[29] D.W. Clarke, C. Mohtadi, P.S. Tuffs, Generalized predictive control Part I. The
basic algorithm, Automatica 23 (1987) 137148.
[30] D.W. Clarke, C. Mohtadi, P.S. Tuffs, Generalized predictive control Part II.
Extensions and interpretations, Automatica 23 (1987) 149160.
[31] T.-W. Yoon, D.W. Clarke, Observer design in receding-horizon predictive control, International Journal of Control 61 (1995) 171191.
[32] T.R. Fortescue, L.S. Kershenbaum, B.E. Ydstie, Implementation of self-tuning
regulators with variable forgetting factors, Automatica 17 (1981) 831835.
[33] A.O. Cordero, D.Q. Mayne, Deterministic convergence of a self-tuning regulator with variable forgetting factor, IEEE Proceedings D: Control Theory and
Applications 128 (1981) 1923.
[34] G.J. Bierman, Measurement updating using the U-D factorization, Automatica
12 (1976) 375382.
[35] J. Mikles, M. Fikar, Process Modelling, Identication and Control, SpringerVerlag, Berlin, 2007.
[36] G.H. Cohen, G.A. Coon, Theoretical Considerations of Retarded Control, Transactions of the ASME 75 (1953) 827.
[37] R. Shridhar, D.J. Cooper, A tuning strategy for unconstrained SISO model predictive control, Industrial and Engineering Chemistry Research 36 (1997) 729746.
[38] D.E. Seborg, T.F. Edgar, D.A. Mellichamp, Process dynamics and control, second
ed., John Wiley & Sons, Inc., New Jersey, 2004.
[39] I.L. Chien, P.S. Fruehauf, Consider IMC tuning to improve controller performance, Chemical Engineering Progress 86 (1990) 3341.
[40] J.G. Ziegler, N.B. Nichols, Optimum settings for automatic controllers, Transactions of the ASME 64 (1942) 759768.
[41] A. Lucas, J. Arnaldos, J. Casal, L. Puigjaner, Improved equation for the calculation
of minimum uidization velocity, Industrial & Engineering Chemistry Process
Design and Development 25 (1986) 426429.
[42] D. Kunii, O. Levenspiel, Fluidization Engineering, second ed., ButterworthHeinmann, Boston, MA, 1991.
[43] N. Mostou, H. Cui, J. Chaouki, A comparison of two- and single-phase models for uidized-bed reactors, Industrial & Engineering Chemistry Research 40
(2001) 55265532.
[44] K. Hilligardt, J. Werther, Local bubble gas hold-up and expansion of gas/solid
uidized beds, German Chemical Engineering 9 (1986) 215221.
[45] Z.-H. Luo, P.-L. Su, D.-P. Shi, Z.-W. Zheng, Steady-state and dynamic modeling of commercial bulk polypropylene process of Hypol technology, Chemical
Engineering Journal 149 (2009) 370382.