CHEMCON-2014

IICHE
Date:

Santosh Shinde1*, Prashant Shevgaonkar2

1
Chemical Engineering Department
Vishwakarma Institute of Technology
666, Upper Indiranagar, Bibwewadi, Pune 411037, India.
E-mailsantvss@gmail.com; prashant.shevgaonkar@vit.edu

( Research Program: Process Dynamics and Control )

Objective nonlinear system. The purpose is to obtain estimates of

In this work the Unscented Kalman Filter algorithm is
applied for state estimation for solution polymerisation of
styrene in a Continuously Stirred Tank Reactor (CSTR)
using multirate strategy.
Introduction
Estimation of states of a process is of vital importance
for implementation of model-based control strategies.
This is especially true for polymerisation reactors which
present challenges such as changing process conditions,
infrequent availability of vital measuerments such as
various concentrations and polymer moments etc. Soft
sensing state estimation techniques can be of aid for such
processes while implementing control strategies. State
estimator implementations can be used on these processes
to estimate unmeasured or slowly measured states and
the effect of disturbances on-line while taking into
consideration the process-model mismatch. The use of a
state estimator can overcome many of the problems
associated with the control strategies and can make the
closed-loop system more robust to process-model
mismatch. Recursive stochastic state-estimation
techniques, like Extended Kalman filter, are now being
extensively used for the state estimation of nonlinear
systems in chemical processes.
One major drawback of the EKF is that, it assumes
the successive local linearization of the model. This
strategy may not work effectively on highly nonlinear
processes. Thus more robust methods for the estimation
of highly nonlinear processes are needed. One such
technique is the Unscented Kalman Filter.
In this work the Unscented Kalman filter is applied to a
styrene polymerization CSTR, which is a highly
polymer production rate, degree of polymerisation,
polydispersity index etc based on measurements which
are available at a slow rate.
The process for which state estimation was implemented
was solution polymerisation of styrene in a CSTR. The
dynamic process model consisted of a set of eight
ordinary differential equations, the corresponding eight
states being concentrations of monomer, initiator, solvent;
reactor and coolant temperatures; and the first three
polymer moments. It was assumed that six measurements
are available: the two temperatures sampled at a rate of 5
minutes per sample, whereas the solvent concentration
and the three polymer moments sampled at a slower rate
at an interval of 30 minutes. Normally distributed noise
was added to the measurements to simulate process noise.
In the implementation of the UKF algorithm, a total of 53
sigma points were generated every sampling instant.
These were passed through the estimator process model
ODE system which was assumed to be identical to the
process model. Prior mean and covariances were
implemented based on these values. For estimation of
posterior means and covariances a multirate strategy was
used. Since two temperature measurements were
available at every 5 minutes and all six measurements
were available every 30 minutes, the kalman gain matrix
varied between dimensions of two and six. Monomer
conversion, polymer production rate, degree of
polymerisation and polydispersity index were calculated
based on the state estimates every 5 minutes. Root mean
squared error of estimation (RMSEE) was calculated at
the end of the simulation for each state.
Results and discussions

*Presenting author: Santosh Shinde

1
The UKF estimator was tested in various ways : by giving
disturbances to the manipulated and load variables, by
introducing mismatch between process and estimator
model parameters such as polymerisation propagation
rate constant etc. Figure 1 shows the results for a 10 %
reduction in coolant flowrate to the process. As can be
seen the estimator follows the process state changes very
well. The RMSEE for the zeroth polymer moment is
0.50093% whereas the value for the unmeasured
monomer concentrtion is 0.43901 %.
Figure 1 – shows the measured(‘dashed lines) and
estimated values(‘solid lines) of the state variables.
The manipulated variable QC was reduced by 10% of
the steady state value for the plant model. Plots of
state variables versus time,(i) Initiator conc, (ii)
Reactor temperature, (iii) Monomer conc, (iv)
Solvent conc, (v) Jacket temperature, (vi) Zeroth
moment, (vii) First moment, (viii) Second moment.

0.075
Initiator conc(mol/l) vs Time (s) Conclusions
0.07 We have studied the application of Unscented
0.065 Kalman filter for state estimatiom of the styrene
0.06

0.055
polymerization model under different circumstances such
0 1 2 3 4 5
Monomer conc (mol/l) vs Time (s)
6 7 8 9 10
4
x 10
as plant-model mismatch, change in initial covariance,
0.95
step change in manipulated variables and under
0.9

0.85
catastrophic behavior of the plant, it was found that the
0.8 filter performed well under all these conditions making
0.75
0 1 2 3 4 5 6 7 8 9 10 filter a robust filter.
4

Solvent Conc (mol/l) vs time (s)

x 10
Single rate filter was studied in comparison to the
7
multirate filter. Even though the concentrations were
6.5 measured at the of 30 min per sample the filter performed
6
well to estimate the values of this concentrations between
0 1 2 3 4 5 6 7 8 9 10
4
x 10
this interval.
Temperature (K) vs time (s)
356

355
Acknowledgements
354

353
The Authors S.Shinde. and P.Shevgaonkar would like to
352 thank Chemical Engineering Department for their
351
0 1 2 3 4 5 6 7 8 9 10 extended help in completing this work.
4
x 10

References
Cooling Jacket Temp (K) vs time (s)
319

318 [1] V.Prasad, M.Schley, L.P.Russo, B.W.Bequette.

317
(2002).Product property and production rate control of
316
styrene polymerization,Journal of Process Control, 12,
353-372.
315
0 1 2 3 4 5 6 7 8 9 10
Zeroth Moment(mol/l) vs time (s) 4
0.022 x 10

0.02

0.018
[2] S. J.Julier, J. K. Uhlmann. (2004).Unscented filtering
0.016 and nonlinear estimation,Proceedings of The IEEE, 92,
0.014
0 1 2 3 4 5 6 7 8 9 10
401–422.
First Moment(mol/l) vs time (s) 4
0.95 x 10

0.9
[3] L.P. Russo, B.W. Bequette.(1998). Operability of
0.85

0.8
chemical reactors:multiplicity behaviour of a jacketed
0.75
styrene polymerization reactor,Chemical Engineering
Science, 53, 27–45.
0 1 2 3 4 5 6 7 8 9 10
Time(s) 4
x 10
Second Moment(mol/l) vs time (s)
70

65

60

55
0 1 2 3 4 5 6 7 8 9 10
Time(s) 4
x 10