CHEMICAL ENGINEERING TRANSACTIONS
VOL. 32, 2013
A publication of
The Italian Association
of Chemical Engineering
Online at: www.aidic.it/cet
Chief Editors: Sauro Pierucci, Jiří J. Klemeš
Copyright © 2013, AIDIC Servizi S.r.l.,
I SBN 978-88-95608-23-5; I SSN 1974-9791
Application of Artificial Neural Networks in an Experimental
Batch Reactor of Styrene Polymerization for Predictive Model
Development
Raphael R. C. Santos
*
, Brunno F. Santos, Ana M. F. Fileti, Flávio V. da Silva,
Roger J. Zemp
State University of Campinas, Chemical Engineering Faculty. Cidade Universitária Zeferino Vaz , Av. Albert Einstein, 500 -
CEP 13083-852 - Campinas - SP - Brasil
rrcs.al@gmail.com*
Batch reactors are widely used in the polymer industry, especially for multi-purpose processes where different
types of polymers are produced on demand. Batch polymerization reactors impose rather great operational
difficulties due to the complex reaction kinetics and inherent process nonlinearities. Thus it is a difficult task to
develop mathematical models for polymerization processes. If required for process control purposes the
model should be accurate but simple, so that it can be used in a control loop. The present work shows the
application of the neural network approach in the development of a predictive model for a styrene
polymerization pilot plant, located at the Laboratory of Chemical Systems Engineering, School of Chemical
Engineering at UNICAMP. Artificial Neural Networks have become a usual application in many areas of
engineering, and are well suited for chemical processes due to their ability to describe multi-variable non-
linear models. However to control purposes, the consideration of the variation of the process variables in real
time is required as input to the model, to ensure the representation at the dynamics of the process. The
experimental prototype consists of a jacketed stirred reactor, using thermal fluid as a heat source. Reaction
progress was measured by a density sensor situated in a external recycle loop. Temperature sensors were
positioned both inside the reactor and in the inlet and outlet of the jacket. Traditional feedforward neural
networks with back-step inputs and the Elman network were applied to obtain the best model to be employed
in a control loop. A comparison between the networks was performed, showing that, for process dynamics
modeling, both networks were able to create suitable polymerization models.
1. Introduction
One of the many advantages of batch processes is ability to allow for the change of operating conditions, both
at start up and along the reaction, in order to conduct them to desired products and material properties.
According to Hosen et al. (2011), it is common to manufacture different products in the same equipment, for
instance the polymer industry, where batch reactors are used to produce materials aggregate to demands of
all potential customers in a specific market.
Many batch processes present serious difficulties for control due, especially, to the inherent complexity of the
involved reactions, like the nonlinear behaviour and with properties that vary whole batch period, namely,
require the frequent change of operating parameters (Galván et al. 1997). Lastly, the results are products yield
lower than expected.
Searching for performance improvements of process is usual to develop mathematical models for the dynamic
simulation of the systems, so that studies to improve yield, cost minimization and test of new control
strategies can be performed.
The development of phenomenological models requires much time and can lead to high costs due the range
of operating conditions that have to be considered, and the large number of components and reactions
involved. Also, the additional cost might not be compensated for, as many batch reactors are used for smaller
scale production.
1399