Copyright © IFAC Dynamics and Control of Process Systems,

Corfu, Greece, 1998

MODELING, OPTIMIZATION AND CONTROL OF BATCH

CHEMICAL REACTORS IN FINE CHEMICAL PRODUCTION
M .V. LE LANN, M. CABASSUD, G. CASAMAITA
Ecole Nationale Superieure d'Ingenieurs de Genie Chimique
Laboratoire de Genie Chimique -UMR CNRS 5503
18, Chemin de la Loge, 31078 TOULOUSE Cedex 4 - France
Tel : 33 5 6225 23 68 - Fax : 33 5 62 25 23 18,
email: MarieVeronique.LeLann@ensigctJr

Abstract : Operation of a batch reactor covers a wide domain from

optimization of operating conditions to on-line control and monitoring. This
paper focuses on the overall methodology developed in our laboratory for
more than 10 years, in the field of fine chemical industry. This methodology
lies on tendency modeling for stoichiometry and kinetic determination,
optimization of operating conditions, and control of batch reactor by linear or
non-linear Model Predictive Control. Copyright © 1998 IFAC

Keyword : Batch reactor, Modeling, Optimization, Model Predictive

Control, Fine Chemistry

1. INTRODUCTION reactions are usually complex, their kinetics <J'e

With the emergence of generic drugs, pharmaceutical
industries which were not used to suffer from a severe
concurrence are now pushed to compete in order to be
the first on the market. So, they are now interested
in any technique or methodology which permits them
to reduce the time to market. The heart of a drug
manufacturing process is the batch reactor which is
still widely used in fine and pharmaceutical
industries. It is often characterized as a flexible art
multipurpose equipment. That means that a same
apparatus is used to carry out different reactions art
operations under various operating conditions
involving chaining of sequences.
In this paper, we want to focus on the overall
methodology we have developed for modeling,
optimization and control of batch reactor for more
than 10 years . We have successfully applied these
developments to industrial processes particularly in
the SANOFI company with which a long time
partnership has been established ..
Operation of a batch reactor covers a wide domain
from optimization of operating conditions to on-line
control and monitoring. In each of these domains,
computational procedures are generally based on a
process model. The numerical problem of batch
reactor performance optimization has attracted a lot of
attention (Bonvin, 1997). Since fine chemical
poorly known. Classical kinetic studies are not
possible because they are time-consuming in
comparison with the duration of the campaigns and
the economical objectives. An alternative approach is
proposed which aims to conserve the numerical
approach of the optimization, based upon a structured
non-linear model. The model will not be a
fundamental model but a "tendency model" i.e. an
approximate stoichiometric and a pseudo-
homogeneous kinetic model of the transformation
(Filippi, et al., 1986). The tendency model allows to
propose a mathematical support to integrate available
knowledge on the transformation thus providing a
better insight into the process than that offered by the
approach based on a black box model (Rastogi et al.,
1992). The aim of this tool is first of all, to represent
satisfactorily the influence of the most important
operating parameters in order to elaborate an
operating protocol leading to the improvement of a
given criterion. This approach will be detailed in the
first part.
Once the stoichiometric and kinetic models have been
determined, it is therefore possible to focus on the
optimization of operating conditions. Let us notice
that this optimization is rather different from the
design procedure commonly used for continuous
processes. In our case, the industrial reactor in which
the reaction will be carried out already exits. So, it is
necessary to match this requirement by either

715
introducing a perfect description of the apparatus
itself inside the model used for the optimization or by
defining physical constraints relative to the use of
this specific reactor. This optimization procedure is
described in the third part. The success of such
methodology in fine chemical industry depends on the
time needed to carry out the experiments. To
minimize this time we now develop a planning
strategy to minimize the number of experiments. In
the fourth part this strategy is discussed for the
problem of kinetic parameters determination. The
final part deals with the control of the reactor with a
particular emphasis on a new strategy developed for
temperature control and recent experimental results
obtained with Non-Linear Model Predictive Control.
2. STOICHIOMETRY AND KINETICS
MODELING
The minimum necessary and essential data for the
development of the tendency model are : the total
batch time, the reactor temperature, the initial
compositions, the feed rate of the reactant (in the case
of semi-batch operation) and the final concentration
of the different components . In the case where
material losses are observed due to unmeasured
products, the definition of pseudo-constituents allows
one to satisfy the molar component balances.
2.1. Identification of a stoichiometric model
The first step is the identification of an approximate
stoichiometric model of the transformation which
correctly fits the batch data. The stoichiometry of a
reaction system involving NC species Aj (j;I,NC)
and NR reactions Ri (i;I,NR), can be written :
Ne
'vA
L..J .=0I) )
j=l
where vij is the stoichiometric coefficient of Aj in
the reaction Ri (if vij>O then Aj is a product, if
vij<O then Aj is a reactant, if Vij;O then Aj is not)
For a batch reactor and a data base of NE experiments
(1; I,NE) , the number of moles of the component Aj
in the chemical transformation represented by several
reactions Ri, is given by :
NR
njl; nO jl + n° L. VijXil (2)
i;1
where n jl is the number of mole of Aj in the
experiment I, n O jl is the initial number of mole of Aj
in the experiment I, Xii is the extent of the reaction
Ri in the experiment I and n° is a normalizing factor.
Introducing the molar ratio :
n).1 _no )1
Yj1 == Yjl - yO jl == ""':'-n-o""";-
NR
leads to : Yjl = L V ij Xii (4)
i=l
This system may be written in a matrix form as
follows :
(5)
The stoichiometric vectors of each reaction and the
corresponding reaction extents have to be identified.
Two algorithms have been proposed in the literature:
Step by step method
This method was developed by Filippi et al. (1986)
and used for several industrial applications. It
identifies the stoichiometric coefficients of one
reaction at a time. Firstly, the algorithm assumes
that one unique reaction can fit the data. The extents
are computed by the conjugate gradient technique aM
the stoichiometric coefficients by a least square
method. The search is stopped if the representation is
sufficiently good. Otherwise, a new reaction is aihl
and its stoichiometric coefficients are identified,
maintaining those of the previous reaction
unchanged. The number of reactions is increased by
one at each step. The procedure is stopped when the
model accuracy, specified by the user. is obtained.
However, this technique only leads to nonsensical
stoichiometric network. The first reaction is a global
representation of the transformation and the others are
added to complete the mass balances. As result, the
obtained stoichiometric matrix is generally a linear
combination of the real one (Garcia, 1993) and the
difficulty lies in the recombination of the identified
matrix. It is important to notice that at each step, the
operator can propose a stoichiometric vector which is
either in agreement with the computed values (by
0) rounding the real values proposed by the algorithm)
or not. In this way, it is possible to incorporate
available information on the transformation to assist
the determination of the stoichiometric network.
Singular Value Decomposition method
The second algorithm used factorial analysis
techniques to separate stoichiometric vectors aM
extents of reaction from the data matrix Y . Hamer
(1989) has applied this technique to fed batch
fermentation examples. Moreover, as in the previous
method, they must be combined to obtain a realistic
model. Since information about the transformation is
often available from the chemist's experience, an
approach called Target Factor Analysis (TF A)
(Bonvin and Rippin, 1990). This technique enables
one to know whether a postulated stoichiometry from
a priori information is compatible with the abstract
factors. If the postulated stoichiometry is consistent
(3) with the data, the postulated and the transformed
matrix are therefore identical. Otherwise, the target
716
stoichiometry is not consistent with
observations.
In conclusion, with both methods, the main problem
is the combination of the identified vectors so as to
obtain physical meaningful vectors. In both cases,
the problem may be solved by testing the proposed
stoichiometries :
- in the iterative method, by forcing at each step
one reaction of the postulated stoichiometry
(beginning with the principal reaction followed by
the main side-reactions)
- in the factor analysis method by testing postulated
stoichiometry using the TFA .
2.2. Identification of a kinetic model
If the reactor operates in batch or semi-batch mode a
molar balance gives:
dn j
- = Fe + QV (6)
dt J' ~
where nj is the number of moles of Aj at instant t,
Fej is the feed rate of the compound Aj (j=I, NC) CIld
V is the reactor volume.
NR
and we have: Rj = .L Vi/i where
i=]
ri represents the rate of the reactions considered.
In this work, the transformation is supposed to be a
pseudo-homogeneous one and the kinetic law is
written as a classical Arrhenius' law. It is important
to emphasize that the form of the kinetic law and its
degree of complexity depend on the user and the
desired accuracy of the tendency model. So, we have
-E Ne
r I = kOI eRfIlC aj (7)
J
j=!
Preexponential factors (kOi), activation energies (Ei),
order (aij) with respect to each of the reactants for
each reaction of the stoichiometric model and the
form of the kinetic laws are unknown. The orders are
assumed to be part of the data of the problem and are
chosen a priori to be equal to the absolute value of
the stoichiometric coefficients of every reactant. The
problem is thus the determination of the
preexponential factors and the activation energies for
each reaction in order to minimize the difference
between the experimental concentrations and those
computed with the identified parameters for the
different constituents. The identification is carried out
over the entire NE experiments. Classical parameter
identification techniques can be used.
A software named BATC HM OD performing both
stoichiometry and kinetic parameters identification
has been developed in VISUAL Basic and Fortran. It
proposes both methods for stoichiometry research:
the step by step and SVD decomposition . The
identification of kinetic parameters is performed using
SQP algorithm from IMSL. Designed for chemists,
it has been implemented on 4 sites of SANOFI
company: Sisteron, Aramon, M ontpellier am
Budapest ( Hungary).
2.3 Exnmple of complex synthesis mode ling
To examplify the interest of this methodology, its
application to a complex chemical transformation is
presented. The alkaline degradation of fructose may be
achhieved in a simple way by introducing fructose in
an alkaline solution. Nevertheless, the alkaline
degradation of fructose is a non-selective
transformation resulting, by means of a very complex
reaction scheme, in numerous side-products, such as :
- the isomers of the fructose : glucose, mannose,
sorbose, psicose, galactose, tagatose, gulose ...
- the formic , acetic, glycolic, glyceric, lactic,
threonic, 2,4 and 3,4 dihydroxybutyric, 2-
deoxypentonic, saccharinic acids ...
products of degradation such as the
cyclopentadion, furanic compounds, phenolic
compounds, volatile compounds, oligomers,
polymers, coloured products ...
For many years the degradation of fructose in alkaline
medium has attracted the attention of chemists. For
simplification, we can say that the transformation can
be roughly decomposed in two kinds of reactions :
- isomerizations leading to the different
isomers of the fructose (reversible reactions),
- degradations leading to the different acids
(irreversible reactions).
The first step of these two reactions is the formation
of an intermediate called enediol (enediol 2,1 or 2,3
.. . according to which carbon it is attached to). The
enolization may be followed by numerous reactions :
elimination, retroaldolization, rearrangements
Through the enediols, the different acids may
originate from either fructose , glucose, mannose or
sorbose ... An experimental planning has been carried
out by Dubois (1992) to improve the yield of lactic
acid by the degradation of fructose. The experiments
have been run for a total batch time of 9 hours. The
amount of 6 sugars (fructose (F), glucose (G),
man nose (M), sorbose (S), psicose (Ps), galactose
(Gal)) and 5 adds (lactic add (AL), acetic add (AA) ,
formic acid (AF) , glycolic acid (AGlyc) and glyceric
acid (AGlyc) was determined by an HPLIC analysis.
The sugars essentially produced from fructose are :
glucose, mannose, sorbose and psicose. The analysis
of the results issued from the experimental planning
indicates the following favourable conditions for
lactic acid production : a high initial hydroxide
sodium concentration ([NaOH]O = 2m01.l- 1), a low
initial fructose concentration ([F]0 = 25 g.t 1) and a
low temperature (T = 40°C). In optimal conditions,
717
the lactic acid selectivity is equal to 65.8% and its
yield to 64.6%. The experimental data show
sometimes a very important loss in raw materials,
due to the non analyzed products. In order to take this
parameter into account in the model, this quantity
was expressed by the mean of a pseudo-constituent
called P. Its molar concentration was roughly
estimated assuming that the production of lactic,
acetic and formic acid is essentially achieved by
means of a triose and the production of glyceric arl
glycolic acids by means of pentose or hexose ; thus
one obtains :
Cp = COF-[CF + CG + CM + CS o + Cps + CGal +
CAGlyce + CAGlyco + (CAL + CAA + CAF)/2)
Some assumptions were made in order to decrease the
number of constituents present in the mixture. In this
work we have assumed that all the sugars, which leaj
to an enediol intermediate, present similar
sensitivities with respect to the operating variables.
All the hexoses were gathered to form a pseudo-
constituent FS : FS = F + G + M + Ps + Gal.
The secondary acids were also grouped to obtain:
A = AGlyco + AGlyce + AA + AF.
The sodium hydroxide and the lactic acid (product of
interest) were kept unchanged.
This complex system has been reduced to 5
components only : NaOH, FS, AL, A and P. The
tendency model had to establish as well as possible
the influence of the operating variables (temperature,
initial fructose and sodium hydroxide concentration)
on the result of the transformation through the
behavior of these 5 constituents or pseudo-
constituents. For the identification of an approximate
stoichiometric model of the transformation based on
the five components described above, the two
techniques previously presented were used (step by
step and SVD methods) (Garcia, 1993). Since the
combination of the nonsensical reactions is a difficult
problem we have chosen to test an a priori
stoichiometry stemmed from deductive reasoning
based on the available knowledge about the
transformation. The determination of the singular
El =65690 J .mol- 1 E2 =64698 J.morl
kOl =2.42 106 12.mor2 .s-1 k02 =2.00 106 s-1
Table 2 : kinetic parameters identified
3. OPTIMIZATION OF OPERATlNG
CONDmONS
Optimization of an operation in the field of fine
chemistry is characterized by the fact that generally
the apparatus devoted to this operation already exits.
Therefore the optimization step has to include not
only the model of the chemical operation involved
but also a model of the apparatus. Moreover as the
operation are performed in batch mode, the models
718
value by SVD of the data matrix composed by the
dimensionless mole fraction of each constituent gives
: (3 .6478 0.7820 0.2154 0 0.0123). The last
two singular values are much smaller than the
previous ones. This indicates that the data may be
described adequately by three reactions. According to
the available knowledge about the transformation we
have postulated a stoichiometry with 3 reactions :
- the first reaction represents the lactic acid
formation from fructose and hydroxide sodium (one
mole of fructose gives two moles of lactic acid)
- the second reaction represents the unknown or
non-measured by-products formation only from
fructose , without hydroxide sodium consumption.
- in the third reaction the secondary acids are
produced according to a ratio of 2 fructose moles for
three sodium hydroxide moles to form three moles of
acids.
NaOH FS AL A P
RI -2 -1 2 0 0
R2 0 -1 0 0 1
R3 -3 -2 0 3 0
Table 1 : Postulated Stoichiometric matrix
The identification of a kinetic model based upon the
stoichiometry described above, requires the
determination of six parameters: three preexponential
factors and three activation energies. The kinetic law
is expressed as a classical Arrhenius' law and the
orders are assumed to be equal to the stoichiometric
coefficients of reactants in absolute value. The
computed kinetic parameters are given in Table 2.
The stoichio-kinetic model has been identified using
data corresponding only to initial and final
concentrations. Nevertheless, the structure of the
model allows to simulate time evolution of the
concentrations as it has been shown in (Garcia et al.,
1997). Especially a good representation of the time-
evolution of the lactic acid concentration was
obtained.
E3 =78156 J.mol- 1
k03 =2.70 108 14.mo l-4. s-1
reeded are dynamic ones. These two characters have
led to more complex problems in comparison to
optimization of continuous processes which generally
involves only steady-state models of the single
chemical operation. This is the reason why the
optimization of batch process operations in this field
has known development relatively recent compared to
works done on continuous processes (Terwiesch, er
al., 1994 ;Bonvin, 1997). In the literature, works
have been performed on batch processes optimization
by using the well-known Maximum Principle
technique. Nevertheless, as this approach is unable to
handle complex models and constraints, the results
are of poor interest since they are generally not
experimentally applicable. Recent developments on
non-linear optimization allowed to defme a new
methodology in better agreement with the apparatus
limitations .
Optimization of batch reactor consists in the
determination of temperature or additional reactant
rate profiles, or the determination of these two
profties simultaneously. The different variables of the
problem are lower and upper bounded. These bounds
represent some physical limits (for example,
concerning the temperature, the lower bound can
represent the minimum temperature of the reaction,
the upper bound the boiling point of the solvent) .
A classical optimization problem with bounded
and constrained variables takes the form :
Minf(X)
j C(x) = 0
I ~ x~ u
C: R m ----7 RP (m ~ p)
where C is a vectoriallinear/non-linear continuous
application, I and u are the lower and upper bounds.
To treat the optimal control problem by non linear
programming, a finite number of discrete parameters
representing the control function u(t), t E It(), tf]
(temperature or feed rate profties), has to be defined.
For this purpose, the interval [t(),tf] is divided into a
finite number nint of subintervals. In each
subinterval, the control function is represented by a
function f as :
u( t) = f(t,Zj) t E [tj-l , tj] j= 1,nint (9)
The profiles are defined by the parameters Zj and the
switching times tj . In order to avoid too much
sophisticated profiles, f is assumed to be a linear
function versus time in each subinterval. The
temperature and feed rate variations in the range [tj-l,
giVu7.)b~ Zj-l + (t _ t j-l l( ~
tjl :ne Z - Z j-1]
t - t
J J-1
Note that this formulation allows for a discontinuity
and change of functional shape in each subinterval.
The optimization procedure needs a model of the
reactor. In the first studies the model used was
restricted to the mass balance assuming no change of
state, only modeling the kinetic aspect (Eq. 1 to 5
with cv=cl=O). Nevertheless to get feasible operating
conditions it is necessary to consider the limitations
of the heating and cooling capaCities of the system
fitted out the real industrial reactor. These limitations
can be expressed by two ways :
- constraints on the rate of change in the
temperature :
(Zj-Zj-J) - bmax(tj-tj-I> < 0
and (Zj_l-Zj) + bmin(tj-tj-J) < 0 (1)
- a constraint on the rate of the heat generated by
the reactions in order to avoid runaway.
Chemical constraints can be used as well :
- constraint to control the amount of introduced
reactant
- constraints of purity restricting the formation of a
by-product W under a threshold HI:
- constraints on the formation of the desired product
P in order to obtain a concentration above a threshold
The objective function, given by a simulation model,
is an implicit function of the variables. It could be a
simple objective function to be maximized, for
example the concentration of the desired product.
Other types of functions could be chosen, depending
upon the nature of the problem : it could be the
maximization of the yield versus to the total quantity
of reactant initially present in the reactor or versus to
the total quantity of added reactant. Some other
objective functions including economic factors may
(8)
be interesting, for example, when the cost of the
reactants is the controlling factor or when it is
economically advantageous to minimize the operation
time or the operating costs .... The problem is now to
compute the variables Zj (temperature and feed rate) at
each bound of each subinterval, that minimize the
criterion subject to the bounds and imposed
constraints. The optimal control problem is solved
by nonlinear programming . A sequential quadratic
programming (SQP) procedure (Gill, 1985) has been
used as it is considered as the most effective method
for NLP problems. These studies led to the
development of a software called OPTIBATCH.
Example of chemical synthesis optimization
This example corresponds to the optimization of the
lactic acid synthesis which the stoichiometry and the
kinetics identification has been reported in paragraph
2.3 . The main objective was to simultaneously
improve the lactic acid yield and to minimize the
(10) unknown by-products and acid by-products formation.
The objective function to maximize used is the
C LA (2)
following : c=
CA +C p
In the conditions determined by Dubois (992) (batch
and isotherm at 40°C) the lactic acid yield and
selectivity are :
R~P = 64% and S ~P = 65 .8%
which leads to a value of the criterion described above
equal to 2.577. One or more variables can be
optimized using the optimization algorithm . The
temperature and the feed rate of fructose will be
computed in order to improve the previous criterion.
The hydroxide sodium is kept at it maximum value
in the experimental range, (i.e. 2 moU I ) . The total
batch time is equal to 9 hours. The reaction volume
719

is equal to 1 liter. The amount of fed fructose is fixed
to 25g. The following constraints and bounds on the
variables were employed : maximal temperature
variations: ±2°C.min- 1, temperature bounds :
400 C ~ T ~ 100°C
and feed rate bounds : 0 ~ F ~ 10-4 kgs- 1 .
The optimal conditions obtained are a fed batch
operation with feed rate of fructose associated to an
isothermal operation at the maximum temperature
100°C. For these conditions one can notice a very
important improvement in the criterion, since it has
improved from 2.577 to 5.785. The positive effect of
a low initial fructose concentration, predicted by the
analysis of the experimental planning, suggests that
the fructose may be introduced in semi-batch mode to
favor the lactic acid production. The results of the
optimization confirm this tendency.
The reaction was camed out under these optimal
operating conditions. A syrup with 75 % of fructose
in water was fed during 9 hours (which corresponds to
25g of fructose . The theoretical and experimental
yields and selectivities (expressed in %) are given in
Table 3. Rather better experimental results that those
predicted by the tendency model were observed.
RE;.P Rt."'}. st;,P s[;
Previous conditions 64.6 59.8 65.8 60.8
(batch - 40 0c)
Feeding of fructose 76.3 74.1 83.0 74.4
(maximum isotherm)
Table 3 : Yields and selectivities in lactic acid
4. EXPERIMENTAL PLANNING STRATEGY
In the previous example, the optimization strategy
was developed on the basis of existing data. Then
arised two novel questions:
-Is-it possible to minimize the number of needed
experiments?
- Is-it possible to design experiments affording
maximum information for model identification ?
Both can be achieved by developing experimental
planning strategy . This approach consists in the
determination of the experimental operating
conditions in such of way to maximize the confidence
in the estimated parameters (Walter and Pronzato,
1990). In practice this objective corresponds to the
minimization of the confidence region volume of the
parameters which can be expressed as :
&- O*)T MF (~, 0)(0 - 0* ~ const
where * is the parameter vector estimated for the
e given temperature. The latter is based on heating or
operating conditions ~ . and M F the Fisher' s
information matrix given by:
MF(~' e) = I (a ll(~i' 9)]T I:-l(~d (a T](~i ' e)]
i=! a9 ae (14)
There are several possibilities to express this
optimization problem . One solution generally
adopted is the maximization of the M F determinant.
Therefore, the determination of the optimal operating
conditions ~D for parameters estimation can be
expressed as :
~D = Arg max ret(MF(~ ' 8) (5)
.;
This objective criterion is a non-linear function of the
unknown parameters. So, a classical approach
consists in alternating the estimation step procedure
and the procedure of experimental operating
conditions determination. This approach is called the
sequential experimental design strategy in which it is
necessary to include all the constraints limiting the
feasible experimental domain . For example, the
minimum time needed between two consecutive
measurements, the maximum time of the experiment
and also all the technical constraints relative to the
reactor in which the experiments are carried out (such
as maximum heating or cooling rate and/or
maximum reactant feeding, maximum released heat
capacity).
5. CONTROL OF BATCH REACTOR
If we consider the optimization step described above
the control of batch reactor should be a temperature
control coupled with a concentration control.
Nevertheless excepted a few applications it appears
generally impossible to implement an on-line control
of concentration. This is due essentially to the lack of
on-line concentration measurements. Even if new
methods have revealed promising (Infra Red
spectrometry measurements (Amrhein et al ., 1996 ;
Wise and GalIagher, 1996» , the problem of batch
reactor control remains essentially a problem of
temperature control (Friedric and Perne, 1995). Thus,
the control performances are mainly dependent on the
heating-cooling system associated with the reactor
(Berber, 1995). A lot of studies have been performed
on control problems and strategies in such a type of
reactors. Chylla and Randall Hasse. (1993) and Juba
and Hamer. (1986) give a good overview on the
challenges in batch reactor control and suggest several
control strategies.
Let us only recall that the two widely used heating-
cooling systems in industry are the multifluid and the
(13)
monofluid. The former one consists in delivering to
the jacket alternatively utility fluids available at
cooling of a single thermal fluid delivered to the
jacket.
720
In the field of fine chemistry the ideal controller
should be as flexible as the apparatus. In such a case
the controller tuning will be attached to the reactor
itself and should be able to handle different operating
conditions or chemical reactions .
5.1 Linear Model Predictive Control
The temptation in the control of batch reactors has
been to use the simplest model i.e. the classical
input-output model (also called the black box
model). In the first works performed in our
laboratory related to the thermal control of batch
reactor, we applied the adaptive Generalized
Predictive Controller with Reference Model
(GPCMR) on pilot and industrial reactors (1000 I
glass-lined reactor (Le Lann et al., 1995» . The pilot
plants' results were very satisfactory as we were able
to operate the reactor for different conditions
(temperature profile, chemical reactions , .. .) with the
same controller design parameters (Le Lann et al.,
1995).We thought that we had reached our objective
to develop a controller as flexible as a batch reactor.
But, the results were not as good as expected for the
industrial reactor fitted out with the multifluid
system. It revealed that the black box model didn't
model the transfers between the reaction mixture arx1
the jacket. A particular problem was the "bain-marie"
effect (effect corresponding to a zero utility flow rate
but with a jacket full of fluid). Let us recall that in
the case of the multifluid system the manipulative
variable is the fluid flow rate. The heat transfer
between the jacket and the reaction mixture
represents the main limiting dynamics in an
industrial-scale reactor.
Realistic Model Based Predictive Controller
A novel predictive controller called "realistic model-
based controller" was developed (Le Lann et al.,
1995). This controller is always based on the GPC
structure but used a nonlinear model which is
obtained from the heat balance on the reaction
mixture (Eq.16) and the jacket (Eq.17).
=Urjarj~j - }~H'+
mrCPr : r Tr
FcCpc (Tc - Tr ) (16)
j Classically this supervision was performed using
m.Cp . df =F. p.Cp.fr. -T.)-
J J ci J J J ~m J
Urj arj ~j - Tr ) (17)
The manipulative variable was the flow rate delivered
to the jacket. In such a case, as shown by the model
the manipulated variable is not a classical one. The
final temperature of the external fluid is independent
of its flow rate, and thus, without a chemical
reaction, the reaction mixture will tend to the
721
temperature of the external fluid . The value of the
flow rate doesn't influence this final state but affects
the "speed" at which the final state is obtained. The
manipulative variable is in fact the time constant of
the process. This is true for all processes where the
manipulated variable is the flow rate (i.e. heat
exchangers for example . The parameters of this model
was identified on-line by using all the available
measurements such as : temperature inside the jacket.
In practice, this to identify the heat transfer
coefficient. An estimator of the heat generation rate
was also developed from past values of reactor arx1
inlet/outlet jacket temperatures. This value, supposed
to be constant on the prediction horizon, was then
used in the model (term ~H') . Such a technique
allows also to elaborate a supervision structure which
was able to predict when the valve of hot fluid has to
be cut-off to prevent from temperature overshoot at
the end of the preheating phase (Le Lann et aI. ,
1995).
Controller based on the thermal flux analysis
By considering Eq.16 one can notice that « the
driving force » for a temperature change in the reactor
is the term U rj a rj ~j - Tr ) which represents the
thermal flux exchanged between the reactor mixture
and the jacket. A new strategy for temperature reactor
control was developed by considering this thermal
flux as the manipulated variable. At a first level, a
predictive controller determines the thermal flux to be
exchanged for temperature set-point tracking. This
thermal flux is then used in a cascade-like structure to
determine the real manipulative variable : valve
opening, heating power ... This structure is based on
a thermal model of the exchanges between the reactor
and its jacket and depends on the heating-cooling
system used. A considerable improvement in
comparison to classical strategy is brought
concerning the supervision of the heating-cooling
system . In fact, in addition to the temperature control
problem itself there is always a problem of
supervision whatever the heating-cooling system is .
Indeed it is necessary to choose on-line either the
right utility fluid in the case of multifluid system or
the thermal device to be used (heat-exchanger, heating
device, ... ) in the case of monofluid system.
temperature alarms or a priori defined range of
operation for each device. With the thermal flux
analysis strategy it is therefore possible to compute at
each sampling period the minimum and maximum
thermal exchange capacities of each fluid or each
device depending on the system used. According to
these thermal flux limits, it is therefore possible to
determine which fluid or device is able to
deliver/remove the thermal flux computed by the
master controller. The choice is thus done properly as
it is based on the thermal capacity and not on a
predefined temperature. This strategy can be applied a
priori to any heating-cooling system.
In parallel we developed a new heating-cooling
system called hybrid system which mixes both
mono fluid and multifluid systems . The objective was
to define a thermal system which takes the advantages
of both systems. The multifluid is many used in
industry as it demands a very low cost to be settled
and offers a large heating (cooling) capabilities of
steam (glycol) when rapid heating (cooling) is needed.
Nevertheless, it makes temperature control difficult
due to changeover of fluids , to a non classical
manipulative variable as it has been explained
previously. The monofluid system is therefore more
convenient in the controllability point of view since
it is better to use a single fluid circulating at a
sufficiently rapid flow rate to get a good heat transfer
coefficient. Nevertheless, it remains expensive since
as it has to be used in a large range of temperature,
the thermal loop is often fitted-out with expensive
devices as heat-exchangers, .... In an other hand it is
unrealistic to think that industry will replace every
multifluid system (which fits 90% of the batch
Air
--1:><:],-----
I
Cold
w water
Figure 1 : Scheme of the 16 liters pilot plant fitted out with the hybrid heating-cooling system
Temperature (0C) Valve opening B
70
FTr"I j O.5
60 L=2J·i 0.4
l~pl 103
50
40
30
20
° °
0 , •.••••••••••••• 0• ••••••••••• • ••••
:WOO 4000 6000 8000
Time (5)
Fig . 2 : Temperature and manipulated variable
reactors) by monofluid one. Therefore, the solution
we adopted was to use a monofluid configuration
called intermediate fluid which is obtained by mixing
steam and cold water while this is possible (in
practice from 25°C to 70°C) and to switch to steam
only when more heating is needed or to glycol-water
in the cas of high cooling (Louleh et al.; 1997).
This system presents the advantages of a multifluid
thermal system (possibility to use utility fluids
directly available on a plant) and those of monofluid
thermal system (characterized by high dynamic
performance and continuity of thermal control). The
switching condition is given by the on-line
computation of the thermal exchange capacities of
each configuration (intermediate fluid, steam, glycol
water) determined as previously described. More
details on this procedure can be found in (Louleh et
al. , 1997). The strategy based on the thermal flux
analysis has been applied to a 16 liter-pilot plant
fitted out with this system. On figure I a scheme of
this pilot plant is given. An example of results of a
temperature control experiment performed on this
pilot plant using the thermal flux analysis strategy
are plotted on figure 2.
GlycolfWater
Tu gw
Tuc
Thermal flux (KcaI.5-1)
5 5
4
3 r---------------------------~ 3
---(lc('llt --. -· Qrn;lx if - . ~ . Q['j'I:~" :
-Qmaxvap ·· · · -Qmmi f 2
2 -Qminvap ...,' - ~):I:.L\!;'h
.i
1
°. ,
.-
o..J : I
-2 -2
2000 4000 6000 8000
Time (5)
Fig . 3 : Limits and control value of the thermal flux
722
This figure shows the evolution of the temperature in
the reactor. the set point and the valve opening degree
(in the case of no recycle of the intermediate fluid). A
good tracking performance of the cascaded model
based controller can be noticed. Figure 3 gives the
evolution of the thermal flux computed by the master
predictive controller (GPCMR) and the upper m
lower flux limit capacities determined on-line for each
configuration. One can see that all the different
configurations have to be used. In particular. at the
end of the heating phase. steam is required instead of
intermediate fluid. Similarly. during the cooling
phase (around 8000 s). the higher cooling rate
implies a changeover from intermediate fluid to
glycoUwater. This experiment shows the capacity of
such a methodology to handle changeovers.
5.2 Non-Linear Model Predictive Control
As a logical continuation to our studies on Linear
Predictive Control and to our works on off-line
operating conditions optimization. we developed a
generic methodology of Non-Linear Model Predictive
Control. This NLMPC is based on the minimization
of a quadratic criterion (Eaton and Rawlings. 1990 ;
Biegler. 1997):
p 2
Min J = I(Y c(k + e + i) - ypred(k + e + i») (18)
u(I) .... u(N) i=l
subject to constraints:
- Upper and lower bounds on manipulated variable
umin(i) ~ u(i) ~ umax(i) i = 1..... N (19)
- Upper and lower velocity bounds on manipulated
variable
u(i-I) - ~umax(i) ~ u(i)~ u(i-I) + ~umax(j) i = 1, .. .,N
(20)
where Yc(.) represents the values of the reference
trajectory (the desired value for the output), Ypred (.) is
the value of the output predicted by the model, u(.) is
the future value of the manipulative variable (they are
the decision variables in this optimization problem).
The values predicted over the time horizon P are
obtained by the integration of an ODE or DAE
system. This strategy has been applied to the control
of temperature in a batch reactor fitted out with the
multifluid system. In fact, which such a system
there were cases where the control of the reactor
temperature can only be achieved by partly filling the
jacket. Therefore in such cases. the filling-up ratio of
the jacket becomes a manipulated variable . The two
equations (Eq. 16 and Eq . 17) are no more sufficient
to model the dynamic behavior of the reactor when
the jacket is not totally full . To model the dynamics
in this case. we have adopted the methodology
formerly developed for batch reactor simulation
(Cabassud et al., 1994). The jacket is lumped in
several tanks . Heat transfer between each jacket-tank
and the reactor is modeled by an energy balance
723
including the filling-up ratio of the tank and the
nature of the fluid (air or utility fluid) . This technique
leads to a ODE system which is solved over the
prediction horizon at every optimization iteration
step. A procedure for model error compensation and
an estimator of the heat generation rate have been
used for eliminating process-model mismatches
(Ettedgui et al. , 1997) .
The main advantage of such technique lies on
considering the phases of jacket filling-up and jacket
purge as any normal operation phase. Before using
this model it was not possible to control the reactor
during theses particular phases and consequently it
was very difficult for example to control temperature
at a constant value when the reaction was weakly
exothermic and the reactor was well insulated. This
technique, firstly tested in simulation (Ettedgui et al.,
1997) has been implemented on a pilot plant (the 16
liter pilot plant described previously but modified to
get a pure multifluid system). Figure 4 shows an
example of experimental results obtained with such a
strategy. One can see that the temperature tracking
has been successfully performed especially during the
constant level temperature step. One can notice that
the predictive controller is able to perfectly anticipate
the changes in the temperature profile and
consequently it can insure a smooth control without
temperature overshoot. This success is mainly due to
the accuracy of the model used for the prediction. An
automatic supervision technique is being developed.
It is based on a similar procedure to the one described
previously : on a prevision horizon, the maximum
and minimum temperature reachable with all the
fluids are computed .It is therefore possible to
determine if the present fluid is accurate or if a
changeover is needed and what fluid can be used. In
such a case, this information is used in the predictive
controller for the computation of the control law.
Temperature ( cC) Manipulative
65r-~-.~~~~~.-~-.~~~v~a~r~ia~ble
55
0.8
45 0,6
35 0,4
0,2
25
o
600 1200 1800 2400 3000 3600
Figure 4 : Experimental results of NLMPC applied
to the 16 liter pilot plant (with multifluid system)
6. CONCLUSION AND RJTURE WORKS
This paper has surveyed more than 10 years of
development we had in modeling, optimization and
control of batch reactors in the field of fine chemicals
industry. These works have been performed in a close
collaboration with industrial partners which permits
to the use of the developed methodology in industry.
Our present studies deal with the development of the
experimental design strategy in order to reduce the
experimental effort. This is a key point to make this
methodology widely applied in fine chemical
industry.
In an other hand. in our previous studies on batch
reactor control, it was assumed that no chemical
reaction model was available. The chemical reaction
was considered as a disturbance. The future works
will focus to include tendency models in the model
based controller. Then, a state estimator based on
temperature measurements coupled to an estimator of
the reaction heats will be developed.
We aim to adapt this methodology to other industrial
fields : biochemical, pharmaceutical.
REFERENCES
Amrhein M., B. Srinivasan, D. Bonvin and M.
Schumacher (1996), Inferring concentrations on-
line from near-infrared spectra : Nonlinear
calibration via mid-infrared measurements,
Comput. Chem. Engng, 20 , S975-980
Berber R.( 1995), Control of batch reactors : a
review, Methods 0/ Model Based Process Control,
R. Berber (ed.), Kluwer Academic Publishers,
Oordrecht, 459-493 .
Biegler L. (1997), Advances in Nonlinear
Programming Concepts for Process Control,
Proceedings o/the lFAC ADCHEM '97. Banff
(Canada), 587-598
Bonvin D. and Rippin D .W.T .. 1990, Target factor
analysis for the identification of stoichiometric
models. Chem. Engng Sci., 45 (12), 3417-3426
Bonvin D. (1997) Optimal operation of
Discontinuous Reactors-A personal view
Proceedings o/the lFAC ADCHEM'97. Banff
(Canada),155-169
Cabassud M. , M.V. Le Lann, B. Euedgui and G.
Casamatta (1994) A general simulation model of
batch chemical , Chem. Engng Technol.,17 (4) ,
255-268
Chylla R. W. and D. Randall Hasse( 1993),
Temperature control of semi-batch
polymerization reactors, Comput. Chem. Engng
17 W 3, 257-264.
Dubois M.F(l992), Valorisation chimique du 0-
Fructose: synthese des acides levulinique,
arabonique et lactique, PhD Thesis, I.N.P.
Toulouse
Eaton J.W . and J.B . Rawlings (1990), Feedback
control of chemical processes using on-line
724
optimization techniques, Comput. Chem.
Engng., 14, 469-479
Ettedgui B., M. Cabassud, M.Y. LeLann, NL
Ricker (1997)Regulation of a Fed-Batch Chemical
Reactor Incorporating Parameter Estimator,
Proceedings o/the lFAC ADCHEM '97, Banff
(Canada), 365-370
Filippi c., JL Graffe , J. Bordet, J. Villermaux, J.L.
Barney, P. Bonte and C. Gerogakis (1986).
Tendency modeling of semi batch reactors for
optimization and control, Chem. Engng. Sci .,
41 , 913-920
Friedric M. and R. Perne, 1995, Design and control
of batch reactors : an industrial viewpoint,
Comp. Chem. Engng 19, S357-S368 .
Garcia V. (1993), , PhD Thesis, I.N.P. Toulouse
Garcia V., M.Cabassud, M.V. Le Lann., L. Rigal,
G. Casamatta (1997), A global strategy for the
optimization of batch reactors : Application to the
lactic acid synthesis by the alkaline degradation of
fructose, lMACS (M2SABI) Mathematical
modelling and simlation in agricultural and bio-
industries, Budapest (Hongrie), , in Proc. 63-68
Gill P.E.,W. Murray, M.A. Saunders and M.H.
Wright (1985) Software and its relationship to
methods in numerical optimization, PT.
Boggs, R.H . Byrd and R.B . Schnabel, Eds.
SIAM, Philadelphia.
Hamer J.W .(989), Stoichiometric interpretation of
multireaction data : application to fed batch
fermentation data, Chem. Engng Sci., 44 (10) ,
2363-2374
Juba M.R. and J.W. Hamer, 1986, Progress and
challenges in batch process control. Proc. CPC
lll. CACHE, Elsevier, Amsterdam.
Le Lann M.V ., M . Cabassud and G. Casamatta,
Adaptive Model Predictive Control (1995), in R.
Berber (ed.) Methods 0/ Model Based Process
Control, Kluwer Academic Publishers,
Dordrecht, pp. 426-458, 1995
Louleh., Z . Cabassud, M.V. LeLann and G .
Casamatta (1997) and G. Casamatta Experimental
Study of a New Thermal Control Strategy for
Batch Reactors, Proceedings o/the lFAC
ADCHEM'97, Banff(Canada), 119-124
Rastogi A., J. Fotopoulos, C. Geaorgakis and H.G.
Stenger (1992), The identification of kinetic
expressions and the evolutionary optimization of
specialty chemical batch reactors using tendency
models, Chem. Engng., Sci ., 47 , 2487-2492
Waiter E. and L. Pronzato,(1990) Qualitative and
Quantitative Experiment Design for
Phenomenological Models-A Survey
Automatica, 26, (2), pp. 195-213.
Wise B.M. and N.B. Gallagher (1996) The process
chemometrics approach to process monitoring aM
fault detection, J. Proc. Control, 6, 329-348