You are on page 1of 4

Computers them. Engng Vol. 22, Suppl., pp.

S981-S984, 1998
0 1998 Published by Elsevier Science Ltd. All rights reserved
Pergamon
Printed in Great Britain
PII: SOO92-b1354(98)00195-l 0098-1354/98 $19.00 + 0.00

MODELING AND OPTIMIZATION OF


PULP AND PAPER PROCESSES USING
NEURAL NETWORKS
Helena Cristina I. L. Aguiar and Rubens Maciel Filho
Universidade Estadual de Campinas - Fat. Eng. Quimica - Depto. Processes Quimicos
Campinas, SP, Brash, CEP 1308 l-970 - CX postal 6066 - e-mail: titina@feq.unicamp.br

Keywords: neural networks, hybrid model, pulping, optimization

ABSTRACT
Two different methods are used to simulate the pulping process in a Kraft digestor - a deterministic model based on
kinetic equations and mass transfer and a neural network model, based on a series of training patterns with usual
process variables as inputs and pulp composition as output. The study indicates which process parameters have
stronger influence on delignification rate, and shows that the neural network and the deterministic model present
equivalent results. 0 1998 Published by Elsevier Science Ltd. All rights reserved.

INTRODUCTION amount of rejects is minimum. A heating zone


Pulp and paper industry is one of the most important follows the impregnation zone, leading the process to
productive activities. As it is the case for a number the suitable temperature for the lignin reaction which
of processes, there is a large incentive for cheaper predominates over the carbohydrates reactions. The
and environmently more friendly operation, but still chips stay in the cooking zone for a period, and then
enhancing product quality. proceed to the washing and discharge zones. The
Wood is a complex structure of fibres glued together liquor is added and extracted from the digester in
by lignin. In order to make pulp, it is necessary to different areas, in order to keep the reaction
separate them, which can be done mechanically, happening properly, and its flow patterns vary from
chemically or a by a combination of the two zone to zone (Assumpclo et a1.,1988).
processes. The focus here is on chemical pulps, more
specifically in Kraft pulps. The Kraft process DETERMINISTIC MODEL
consists of a chemical reaction between the lignin There are several different approaches to the
and a cooking liquor, composed mainly of sodium chemistry of the process in the literature. Most of
hydroxide and sodium sulfide. However, the liquor them use the same general form (Saltin, 1992)
can also attack other wood components, such as although some simplifications and modifications are
carbohydrates (cellulose, hemicellulose and others), usually applied:
degrading them, which is not desirable since they are
the final product. A high quality pulp should have, -dWidt =:(k,[OH’la + k2[OH-]b[HS-]c)(W-W,) (1)
among other properties, fibres in good conditions, for
that will contribute significantly for the resistance of where:
the final paper. W is the wood component concentration
Although the composition of the wood used in the k 1, k2 are the Ahrrenius constants
process will partially determine the properties of the W, is the wood component concentration
paper, these are also altered during the manufacturing after infinite time
process. According to the operating conditions, the a, b, c are the kinectic orders
fibres can be more or less degraded, bleached or [OH], [HS] are the OH’and HS
protected. It will depend on the way in which the concentrations, respectively and t is the time
reactions develop, attacking mostly the lignin, or the
fibres, which is carried out in the digester. Some of them consider a three step reaction, with
The digester is a large pressurized vessel, divided three differential equations of the same type above,
into zones. Although they vary from mill to mill, one for each phase (Gustafsson et al. 1983). Another
digesters basically perform the same tasks. The first approach divide the lignin into high and low
zone is responsible for the impregnation of the wood reactivity lignin, according to the kind of ethers they
with liquor. It is very important that the wood contain (Saltin, 1992 and Mirams and Nguyen,l994)
receives liquor entirely and uniformly, so that at the This study is based on the model developed by
end of the cooking the pulp is uniform and the Mirams and Nguyen (1994) which considers the

S981
S982 European Symposium on Computer Aided Process Engineering--X

simultaneous reactions of two lignins, one reacting ZOH


faster and the other one suffering a slow reaction.
at
There is also a third type of lignin, in a very low
concentration, that does not react with the cooking -3 dL __3 dE j
liquor. +7.0x10
The kinetic expressions for the major X+6’1x10 dt>
carbohydrate components (cellulose and
hemicellulose) consider that these compounds do not
consume the sulphide ion. It is also considered the
presence of two types of cellulose and two types of
hemicellulose. Differently of the most models, this
one considers a simple kinetic equation for the -223xld
extractives. D=60 x 1O-4~~T) x
A good Kraft cook model should include the (10)
transport of the cooking chemicals, expressed in the
diffusion term. The same diffusion coefficient was
used for both, OK and HS-. The thickness is the
critical dimension in Kraft pulping, for it is The ortogonal collocation technique was used to
significantly smaller than length and width and the solve the partial differential equations (the equations
diftusivities are the same in all directions (Jimenez, involving diffusion), and the complete system was
et al., 1989). Therefore, it is usually the only solved by the LSODE integrator.
dimension considered in most pulping models, and so Not only the composition of the wood, and the liquor
in this one. and the diffusion terms are important factors in this
The deterministc model equations can be written as model, but also other characteristics of the wood,
follow: such as porosity and density, as well as the process
temperature. A set of simulation trials was carried
out in order to evaluate the influence of these
EL1 parameters. Two standard patterns (with two
% = kL1 exp(,)[OH]a[HS]bL, (2) different levels of alkali charge) were taken as the
basis, and every parameter was altered, one at a time.
Also, the influence of difusivity on reaction rates was
part of this work.
dL2 EL2 As it can be seen in Figure 1, the chemical charges
- = kL2 exp( ,)[oHlalHslb L2 (3)
dt and temperature are the operational conditions that
seem to have the strongest influences on lignin
reaction rates. However, they have an adverse effect
on carbohydrate consume too, as it is illustrated in
dH1 EHJ Figure 2. Therefore, an optimization of the operating
kH1 exP( =)loHla Hl (4)
dt= policy would be suitable, in order to find the best
combination of reagent concentrations and process
temperature, for a high delignification rate and a low
carbohydrate degradation.
dH2 EH2 The modification of 50% (up and down) in the
-= kH2 eXp( ,,)[oHla H2 (5)
dt diffirsivity value did not seem to alter reaction speed,
as a result of this particular model. Only for low
concentrations of alkali, the difhtsivity change
demonstrated a slight effect on delignification.
dC1 EC1
-= kC1 exp( ,,Wla C, (6)
dt NEURAL NETWORK MODEL
The ability of recognizing and reproducing cause-
effect relationships through training, and filtering
dC2 EC2 noise or irrelevant data, and yet the possibility of
-= kC2 exp( ,,WWaC2 (7) working with multiple input and multiple output
dt
vectors (MIMO), make neural networks efficient in
representing even very non linear systems. That
justifies the choice of neural networks to model the
dE relationship between process variables and final
- = -0.6E (8)
dt product characteristics in complex processes. The
models based on this method may be developed by
using as inputs, variables that are familiar to mill
personnel and easily accessed, instead of parameters
European Symposium on Computer Aided Process Engineering-8 S983

that are not ordinarily measured, in order to predict The results displayed in figures.3 and 4 indicate that
outputs that are usually difficult or expensive to the neural network successfully reproduced the data
obtain. A neural network was developed to model of the deterministic model.
the pulping process, as a mean of comparison with
the deterministic model.
Neural networks are computer programs that Table I- Inputs of the neural network
simulate the learning process of the human brain. Inputs Units
They are also composed by neurons, or process effective alkali 140 g/l of NaOH
elements, that are highly interconnected. In the sultidity 25 %
artificial brain, the neurons are set in layers and may thickness of wood chip 0.003 m
be connected to other neurons of the same or liquor to wood ratio 4 I/kg
different layers through weighted summations and initial temperature 80°C
activation functions of different kinds. A typical cooking temperature 17o*c
neuron is presented in Figure 5. The character of density 0.6 kgdm’
these parameters will differentiate one net from porosity 0.59
another. lignin content 26%
The net developed in the work is a feedforward net hemicellulose content 15.4%
with one hidden layer, that means that the neurons of cellulose content 46.8%
one layer are sending their signals to the neurons of
the following layer. The sigmoid function is used as
0,3 , I
the activation function.
0,25
The ability of learning the relationship input-output is
0,2
conferred to the net by the training step, when a
5 0,15
series of training patterns (input-output sets) are fed
$ 0,l
to the net and the weights in the neurons connections
0.05
are adjusted in order to minimize the error between
-lYh-i 0
the given output and the net output. The error is o~inoom~~-cola
-mmr-m0@4*koolc
backpropagated through the net layers, altering the 7-_-T-

time (min)
weights of each connection proportionally. There are _ Lowconcentration -Standard
several backpropagation training algorithms available ~_ +I 0% effectii alkali - +5% sulfiiity
- +1Ooc
in the literature. The algorithm used in this work is
the generalized delta rule (GDR) algorithm, a kind of Figure 1 - Influence of process variables on
gradient descent method that minimizes the sum of delignitication rate.
squares of the residual (Savkovic-Stevanovic, 1994).
The amount and quality of training sets are essential 0,7
for the success of the artificial neural network
simulations. The number of neurons in the hidden z 0,5
08
layer and the number of hidden layer itself can be P 024
tested in order to obtain better results. Also, the 3 0,3
3 02
learning rate, a coefficient equivalent to a step size
$ 091
and the momentum term which keeps the direction of o-0
descent from changing too rapidly from step to step ii!0 R s 8 s 8
can be altered in order to improve the net efficiency xs,
(Neural Work Sofiare Manual). All these net the (min)
parameters affect not only the accuracy of the
outputs, but also the training time, which can be jz Standard - +1Ooc - -1oocj
considerably long in some cases. Some studies have
been carried out focusing these aspects but they are Figure 2 - Carbohydrates consumption
not shown in this work.
Ordinary industrial process variables, expressed in
mill regular units, and some wood parameters were
used as inputs to the net (table 1). Although the
porosity of the wood is not a frequently measured
value, as well as the wood composition, they should
not vary significantly within the same wood species
and are easily found either in the literature or in mill
process manuals. They are useful parameters when
the network is trained for a more general purpose
such as being appiied to different species.
S984 European Symposium on Computer Aided Process Engineering-4

Both models. deterministic and neural net, produced


similar results, proving that the neural nets can be a
strong tool when the process is partially known, for it
can be developed based only on process variables.
and therefore does not require previous knowledge of
kinetic or mass transfer parameters. On the other
hand, the deterministic model provides a better
process interpretation and it is valid for a wider range
of operation variables, while the neural network is
accurate only in the training region.
-mmr.cn.-m
_ 7 z r-
r z The expected benefit is that the simulation program
training set assists the decision making on process policies
_~__~ according to the desired product characteristics, and
/+dasirad +calculated enables mill personnel to correct the process before
undesired pulp grades are produced. It can also be a
Figure 3 - Comparisom between deterministic model useful tool for training operators and engineers.
results and neural network output for lignin
NOTATION
a,b - kinetic order
t - time
T - temperature
OH - OH- concentration
HS - HS- concentration
Ll e L2 - lignins
Cl e C2 - cellulose
H 1 e H2 - hemicellulose
E - extractives
W - wood component
kW - kinetic constant
EW - activation energy
1~+Cdes +Ccalc ~
D -diffusivity
Figure 4 - Comparisom between deterministic model X - thickness/2
x - distance from the chip center
results and neural network output for lignin
p - density
E - porosity
wlj TJ
a1 \ I
REFERENCES
Assump@o, R.M.V. et al.., 1988. Celulose e Papel -
Tecnologia de fabrica+ da pasta celulbsica,
chapter. VI, Volume 1, IPT.
JimCnez, G., Gustafson, R.R. and McKean, W.T.,
1989. Journal of pulp and paper science, v. 15, no.3.
Mirams, S. and Nguyen, K.Loi, 1994 International
Pan Pacific Conference, Tappi Proceedings, pp 73-
wnJ 82.
Neural Work Sof?are Manual.
Figure 5 - Typical process element in a neural Saltin, J.F., 1992 Pulping Conference, Tappi
network. Proceedings, pp 261-268.
Savkovic-Stevanovic, J., 1994. Computers Chem.
CONCLUDING REMARKS Engngv.18,no.l1/12,pp 1149-1155.
Temperature and concentrations showed to be the
parameters that have the strongest effect on pulping
reaction rates, both for lignin and carbohydrates.
Therefore, in order to get better pulp quality, an
optimization of the process operating policies should
be carried out, aiming for the highest carbohydrate
content and the lower residual lignin. This can be
done by extensive simulation, since the model is
representative of the industrial conditions.

You might also like