Hybrid modeling for continuous production of bioethanol

16th European Symposium on Computer Aided Process Engineering
and 9th International Symposium on Process Systems Engineering
W. Marquardt, C. Pantelides (Editors)
© 2006 Published by Elsevier B.V.
Hybrid modeling for continuous production of
bioethanol
E. Ccopa Rivera, Ivana Mantovaneli, Aline C. da Costa and R. Maciel Filho
Faculty of Chemical Engineering, State Universtiy of Campinas, Campinas, SP, P.O.
Box 6066, 13081-970, Brazil
Abstract
Changes in the operational conditions are quite common in plants of alcoholic
fermentation; they occur not only due to the variations in the quality of the raw material
but also due to variations of dominant yeast in the process. Thus, it is clear that the main
difficulty in model-based techniques for definition of operational strategies, control and
optimization, is the problem of obtaining a reliable model. This work proposes the use
of hybrid models for continuous production of bioethanol plants. The hybrid model
comprises first a system of differential equations describing the mass and energy
balances of the process. Complementary to the mass and energy balances, the kinetic
rates of the biological system are obtained by neural network sub-models. Multilayer
perceptron neural networks for an Extractive Alcoholic Fermentation Process and
Functional Link Networks for an Alcoholic Fermentation Process with Multiples Stages
were used for modeling the kinetic. The results show that typical non-linearities present
in this kind of bioprocesses are adequately represented. Agreement between predicted
and operating data was very good. The models are reliable enough to be used for further
optimization and control studies.
Keywords: Bioethanol, Hybrid modeling, Multilayer Perceptron Neural Network,
Functional Link Network.
1. Introduction
A potential substitute for petroleum in Brazil and in several countries may be biomass,
particularly sugar cane. The sugar cane industry keeps the greatest commercial energy
production in the world with ethanol and the almost complete use of sugar cane bagasse
as a fuel. Brazilian annual production capacity of ethanol is currently in 18 billions of
liters and, with the increase in demand, the idea is to multiply this capacity through the
installation of new factories and optimization of the operation of the existing ones.
Thus, there is an intensified interest in the study of all the steps involved in ethanol
production.
Among the main problems related to the alcoholic fermentation process is the lack of
robustness of the fermentation in the presence of fluctuations in the quality of the raw
material and modifications in microbial metabolism. These lead to changes in the
kinetic behavior with impact on yield, productivity and conversion of the process. The
lack of robustness can be corrected by adjustments in the operational and control
parameters of the process when fluctuations occur. In order to accomplish this, it is
important that a mathematical model be available to aid in the decision making, mainly
when the difficulties of monitoring the key process variables (concentrations of
biomass, substrate and ethanol) are taken into account. However, the operational
changes described make the prediction of the dynamic behavior of the process with a
single model difficult, as they lead to changes in microorganism kinetics. Thus, it would
613

614
be of great advantage to have a mathematical model that could be easily adapted to
changes in operational conditions. A way to deal with this problem is to use hybrid
neural models. These models are expected to perform better than “black-box” neural
network models (pure ANN models), since generalization and extrapolation are
confined only to the uncertain parts of the process and the basic model is always
consistent with first principles. Besides, significantly fewer data are required for their
training. Several authors have applied hybrid modeling for bioprocesses (e.g. Pischogios
and Ungar, 1992; Zorzetto et al., 2000) In particular Harada et al. (2002) and Costa et
al. (2001) dealt with the hybrid modeling of the alcoholic fermentation process.
The objective of this work is to investigate the use of hybrid models to describe
continuous production of bioethanol in order to obtain a fast and easy model to be used
for on-line implementation. Hybrid models have been developed based on first
principles (physical knowledge) associated with Artificial Neural Networks (ANN).
Multilayer Perceptron Neural Networks (MLPNN) were used for an Extractive
Alcoholic Fermentation Process and Functional Link Networks (FLN) were used for
modeling the kinetic of an Alcoholic Fermentation Process with Multiples Stages.
Hence, in this work a tool is developed to build up hybrid models with relatively
smaller amount of data compared to the “black-box” approach required for process
identification exploring the benefits of two different ANN structures.
1.1. Hibrid Models
Psichogios and Ungar (1992) were among the first who proposed the serial approach for
hybrid modeling of bioprocess. The kinetics relations are modeled by a static mapping
of process states to biomass specific rates. Thus, the function to be learned by the ANN
is static, while the process dynamics are represented by the differential equations of the
model. As noted by Psichogios and Ungar (1992), the hybrid approach provides better
predictions than “black box” modeling of bioprocesses with a pure neural network
approach.
In this work, it was used two different ANNs: MLPNN, which are able to approximate
generic classes of functions, including continuous and integrable ones, and FLN, which
have good non-linear approximation ability with the advantage that the estimation of
their weights is a linear optimization problem. The use of the MLPNN structure has
been proved to be a good approach (Zorzetto et al, 2000), however there is a large
incentive to use ANN structures that are quicker to be identified and easier to be
implemented, such as the FLNs. This is particularly interesting for control and on-line
optimization applications.
1.2. Multilayer Perceptron Neural Network (MLPNN)
A MLPNN consists of three types of layers: an input layer, an output layer and one or
more hidden layers. Each layer may have a different number of neurons, and even a
different transfer function. An appropriate architecture would help to achieve higher
model accuracy. It has been shown that all continuous functions can be approximated to
any desired accuracy, with a network of one hidden layer of sigmoidal
−X
f (X) = 1/(1+ exp ) hidden units (nodes) and a layer of linear output nodes. Such
structure was used in this work. Eq. (1) describes mathematically a MLPNN.
M N
1 = 1[ 1j j( jlX l + j0)
+ 10]
j= 1 l=
1
∑ ∑ (1)
y F W f w w W
E. Ccopa Rivera et al.
In Eq. (1), wj,l and W1,j specifies the adjustable parameters; i.e., the weights and biases.
X is the input variable of the network.

Hybrid Modeling for Continuous Production of Bioethanol 615
1.3. Functional Link Network (FLN)
The FLN was developed by Pao (1989) to increase the computing power of neural
networks. The input-output relationship of the FLN is represented by Eq. (2).
M
y (X ) = ∑ w h (X ) , 1 ≤ i ≤ m (2)
i e ij j e
j=
1
In the general structure of an FLN the hidden layer performs a functional expansion on
the inputs which maps the input space, of dimension n, onto a new space of increased
dimension, Mz(Mz>nz). In this case, the expansion results, hj(Xe), are a series of
monomials of Xe. The output layer consists on m nodes, each one, in fact, a linear
combiner, and each wij is an FLN weight.
In these networks, a non-linear functional transform or expansion of the network inputs
is initially performed and the resulting terms are linearly combined. More details about
the FLNs can be found in Costa et al. (1999).
2. Building up the models
This section shows the models development through the hybrid approach for a
continuous extractive alcoholic fermentation process and a continuous alcoholic
fermentation process with multiples stages.
2.1. Case study 1: Extractive alcoholic fermentation process
High concentrations of ethanol inhibit the fermentation process, particularly when a
fermentative medium with high concentration of substrate is used, as is the case in the
majority of the industrial processes. Taking these into account, a process of
fermentation combined with a vacuum flash vessel was considered, which selectively
extracts ethanol from the medium as soon as it is produced. The process is composed of
four interlinked units: fermentor (ethanol production unit), centrifuge (cell separation
unit), cell treatment unit and vacuum flash vessel (ethanol–water separation unit). It was
shown that this scheme presents many positive features and better performance than a
conventional industrial process. The extractive alcoholic fermentation process is shown
in Figure 1. A detailed description of the process and mathematical model can be found
in Costa et al. (2001).
F 0, S 0, T 0
F R , S R , X R , P R , T R
F, S F , X F , P F , T F
F W , T W
F LR , S LR , X LR , P LR , T LR
Fc1 T
R
E
A
T.
air
FERM.
CO2 air
F, S, X, P, T
F P
Fc, S, Xc, P, T
F E , S, X E , P, T
CENTRIFUGUE
Figure 1. Extractive alcoholic fermentation (Costa et al., 2001).
F
L
A
S
H
F V, P V, T V
F L
F LS

616
2.2. Case study 2: Alcoholic Fermentation Process with Multiples Stages
F0
FW
Cooler
FR
1
2
3
4
Cooler
Cooler
5
Cooler
acid
water, Fa
Cooler
Heavy Phase
FS
Purge
Stream
air
Surge Tank
Centrifuge
FV
Light Phase
Figure 2. Alcoholic fermentation process with multiples stages (Andrietta and Maugeri, 1994).
The fermentative process is illustrated in Figure 2. The system is a typical large-scale
industrial process made up of five continuous-stirred tank reactors (CSTR) attached in
series and operating with cells recycle. Each reactor has an external heat exchanger
whose objective is to keep the temperature constant at an ideal level for the fermentation
process. This process operating conditions are real conditions of typical industrial
distillers in Brazil. More details about this process are given by Andrietta and Maugeri
(1994).
2.3. Training and Validation
The training was performed using a full factorial design 2 2 +star configuration, involving
the variables S0 (kg/m 3 ) and T0 ( o C). Both variables had their values changed every 10
hours (this period was sufficient to reach the steady state) within the ranges 90 kg/m 3 to
270 kg/m 3 (S0) and 28°C to 40°C (T0) for Case 1 and 150 kg/m 3 to 210 kg/m 3 (S0) and
28°C to 35°C (T0) for Case 2. Thus, representative data containing the input and output
signals of the process was generated. For validation, the disturbances were also a
sequence of steps with periods of 10 hours within the operational interval 120 kg/m 3 to
230 kg/m 3 (S0) and 30°C to 38°C (T0) for Case 1 and 162 kg/m 3 to 198 kg/m 3 (S0) and
28°C to 32°C (T0) for Case 2.
For the current cases, the specific rates of microbial growth were modeled with an ANN
with four inputs (concentrations of biomass, substrate and ethanol, and temperature).
2.4. Quality of Prediction
The quality of prediction of the proposed hybrid models was given by the residual
standard deviation (RSD), Eq. (3), which provides an indication of the accuracy of the
predictions.
1
RSD= ( ( - ) )
n
2 0.5
yi yPi
n i=1
∑ (3)
In this equation yi is the ‘real’ value (calculated by the deterministic model), yPi is the
value predicted by the hybrid model and n is the number of points. The magnitude of
the RSD changes depends upon the magnitude of the variable to be predicted. It is
worthwhile mentioning that it is easier to analyze the RSD as a percentage of the
average of the real values y i (Atala et al., 2001), Eq. (4).
RSD
RSD(%) = × 100
(4)
y
i
E. Ccopa Rivera et al.

Hybrid Modeling for Continuous Production of Bioethanol
3. Results and Discussions
The developed hybrid models described well the simulation data. Figures 3(a) and
Figure 3(b) for Case 1 and Case 2, respectively, show the results which indicate the
good performance of the hybrid model for ethanol concentration.
Table 1 for Case 1 and Table 2 for Case 2, shows the prediction results of the hybrid
model for concentrations of biomass, substrate and ethanol, and temperature.
The range of deviations were from 0.019% to 0.649% and 0.002% to 1.973% for Case 1
and Case 2, respectively. Deviations below 10% are considered acceptable regarding
bioprocess engineering (Atala et al., 2001). These results show that this model is
reliable enough to be further used for optimization and control studies.
P [Kg/m 3 ]
40
37
34
31
28
25
0 20 40 60 80
(a) (b)
Time [h]
P 5 [Kg/m 3 ]
72
69
66
63
60
57
0 20 40 60 80
Time [h]
Figure 3. Process output (solid line) and synthetic data of the hybrid model (dashed line): (a) Case
study 1; validation data for ethanol concentration, P, into the reactor. (b) Case study 2; validation
data for ethanol concentration, P 5, of last (fifth) reactor.
Reactor
(i)
Table 1. Residual Standard Deviation for Case study 1
Output variable RSD (%)
Biomass (X) 0.019
Substrate (S) 0.649
Ethanol (P) 0.359
Temperature (T) 0.042
Table 2. Residual Standard Deviation for Case study 2
RSD(%)
Biomass (X i) Substrate (S i) Ethanol (P i) Temperature (T i)
1 0.0037 0.151 0.0323 0.0087
2 0.0021 0.309 0.0147 0.0049
3 0.0020 0.754 0.0126 0.0100
4 0.0024 1.429 0.0149 0.0125
5 0.0023 1.973 0.0143 0.0126
4. Conclusions
Bioreactors operating in continuous steady state are quite difficult to model, since
their operation involves microbial growth under constantly changing conditions. Hence,
there is a need for the development of simple though realistic mathematical descriptions
617

618
of alcoholic fermentation process, which can be used to develop optimal operating
strategies. This work demonstrated that hybrid model coupling deterministic
representation with universal approximators (MLPNN) or linear combination of
expanded input variables (FLN) was able to describe the fermentative processes with
good performance. The residual standard deviations (RSD%) showed low values.
Therefore, the developed models may be used to predict system performance and to
design process controllers. It is worthwhile mentioning that the proposed approach for
hybrid modeling is able to deal with disturbances in transient phases of the process and
different yeasts since the identification procedures be carried out adequately with a
suitable set of data representative of the process.
References
Andrietta, S.R. and Maugeri, F., Optimum design of a continuous fermentation unit of an
industrial plant for alcohol production. Advances in Bioprocess Engineering, 1994, 1 47-52.
Atala, D.I.P., Costa, A.C., Maciel, R. and Maugeri, F., Kinetics of ethanol fermentation with high
biomass concentration considering the effect of temperature. Applied Biochemistry and
Biotechnology, 2001, 91-93(1-9) 353-366.
Costa, A.C., Henriques, A.W.S, Alves, T,L.M, Lima, E.L. and Maciel Filho, R., A hybrid neural
model for the optimization of fed-batch fermentations. Brazilian Journal of Chemical
Engineering., 1999, 37 125-137.
Costa, A.C., Atala, D.I.P., Maugeri, F. and Maciel Filho, R., Factorial design and simulation for
the optimization and determination of control structures for an extractive alcoholic
fermentation. Process Biochemistry, 2001, 37 125-137.
Harada, L.H.P., Costa, A.C. and Maciel Filho, R., Hybrid neural modeling of bioprocess using
functional link networks. Applied Biochemistry and Biotechnology, 2002, 98-100(1-9) 1009-
1024.
Pao,Y. H. Adaptive Pattern Recognition and Neural Networks. Addison-Wesley Publishing
Company, California, 1989.
Psichogios, D.C. and Ungar, L.H., A hybrid neural network-first principles approach to process
modeling. American Institute of Chemical Engineers Journal, 1992, 38 1499–1511.
Zorzetto, L.F.M., Maciel Filho, R. and Wolf-Maciel, M.R., Process modelling development
through artificial neural networks and hybrid models. Computers and Chemical Engineering,
2000, 24 1355–1360.
Acknowledgements
The authors acknowledge FAPESP and CNPq for financial support.
E. Ccopa Rivera et al.