You are on page 1of 6

20

th
European Symposium on Computer Aided Process Engineering ESCAPE20
S. Pierucci and G. Buzzi Ferraris (Editors)
2010 Elsevier B.V. All rights reserved.
A hybrid neural approach to model batch
fermentation of dairy industry wastes
Alessandra Saraceno, Sascha Sansonetti, Stefano Curcio, Vincenza Calabr,
Gabriele Iorio
Department of Enginering Modeling-University of Calabria, via P. Bucci-cubo 42/A,
Arcavacata di Rende 87036, Italy, alessandra.saraceno@unical.it

Abstract
In this work, the fermentation of Ricotta cheese whey for the production of ethanol
was simulated by means of a Hybrid Neural Model (HNM), obtained by coupling neural
network approach to mass balance equations describing the time evolution of lactose
(substrate), ethanol (product) and biomass concentrations. The realized HNM was
compared with a pure neural network model (NM) and the advantages gained from the
hybrid approach were emphasized. The experimental data, necessary to develop the
model, were collected during batch fermentation runs. For all the proposed networks,
the inputs were chosen as the operating variables exhibiting the highest influence on the
reaction rate. The simulation results showed that the HNM was capable of an accurate
representation of system behavior by predicting biomass, lactose and ethanol
concentration profiles with an average error percentage lower than 10%. Moreover,
especially if compared with the NM, the HNM showed good forecasting capability even
with fermentation run never seen during the training phase.

Keywords: grey-box models, artificial neural networks, batch fermentation, modeling
1. Introduction
Modeling of biotechnological processes represents a key issue to achieve proper design
and control aimed at process optimization (Lubbert and Jorgensen, 2001). Biochemical
reactions actually involve many parallel-serial reaction steps and depend on several
transport phenomena that may limit the observed reaction rates. Usually, the
fundamental approach, largely considered as the most rigorous, cannot be applied in
reaction kinetic modeling due to inherent non-linearity, lack of information,
experimental inaccuracy, deviations from ideal conditions (Feyo de Azevedo et al.,
1997). On the contrary, black-box models, suggested by the empirical approach, suffer
from a restricted validity domain depending on the range of data collected during the
experiments. A reasonable trade-off between theoretical and empirical approach is
represented by hybrid modeling, leading to a grey-box model capable of good
performance in terms of data interpolation and extrapolation. The main advantage of
hybrid modeling regards the possibility of describing some well-assessed phenomena by
means of a fundamental theoretical approach, leaving the analysis of other aspects, very
difficult to interpret and describe in a traditional way, to rather simple cause-effect
models. Among these, Artificial Neural Networks (ANNs) were successfully used in
bioreactor modeling (James at al., 2002; Laursen et al., 2007; Simutis et al., 1995)
ANNs consist of interconnected computational elements called neurons or nodes: each
A.Saraceno et al.


neuron receives input signals from the related units, elaborates these stimuli by an
activation or transfer function and generates an output signal that can be transferred to
other neurons. ANN can be used both in a pure empirical model structure and in a
hybrid model formulation. When ANNs are utilized in a hybrid model, a very important
step is identifying the respective domain of theoretical and empirical parts in such a way
as to decide which aspects of the process are to be described by a black-box model.
Ricotta cheese whey (RCW),scotta, is a dairy industry waste characterized by a
lactose content of about 45-50 g/l and a very low amount of proteins ranging from 0.22
0.24 % in weight. BOD reduction of scotta can be achieved, for instance, by
alcoholic fermentation of lactose carried out by Kluyveromyces Marxianus. Even if
several authors dealt with biotechnological utilization of cheese whey (Gonzles Siso,
1996; Ozmihci and Kargi, 2006, 2007), as far as our knowledge is concerned, a single
work exists about Ricotta Cheese whey(Sansonetti et al., 2009).
The aim of this paper was to propose a model that could describe batch fermentation of
lactose in ethanol investigating two different strategies. At first, the concentration
profiles of the reacting species were described by a pure neural model; afterwards, with
the aim of introducing theoretical information (mass balance) into the model, a multiple
hybrid neural model was realized.
2. Development of the models
The proposed models, i.e. NM and HNM, are aimed at the description of the batch
fermentation of RCW predicting biomass, lactose (substrate) and ethanol (product)
concentration profiles. A preliminary phase of the model development was performed.
It consisted on the identification of the operating variables that, among all the
parameters that could affect reaction progress, exhibited the highest influence on
process performance. These variables, i.e.: temperature (T), pH (pH), reactor stirring
rate (rpm), initial lactose concentration (C
0
lat
) and reaction time (t), represented the
inputs for both the pure neural and the hybrid neural model.
2.1 Neural model structure
A multiple neural model was proposed to predict concentration profiles. It consisted of
three neural networks, i.e. NM1, NM2 and NM3, respectively capable of predicting
biomass, lactose and ethanol concentration profiles as outputs. The structure of the three
neural networks was very similar: they were characterized by the same number of input
nodes, corresponding to the pre-identified input variables and by a single output node
that had to produce, as output, the corresponding concentration profile. Between input
and output nodes two hidden layers were put and the number of nodes in each layer was
determined according to an iterative procedure.
2.2 Hybrid model structure
The realized hybrid neural model consisted of three models characterized by a common
structure. A mass balance equation was written for each of the compounds and an ANN
was set up to evaluate the kinetic parameters actually necessary to solve the balance
equations. Some logic conditions, expressing the fulfilment of physical constraints,
were also introduced into the model.
The three mass balance equations, written with reference to biomass, lactose and
ethanol assumed the following form:
A hybrid neural approach to model batch fermentation of dairy industry

=
=
=
X
dt
dP
qX
dt
dS
X
dt
dX
P

(1)
where X, S, P were, respectively, biomass, lactose and ethanol concentration profiles,
was biomass specific growth rate, q was the lactose consumption rate function,
p
was
the product growth rate and t was the reaction time. Eqs. (1) were approximated by
Eulers discretization and the discretized form was used recursively to determine the
biomass, lactose and ethanol concentration values at the time instant t+t, knowing all
the information concerning time instant t. The values of kinetic parameters at time
instant t, strictly necessary to solve the discretized form of the equations, were provided
by three neural networks named HNM1, HNM2 and HNM3, respectively predicting ,
q and
p.
The three neural networks were set up according to an iterative procedure. The
kinetic parameters, evaluated using HNM1, HNM2 and HNM3, were subsequently
processed by some logic conditions that verified the agreement between the net outputs
and some well-assessed process information. The logic condition assumed the following
form:

= <
= <
= <
0 0
0 0
0 0
p p
if
q q if
if


(2)
Eqs. (2) faced the fact that, during the fermentation runs, biomass and ethanol
concentrations always grew, whereas lactose concentration always decreased.
3. Materials and methods
3.1. Experimental design
The experimental data necessary to develop the models were collected from a set of
anaerobic fermentations performed on RCW. According to the factorial design method
(Box et al., 1978), the following variables and operating condition were chosen: 1)
Temperature in the range 32- 40 C; 2) pH in the range 4-6; 3) Stirring rate in the range
100-300 rpm; 4) The lactose concentration was varied between 50-90 g/l since this is
the expected range in which the process could be operated. The lowest value of the
range was set to 50 g/l due to the typical lactose concentration of RCW, i.e. 45-50 g/l.
The highest value of the range, i.e. 90 g/l, derived from the consideration that it could be
useful to concentrate RCW before the fermentation process so to achieve a more
concentrated ethanol solution flowing out the reactor. Nevertheless this value cannot be
increased as much as one likes since a limit imposed by the osmotic pressure at the cell
wall does actually exist. The value of 90 g/l value was chosen according to several
preliminary experiments (not reported in the paper), which showed the technical and
economical feasibility of RCW pre-concentration (up to 90 g/l) by nanofiltration and,
therefore, by a well-assessed procedure.
A total of 16 batch runs, having a duration of 18 hours with a sampling time of one
hour, were therefore performed and a total of 912 experimental points, expressing the
A.Saraceno et al.


time evolutions of biomass, lactose and ethanol concentration, were collected during
fermentation.
3.2. Neural network development
To realize both the NMs and the HNMs, six different neural networks were developed
following a common iterative procedure implemented according to Matlab Neural
Network Toolbox Ver. 4.0.1. For the neural model only the concentration values
relative to the batch fermentation runs were sufficient to train the networks. For the
hybrid model development it was necessary to determine the experimental values of ,
q,
p
by interpolating the collected experimental data by cubic splines. Subsequently,
the experimental points relative to 15 of the 16 available runs were used to train/test the
developed networks: the data points were randomly split into two groups, reserving 2/3
of data (570 points) to the training phase and remaining 1/3 (285 points) to test the
neural networks. Finally, the six realized neural models were validated using two
complete experiments never exploited during both the previous training/test phases. The
neuron transfer function was always the hyperbolic tangent, except for the output layer
where a linear transfer function was used. To develop the ANNs, an iterative procedure
was implemented. The performance index, chosen to evaluate simulation results
reliability, was the percentage error between each predicted (C
p,k
) and measured (C
m,k
)
concentration value: in a whole batch run the average value of the percentage error had
to be lower than 10%. The iterative procedure consisted in placing a layer with a single
neuron between the input and output layers, thus verifying the fulfillment of
convergence criterion; if at the end of the training procedure the performance objective
was not achieved, a neuron was added into the hidden layer, repeating the procedure
until the performance objective was achieved. If the neurons of the 1
st
hidden layer
reached the number of 10, an additional hidden layer with a single neuron was put into
the network and, the procedure described for the first hidden layer iterated.
4. Results and discussion
On the basis of the previous discussion, a multiple neural model and a multiple hybrid
neural model were realized. Table 1 summarizes the architecture of all the developed
networks.
Table 1. ANNs architecture
Neural Network
Number of neurons
1
st
hidden layer
Number of neurons
2
nd
hidden layer
Number of neurons
output layer
NM1 8 5 1
NM2 10 10 1
NM3 10 4 1
HNM1 6 5 1
HNM2 10 3 1
HNM3 10 2 1

A comparison between models predictions and experimental data is reported in Fig.1.
Fig. 1 reports the simulation results relative to a batch run used during training/test
phase of the models. A remarkable agreement between model predictions and the
corresponding experimental data was found, both with the pure neural and the hybrid-
neural approach so that an average percentage error much lower than 10% was found.
A hybrid neural approach to model batch fermentation of dairy industry



Figure 1. Predictions during training/test. Run conditions: T=32C, pH=6, rpm=100, C
0
lat
=90 g/l
0.00
30.00
60.00
90.00
0 2 4 6 8 10 12 14 16 18
time [h]
c
o
n
c
e
n
t
r
a
t
i
o
n

[
g
/
l
]
biomass (experimental) lactose (experimental) ethanol (experimental)
biomass (NM1) lactose (NM2) ethanol (NM3)
biomass (HNM1) lactose (HNM2) ethanol (HNM3)

Moreover, it should be observed that even if models performance is very similar, NMs
were more accurate in predicting the system dynamics with the experimental point
reserved to training/test phase. This result emphasizes how neural networks, being data-
based models, are able to perform very well on training data. Indeed, the training phase
of a neural network consists in structuring it so as to predict the training points
extremely well. Fig. 2. shows two different validation tests aimed at comparing hybrid
neural model and pure neural model predictions under a combination of operating
conditions never exploited during the training and the test phases, even if included in
the same range chosen to perform the experimental design. From Fig. 2 it is possible to
appreciate how NMs and HNMs validation results were quite different. The hybrid
model was capable of giving good simulation results even during validation: the model
performance, measured in terms of %, is comparable to that obtained with the
experimental data used to perform the training and the test phases.

Figure 2. Models validation. Run conditions: a) T=32C, pH=4, rpm=100, C
0
lat
=50 g/l; b)
T=37C, pH=5, rpm=300, C
0
lat
=50 g/l
0
30
60
0 4 8 12 16
a) time [h]
c
o
n
c
e
n
t
r
a
t
i
o
n

[
g
/
l
]
0
30
60
0 2 4 6 8
b) time [h]
c
o
n
c
e
n
t
r
a
t
i
o
n

[
g
/
l
]




This result shows how hybrid models do not just learn in order to recognize the training
points; instead, the models are able to predict the system behaviour even when they are
operated in conditions unexploited during training phase. On the contrary, neural
A.Saraceno et al.


models, during the validation, give worse performance that those obtained during the
training phase. Indeed, even if the neural models are able to predict smoothly the system
behaviour, the error percentage is significantly higher than 10% thus confirming how
pure black-box models are very reliable into their definition domain whereas they
become less accurate when used in interpolation and extrapolation. Therefore, the
results obtained in this paper, indicate how the introduction of theoretical information
into black-box models can lead to a more robust modeling approach.
5. Conclusions
The purpose of this work was to make a comparison between a pure neural and a hybrid
neural approach in modeling the batch fermentation of RCW. The hybrid modeling
approach showed better forecasting capability than the pure neural model, being capable
to predict the system behavior even in operating conditions never exploited during the
training phase. On the other hand, NMs did not show the same reliability during the
model validation thus stating how the introduction of simple theoretical information into
black-box models can lead to a strong improvement in model performance. As a matter
of fact, HNMs turned out to be a very efficient tool for the simulation of the RCW
fermentation process and, more generally, it could be used in modeling several complex
biotechnological processes. Overcoming the difficulties to describe very complex
reaction mechanism, HNMs are able to exploit all the available theoretical knowledge
all the same. This modeling effort can lead to more robust models than empirical
models and this feature make the hybrid neural approach a promising alternative in
process engineering because most of the advanced control and optimization strategies
are based on mathematical models.
References
G. Box, W. Hunter and S.Hunter , 1978, Statistics for Experimenters, An Introduction to Design,
Data Analysis and Model Building, John Wiley and Sons, New York.
S. Feyo de Azevedo, B. Dahm and F.R. Oliveira, 1997, Hybrid modelling of Biochemical
Processes: A comparison with the conventional approach., Comput. Chem. Eng., 21,751.
M. I. Gonzles Siso, 1996, The biotechnological utilization of cheese whey: a review, Bioresour.
Technol., 57, 1.
S. James, R. Legge and H.Budman, 2002, Comparative study of black-box and hybrid estimation
methods in fed-batch fermentation, J. Process Control, 12, 113.
S. O.Laursen, ,D. Webb and W. F. Ramirez, 2007, Dynamic HNM of an industrial fed-batch
fermentation process to produce foreign protein, Comput. Chem. Eng., 31, 163.
A. Lubbert and S. B. Jorgensen, 2001, Bioreactor performance: a more scientific approach for
practice, J. Biotechnol., 85, 187.
S. Ozmihci and F. Kargi, 2006,Utilization of cheese whey powder for ethanol fermentation:
effects of operating parameters, Enzyme and Microbial Technol., 38, 711.
S. Ozmihci and F. Kargi, 2007, Kinetics of batch ethanol fermentation of (CWP) solution as a
function of substrate and yeast concentrations, Bioresour. Technol., 98, 2978.
S. Sansonetti, S. Curcio, V. Calabr and G. Iorio, 2009, Bio-ethanol production by fermentation
of Ricotta Cheese Whey as an effective alternative non-vegetable source, Biomass and
Bioenerg., 12, 1692.
R. Simutis, M. Dors and A. Lubbert, 1995, Bioprocess optimization and control: Application of
hybrid modelling, J. Biotechnol., 42, 285.