J. Inst. Brew., March-April, 1994, Vol. 100, pp.

99-104

TEMPERATURE CONTROL IN FERMENTERS: APPLICATION OF NEURAL NETS AND

FEEDBACK CONTROL IN BREWERIES

By G. Gvazdaitjs, S. Beil, U. Kreibaum, R. Simutis, I. Havlik, M. Dors, F. Schneider,

and A. LGbbert

(Institut far Technische Chentie, Universiliit Hannover, Callinslrasse 3, D-3000 Hannover, Germany)

Received 19 April 1993

The main objective of on-line quality control in fermentation is to perform the production processes
as reproducible as possible. Since temperature is the main control parameter in the fermentation
process of beer breweries, it is of primary interest to keep it close to the predefined set point. Here,
we report on a model-supported temperature controller for large production-scale beer fermenters.
The dynamic response of the temperature in the tank on temperature changes in the cooling elements
has been modeled by means of a difference equation. The heat production within the tank Is taken
into account by means of a model for the substrate degradation. Any optimization requires a model
to predict the consequences of actions. Instead of using a conventional mathematical model of the
fermentation kinetics, an artificial neural network approach has been used. The set point profiles for
the temperature control have been dynamically optimized in order to minimize the production cost
while meeting the constraints posed by the product quality requirements.

Key Words: Fermentation, temperature control processes
Introduction
Temperature control is a practical problem in production-
scale brewery fermenters, when the vessels are very big and
not able to immediately follow changes in the temperature
set point profiles. Since in practice the control strategies are
very simple, considerable temperature fluctuations may be
observed in big fermenters.
Temperature is the variable by which the fermentation
process in beer breweries is controlled. It is, thus, essential
for a reproducible fermentation process (the only practical
means to ensure constant quality during fermentation) that
the temperature exactly follows the predefined profile.
Hence, the temperature control is of considerable importance
for any quality assurance of the beer fermentation process. In
particular, for a reliable prediction of the process behaviour,
which becomes necessary if the length of the fermentation
process and thus the tank logistics is to be controlled, one
must presume that the temperature can be kept close to the
predefined temperature profile.
With larger production vessels, the time constants of the
temperature response behaviour increase, and it becomes
more difficult to control the temperature in the tanks. Then,
advanced control strategies become indispensable in optimiz
ing the quality/cost relation of the fermentation process.
Energy conservation is on everybody's agenda these days.
In new plants for beer fermentation and maturing measures
for reducing the energy for cooling is of high importance. For
daytime cooling operations, a big Japanese brewery recently
installed ice banks frozen at night using cheap 'off peak'
electricity to minimize the refrigeration costs.
In this paper, we describe an advanced temperature control
strategy for brewery fermenters by the example of a 300 mJ
cylindroconical fermenter installed in a German brewery.
Temperature Control
cess in breweries, this primarily means to control the fermen
tation temperature.
Advanced quality assurance requires continuous state pre
diction as the essential basis of control. In the beer fermen
tation process, prediction of the substrate degradation rate
and prediction of those components in the fermentation
medium that influence the taste of the beer are of primary
importance. It would be of considerable value to monitor
and predict the development and degradation of the key
component of the different compounds influencing the taste,
diacetyl. This is a requisite for every measure of keeping it
below some threshold concentration.
In an earlier publication, we described the supervision and
prediction of the extract concentration in a beer fermentation
by means of fuzzy-supported extended Kalman filters4 and
by fuzzy-aided artificial neural networks5.
In the course of these investigations, it has been found
that the temperature fluctuations limit improvements in the
state prediction of fermentation processes. Hence, advanced
temperature control becomes imperative.
Any model-supported control of the fermentation tempera
ture must be based on a reliable description of the cooling
behaviour of the big fermentation tanks as well as on a
reliable description of the temperature generation process,
i.e., the fermentation process itself.
As a representation of the fermentation process we use
an extended form of the neural network software described
by Simutis3. The dynamic response behaviour of the 300 m3
cylindroconical fermentation tank used in the corresponding
fermentation runs was modeled by means of a difference
equation. Both models have been incorporated into a model-
supported adaptive temperature controller. The models and
the controller construction are described in the following
sections.
Based on the process models and the controller, a minimiz
ation of the cooling expenses under the constraints of prede
fined tolerances of some quality parameters has been perfor
med by construction of an optimal feedback controller.

Objectives Neural Net Representing the Extract Degradation

Quality assurance is imperative for all industrial compa
nies. The main aim of quality control is to keep those process
parameters which characterize the product quality strictly
within predefined tolerance bands. In most processes where
living organisms are involved, this can only be obtained by
running the process under reproducible conditions, i.e., close
to the desired parameter profiles. For the fermentation pro
In their previous work, the authors4 compared different
ways of representing the microbial conversion of malt into
ethanol and carbon dioxide in the brewery fermentation
process. One result was that process simulations based on
neural networks are at least as reliable as simulations using
mathematical process models, but the networks turned out
to be much easier to establish and to maintain. Consequently,
100 TEMPERATURE CONTROL IN FERMENTERS
we used a neural network approach in the work described
in this article.
In the model presented by Simutis et al.5 the only variable
predicted was the substrate degradation rate or the extract
concentration, respectively. In order to supervise the key
quality indicator as well, the diacetyl concentration was incor
porated into the model. Since the available expert knowledge
about the diacetyl development and its degradation is rather
limited, we did not use a fuzzy-aided neural network as in
the reference mentioned. Instead, we used a more compli
cated net and an advanced training technique to fit the
network to the requirements. The training procedures used
are described in Simutis et al.6
As can be seen in Figure 1, the network contains 9 input
and 2 output nodes. These numbers, characterizing part of
the network topology are fixed to the available data and
required output quantities. The remaining question is the
.structure of the hidden layers. Only a single hidden layer
was chosen. 10 nodes in the hidden layer proved to be
sufficient for modelling the fermentation process and to keep
the time needed for the training in acceptable limits. The
number of nodes in the hidden layer were determined by
means of cross validation techniques3. The essential idea
behind this technique is to divide the available data sets into
two parts. About 70% of the data sets are used for training,
the remaining 30% were kept aside to validate the trained
network with independent data.
The advanced training procedure used in the experiments
is based on an evolutionary optimization algorithm, i.e., the
ideas behind the optimal development of organisms in nature
(mutation, cross-over and selection in a set of individuals)
have been applied to the optimization of the artificial neural
network representing the dynamics of the substrate degra
dation and diacetyl formation/degradation in a beer fer-
menter. A typical result of the simulation is depicted in
Figure 2. As can be seen, the simulations very closely follow
the measured experimental data.
Model for the Fermentation Tank Dynamics
Fermentation tanks with 300 m3 total volume have been
taken as examples. Their diameters are 4.1 m, and they are
equipped with three wall cooling segments each. The cooling
is performed by vaporization of ammonia.
The dynamical behaviour of big cylindroconical fermen
Time
Fio. 1. Structure of the neural network used lo represent the substrate degradation and diacetyl formation and degradation in the
fermentation tank.
[J. Inst. Brew.
tation tanks is complicated by several mechanisms. The stir
ring action of the CO2 bubbles rising in the liquid phase is
by no means continuous, considerable eruptions characterize
the carbon dioxide gas production. Supersaturation effects
and chaotic flow properties of the dispersion may be respon
sible for this behaviour. As the cooling sections at the fer-
menter wall are switched on and off in a binary way, the
boundary conditions concerning the temperature at the fer-
menter wall are abruptly changed.
Since most of the parameters determining the heat transfer
coefficients are unknown, the cooling behaviour of the tank
was determined experimentally by analysing the temperature
response to changes in the cooling control signals. In order
to keep the dynamical representation of the results simple,
a linear difference equation was used. In this equation, the
temperature sampling values Tk, where k is the time index,
are related to the registered control signals Sik of the i-th
cooling element.
Tk
a, Tk_, + a2 Tk_2 + a3 Tk_3 aj Tk_4 + a5 Tk_5
b, b3 Si,k_j b4 Si,k-4 bs Si.k_5
b| S2,k_i b2 S2.k-2 b3 S2-k_3 b4 S2.k_4 + b5 S2-k_5
bi S3,k_i b2 S3,k_2 b3 S3,k_3 b4 S3,k_4 + b5 S3-k_5
Kk
with the parameters ai, bj, Cj, determined by the identification
procedure. The variables Kk are characterizing the heat pro
duction and the Ck are characteristic for the CO2-off-gas
mass flow. As a characteristic CO2 mass flow through the
off-gas line we took a short-term integrated average (over
2-3 hours in our application) of the CO2 massflow. The
characteristic heat production parameter Kk can be deter
mined in two alternative ways, either from the long term
CO2 development, which is directly proportional to the heat
production, or by directly incorporating the output of the
neural net describing the kinetics of the substrate degra
dation, as described above. Once again, the fact is used that
heat production is proportional to the substrate conversion
rate.
The identification of the other parameter was performed
Vol. 100, 1994| TEMPERATURE CONTROL IN FERMENTERS 101

120 0.6

110
-- 0.5
"*N£iacetyl
100
Extract^
I
\
\
r i
-
0.4 §
80

/\
80 0.3

70
-- 0.2

I
/
60

■- 0.1

/
SO

40
SO 100 150 200

Fermentation Time [h]

Fig. 2. Typical examples of the predicted variables substrate and diacetyl concentration. The symbols are the experimental data measured
off-line. The lines represent the output of the neural network simulation.

with a standard optimization procedure taken from the well-
known NAG scientific subroutine library.
In order to get an impression of the performance of this
model, the temperature was predicted for periods of two
hours in advance using the available measuring data sampled
up to the actual time point. These simulated temperatures
are depicted in Figure 3 together with the temperature
measurement values obtained two hours later.
As can be seen from Figure 3, the temperature of the tank
can be predicted quite accurately two hours in advance.
Hence, this model can be used to construct a model-
supported temperature controller.
Since it has been observed that the dynamical behaviour of
the big fermentation tank is different at different temperature
intervals, the parameters of the model have been separately
identified for intervals around the normal fermentation tem
perature and the temperature which is kept in the end phase
of the fermentation, the so called diacetyl step. The two
different temperature intervals in our application are
12°C > T > 7"C and 7°C> T > 4°C.

50 100 150 200

Fermentation Time [h]

Fig. 3. Comparison of the simulated temperatures together with the temperatures measured two hours later. Also, the control profile of
the cooling system is shown in Ihe graph. Its values 1, 2, and 3 denote that only the upper cooling section, the upper two only, and
all three cooling sections are switched on.
102 TEMPERATURE CONTRO1. IN FERMENTERS [J. Inst. Brew

DATA MEMORY

J SWITCH
Lpcto ERROR /
FUZZY
/I
Y CONTROLLER / PLANT

/
, /
WATCH

DOG
J CCS CO 2

DATA
MEMORY

ERROR ^

i S
CONTROLLER
y
To

PROOESS
MODEL

7
Fig. 4. Schematic view of the temperature controller consisting of the adaptive controller, the fuzzy controller and the watch-dog which
works as supervisor.

Controller Design
The feedback controller used is based on a prediction of
the temperature by means of the above-mentioned models.
The required control vector for the three cooling elements
is computed every 8 sampling time steps (12 min each) only,
in order to reduce the controller actions. In Figure 3 it was
shown that the model can reliably predict the temperature
over such long intervals. However, time between is used to
adapt the model parameters to the process. Hence, after the
8 time steps, the next control profile segment is computed
with an updated model.
The control actions for the 8 time steps are chosen from
all possible combinations of activating the three cooling
sections in such a way that the mean squared deviation
from the set point profile in that interval becomes minimal.
In the case that the actual temperature is far below the
set point temperature, nothing is done, because it is not
possible to heat the fermenter. In this case, one must wait
until the internal heat production enhances the temperature
to a value near the set point.
The most important parameter in the model-supported
controller is the adaptation gain by which the deviations

ON-LINE (a)

aolpolnl
FUZZY-AIDED CCS STATE X(t)

CONTROLLER

APPRIORI
KNOWLEDGE
ERROR

FUZZY-AIDED CCS PLANT

CONTROLLER MODEL

OPTIMIZATION

1 opllm
J
Flo. 5. Schematic of the optimal feedback controller used to minimise the fermentation cooling cost in a running production.
Vol. 100, 1994] TEMPERATURE CONTROL IN FERMENTERS
120
Optimized
110
100
\
90
\ -
O 80
W 70
\ \
60
50
40
0 SO 100
Fermentation Time [h]
Fig. 6. Typical example of the controlled temperature with and without optimal feedback control.
AT of the actual temperature value from the temperature
predicted by the model influences the model parameters used
in the controller. If the gain is high, i.e., the influence of
the temperature deviations on the model adaption is high,
the controller might become unstable. If it is too low, the
controller might not follow fast changes in the system behav
iour. Since the response of the reactor on the actions in the
cooling system is rapidly changing during the fermentation,
it was decided to use a high gain.
Any risk of obtaining an unstable controller behaviour was
kept extremely low by the introduction of a robust fuzzy
expert controller which was constructed from the available
knowledge on the cooling process. This fuzzy controller is
able to compensate for larger deviations of the actual tem
perature, however, its performance is not as good as that of
the adaptive model-supported controller.
In order to supervise both controllers, a watch-dog module
was installed in the software which detects whether the adapt
ive controller works correctly or not. If the adaptive control
cannot cope with the control, the watch-dog switches over
to the fuzzy controller. Figure 4 depicts a schematic view of
the controller structure.
Profile Optimization
Objectives
The temperature fluctuations, observed over a longer
period showed that random variations of ±1 °C around the
set point do not influence the beer quality significantly.
However, the predictability of the extract degradation and
its reproducibility depend on the accuracy of the temperature
control.
At first glance, an improved control would thus primarily
support the process predictability. This might not be a strong
enough argument to justify investments for an improved
temperature control. The value of an improved controller
becomes more obvious if one considers its integration into
an optimized feedback control strategy, where the quality
requirements as well as the cooling costs are taken into
account. In other words, the possibility to more precisely
guide the temperature within a given tolerance band can be
103
12
11
10
L 9
E
-
8
I2
1
7
6
53
a.
H
5
4
150 200
used to minimize the cooling costs under the condition that
the quality parameters stay within guaranteed intervals.
Since in our application the cost of electrical energy is
lower overnight, the cooling actions are controlled, so that
cooling is predominantly performed during night hours. The
process state prediction developed is thus used to fit the
process control profiles to match this requirement.
Implementation
The cost functional J used in the optimization is defined
as follows:
J = k, (Sf - Sg (Deviation from extract set
point)
k2D2 + (Deviation from diacetyl set
point)
2 uk P At (Cooling cost)
— The first component describes the influence of a deviation
of the actual extract concentration from the model value.
The coefficient kt is typically chosen to be 0.05.
— The second term describes the influence of the diacetyl
concentration D. It is assumed that concentration D influ
ences the loss quadratically. Parameter k2 is zero as long as
the concentration is below some threshold value (usually
0.2 mg/1). Above that concentration it might be chosen to
be 100.
— The third term sums up the cooling cost up to the actual
point of time. At is the time increment. uk, with k = 1, 2, 3,
respectively, is a binary control signal indicating whether the
individual cooling elements are switched on or off. P = P(t),
a function of time t, is the price of electric power for the
cooling aggregate.
The minimization of the cost functional was obtained with
classical optimization techniques. A NAG library routine
(E04DBF) was used in the optimization.
Figure 5 shows the structure of the feedback controller
schematically. The controller contains an on-line part and an
off-line component. In the off-line part, the optimal control
profile is calculated every 24 hours in advance. Then, the
on-line part controls the temperature along this profile. Dur
ing this closed loop control, the deviations of the temperature
104 TEMPERATURE CONTROL IN FERMENTERS
from the predefined profile are used to correct the para
meters of the model within the controller. If, every 24 h,
new extract values are available, the kinetic representation
by the neural nets is also improved. So, for the next days
the profile can be calculated using updated models.
Figure 6 presents a typical example of the controlled tem
perature with and without optimal feedback control. The
optimized controller fully exploits the temperature interval
allowed during the extract degradation. As far as possible,
cooling actions are performed overnight.
For a single fermenter, the cost reductions may come up
to more than 20% in this way. This is in the same order of
magnitude as the energy reduction claimed by Asahi and
Kinoshita, who used an unspecified AI computer system
in combination with a cold storage system to maintain an
optimized economical balance.
Conclusion
A model-based adaptive controller has been constructed
based on a neural network representation of the substrate
degradation and the diacetyl formation/degradation process
and a difference equation to model the dynamical tempera
ture behaviour of the fermenter. The feedback control profile
could be optimized in order to minimize the cooling expense
while fully maintaining the quality requirements.
The investigation shows that advanced software techniques
combining traditional and new approaches can be used to
improve production processes. Since the development time
[J. Inst. Brew.
for these powerful tools is considerably shorter than that for
the construction of conventional software approaches, the
new techniques become economically interesting for many
industrial companies.
The controller was developed using the practical example
of a fermentation process of a German brewer. The software
runs under the process management system RISP on a simple
small VAX workstation connected to several fermentation
tanks via a microprocessor system.
Acknowledgements. The financial support of the AIF
(Arbeitsgemeinschaft industrieller Forschungsvereinigungen,
Koln) is gratefully acknowledged. This AIF grant Nr. 08
was financed by the German Federal Ministry of Economic
Affairs. The Gilde Brauerei AG, Hannover, generously sup
ported the developments described.
References
1. Asahi, K., Kinoshita, M. Technical Quarterly of the Master Brew
er's Association of the Americas, 1992, 29, 48-52.
2. Beil, S., Dors, M., Havlik, I.. Krcibaum, U., Schneider. F..
Simutis, R., Lubbert, A. Brauwissenschaft, 1992, 45, 196-204.
3. Nahas. E. P.. Henson, M. A., Scborg. D. E. Computerised
Chemical Engineering, 1992, 16, 1039-1057.
4. Simutis, R., Havlik, I., Lubbert. A. Journal of Biotechnology,
1992.24,211-234.
5. Simutis, R., Havlik, I., Lubbert, A. Journal of Biotechnology,
1993, 27, 203-215.
6. Simutis, R., Havlik, I., Dors, M., Lubbert, A. Process Control
and Quality, 1993, 4, 211-220.