Bioprocess Biosyst Eng (2006) 29:399407

DOI 10.1007/s00449-006-0089-5

ORIGINAL PAPER

Predictive controller evaluation including non-stationary

high frequency noise and outliers for batch solid substrate
fermentation bioreactors
J. R. Perez-Correa M. Fernandez-Fernandez

Received: 29 September 2006 / Accepted: 29 September 2006 / Published online: 3 November 2006

Springer-Verlag 2006

Abstract Optimum operation and automatic control

of large-scale solid substrate fermentation (SSF) bioreactors is difficult. Though advanced control algorithms can handle most challenges encountered
properly, for real-time SSF processes such controllers
are expensive and time consuming to design and tune.
With these considerations, advanced control algorithm
tests using realistic simulations appear more appropriate. We used a phenomenological process model of
an SSF pilot bioreactor, coupled with a realistic noise
model, to test linear model predictive controllers. We
focused on the effect noise has on the performance of
the control algorithms, and how to enhance performance using a combination of low-pass (Butterworth)
and outlier shaving (Hampel) filters. In simulations
undertaken directly with the phenomenological model
it was relatively straightforward to achieve good control performance. Nevertheless, control degraded
sharply when the output of the phenomenological
model was contaminated with noise using our realistic
noise model, even with proper signal filtering.
Keywords Non-linear dynamics
Internal model
control
Model predictive control
Solid substrate
cultivation

J. R. Perez-Correa (&)
Departamento de Ingeniera Qumica y Bioprocesos,
Pontificia Universidad Catolica de Chile, Casilla 306,
Santiago 22, Chile
e-mail: perez@ing.puc.cl
M. Fernandez-Fernandez
Departamento de Ciencias de la Ingeniera, Universidad de
Talca, Camino Los Niches Km. 1, Talca, Chile

Introduction
Solid substrate fermentation (SSF) has been used for
centuries to produce pigments and fermented foods
such as Koji. Nowadays, facilities abound for producing other high value microbial metabolites by SSF;
antibiotics, biopesticides, aromas, hormones, and enzymes and the like [1]. SSF processes are particularly
attractive industrial applications, as they offer higher
product yields and a different expression of microbial
metabolites than submerged fermentations (SmF). In
spite of such promise, several unresolved technological
problems, such as the lack of affordable and reliable
instrumentation and inherent process complexity, hinder SSF processes at industrial scales [2]. In particular,
at both industrial and pilot scales, metabolic heat removal from SSF bioreactors is extremely difficult
resulting in disappointingly low productivity, growth of
competing microorganisms (contaminants) or complete fermentation run failure [3]. Consequently, few
of the many SSF processes developed at laboratory
scale have been scaled up successfully to industrial
production. The situation would be very different if
effective control strategies for regulating the bed
temperature were available.
Temperature control in large-scale SSF bioreactors
involves long delays, strong non-linearities, loop
interaction, and varying responses from batch to batch.
Settings for the classical proportional integral derivative (PID) controller need to be adjusted often since
specific tuning is only effective for a short period in a
given batch [3] and, therefore, advanced control algorithms have a better chance of achieving optimum
bioreactor performance. A few studies report applications of advanced techniques for controlling SSF

123

400

Bioprocess Biosyst Eng (2006) 29:399407

bioreactors. A successful application of an adaptive

algorithm to control the temperature of the solid bed in
a laboratory-scale rocking reactor is described in [4].
Elsewhere, Dynamic Matrix Control (DMC) has been
tested in simulations using SSF lumped parameter and
distributed parameter models [5, 6], and in real-time
experiments [3]. Measurements in real SSF bioreactors
exhibit complicated noise patterns; however the models used in these simulation tests overlooked the effect
of noise. In our experience from experimental runs of a
pilot-scale SSF bioreactor, it has been impossible to
reproduce the good control performance observed in
simulations using models that do not explicitly include
the complex noise patterns inherent in the real process.
Here, we explore the effect realistic noise had on the
performance of model-based controllers, and assess
how suitable filters could compensate for the falloff in
performance due to noise. Specifically, linear Model
Predictive Control (MPC) and Internal Model Control
(IMC) algorithms were tested and their performance
compared in regulating the average bed temperature of
a simulated SSF reactor that included a real-time dataderived noise model.

Process description
Simulations were performed with a model of a 200 kg
capacity aseptic pilot SSF bioreactor. The bioreactor
was used for cultivation runs of the fungus Gibberella
fujikuroi on wheat bran to produce Gibberellic acid.
For optimum growth, bed temperature should be held
at 28C during the exponential phase, while for optimum production it should be held at 31C for the
stationary phase. Emulating common practice in largescale SSF bioreactors, evaporative cooling was adopted
to control temperature [7] with inlet air temperature
and relative humidity manipulated simultaneously.
We developed a realistic SSF bioreactor model
comprising two sub-models: a simplified lumped
parameter model to simulate the main features of the
dynamic response of this reactor [8] and a noise model
to mimic the outliers and non-stationary high frequency noise observed in reactor measurements [9]. A
brief summary of the model follows.
Measured biomass, Xtot, considering active (X) and
inactive fungus is expressed on a dry mass basis (kgdb):
dXtot
lX
dt

kgX =h kgdb :

The specific growth rate, l, considers the effect of

the limiting nutrient (available nitrogen) on biomass

123

growth, using Monods rate expression. The change in

available nitrogen, NI, is expressed by:



dNI
X
/
kl

YX=NI
dt

kgN =h kgdb ;

where / is the percentage of nitrogen the microorganism assimilates from wheat bran, k is the conversion
factor between wheat bran nitrogen and available
nitrogen and YX=NI is the mass yield between biomass
growth and available nitrogen. Wheat bran degradation, in turn, follows zero-order kinetics.
Bed temperature (Tb) dynamics are given by the
energy balance:


dTb
1

QFw DHr  Qg  Qwall  C/h

Cs
kexp
dt
3
Metabolic heat production, DHr, can be estimated
from the evolution of respiratory gases. The heat removed by gas, Qg, includes evaporation, forced convection between the bed and the airflow, and any
sensible heats associated with respiratory gases. Heat
losses through the reactor wall, Qwall, take into account
natural convection and radiation. Finally, QFw represents the enthalpy contribution of added water.
The noise model generates time-variant high frequency noise and outliers, which can be high, medium
or low. High and medium outliers are of constant
amplitude and are distributed uniformly throughout
the fermentation. In turn, low outlier amplitude distribution is normal and their distribution over time is
exponential. High frequency noise is generated in
clusters of varying amplitude and frequency.
The whole model was coded in Matlab/SimulinkTM.
Model parameters were taken from Araya [8] and
Bustamante [9].

Control algorithms
Model predictive control
We used the non-linear systems control block,
NLMPCSIM, from the Model Predictive Control
Toolbox of Matlab/Simulink. In this algorithm, control
computations are performed in two stages: (a) future
controlled variables are estimated from past measurements and control moves, and (b) an optimum trajectory is computed over a prediction horizon of P steps
applying C control moves (control horizon). At each
sampling time, the whole procedure is repeated and

Bioprocess Biosyst Eng (2006) 29:399407

401

only the first control move is applied to the plant, in a

receding horizon form with P and C the respective
tuning parameters. The minimization problem can be
formulated in the following form:
P
X

min

Duk ;...;DukC1

e
eki
WD

eki T

i1

C
X

Du
Duki1
WD

Duki1 T

i1

s.t.

uL  uki1  uU 8i 1; . . . ; C
yL  ^
yki  yU 8i 1; . . . ; P
Duki 0

8i C; . . . ; P

Here, ^
yk is the vector of predicted plant outputs at
time interval k, ek is the vector of output deviations
from the set points and Duk is the vector of control
moves. WeD and WDu
D are diagonal matrices with weights
penalizing output deviations from set points and control movements, respectively. The prediction horizon,
P, defines the period over which the cost function will
be minimized. In turn, the control horizon, C, establishes the length of the sequence of future control
moves. The minimization problem also includes constraints on inputs and outputs. In the expression above,
uL and yL represent the lower bounds, and uU and yU
represent the upper bounds. This optimization problem does not have an analytical solution, other than
when a linear predictive model is used that includes no
constraints. Thus, as a rule, the problem has to be
solved numerically. Further details regarding MPC
algorithms and their use within Matlab can be found in
Maciejowski [10].
In this work we obtained a continuous nominal linearized model of the SSF non-linear model using
Matlabs TRIM and LINMOD commands. Manipulated variables were considered constant over the entire fermentation run (Tgi = 25C and Hgi = 85%)
while the remaining input disturbances, such as pulsed
water addition and inlet air flowrate, were assigned
variable values from real pilot-scale experimental run
data [3]. We then applied the C2DMP and SS2MOD
Matlab commands to get the discrete linear model into
the required NLMPCSIM block format. Finally, the
MPC controller gain was computed with the MPCCON
command.
Split-range internal model control
The input variables Tgi and Hgi ought to be manipulated simultaneously to achieve good bed temperature
control. However, we wished to assess performance of

a single-input single-output (SISO) IMC control algorithm in the loop. So we applied a split-range scheme
to manipulate both inputs simultaneously, which were
constrained, as in the real process, between 20 and
30C for the inlet air temperature and between 55 and
100% for the inlet air relative humidity. To simplify the
application of the split-range algorithm, both inputs
were scaled as
Tgi

Tgi  20

0:5
30  20

Hgi

Hgi  55

0:5:
100  55

Therefore, given a control action normalized between 0 and 1, the manipulated variables are defined by
0  u  0:5
(


55  Hgi 55 10055
0:5
u  100
Tgi 20 C
0:5  u  1
(


20  Tgi 20 3020
0:5
u  30:
Hgi 100

When a given manipulated variable, say the inlet air

humidity, is saturated, the controller will start changing
the other input (the inlet air temperature in this case),
and vice versa. When maximum cooling is required,
Tgi takes the value 20C and Hgi = 55%, and when
maximum heating is required, Tgi will be 30C and
Hgi = 100%.
Then, a one degree of freedom SISO IMC controller
[11] was designed, taking u as the manipulated variable. Now, a transfer function that describes the dynamic behaviour of the bed temperature is needed.
Several positive and negative step changes were applied on the new manipulated variable (u) at different
simulation times, and a transfer function model was
fitted relating input and output deviations from a
nominal reference trajectory (defined by u = 0.5). A
first-order model was satisfactory for representing the
step responses although different model parameters
were obtained depending on the time and direction of
the step change. The range of models produced can be
represented by
ps

13:3  11:3
:
1:8  0:9
s 1

The control function, q(s), is given by the plant inverse in series with a first-order filter term,

123

402

qs

Bioprocess Biosyst Eng (2006) 29:399407

2:25s 1
1


:
24:6
es 1

Here, the nominal gain of the invertible model is the

maximum process gain, while the nominal time constant is the average value, as recommended elsewhere
[11]. In a preliminary step, the first-order filter
parameter, e, was tuned in order to reduce the control
gain at high frequencies (e = 0.1). This value also
achieves robust performance contemplating the variations in our model (Eq. 8).

Results
Noise model and filtering
Characteristics of the noise of each of the measured
signals in the SSF pilot bioreactor described above
were determined using data from nine fermentation
runs, each around 100 h [3]. Two kinds of disturbance
were found, outliers and high frequency noise. Outliers
were classified as high, medium and low, according to
magnitude. High frequency noise presented clusters
of different magnitudes. Further details are given in
Bustamante [9].
Bed temperature noise is the main concern in SSF
bioreactor control. Figure 1 shows an example of such
noise that was used in the simulations below.
Fig. 1 a Complete noise
signal (outliers plus high
frequency noise). b Clusters
of high frequency noise

Fig. 2 Noise signal filtered

with second-order
Butterworth: a cutoff
frequency of 0.12 Hz;
b cutoff frequency of 1.2 Hz;
c cutoff frequency of 12 Hz

123

Differences in outlier magnitudes are clear in Fig. 1a

while Fig. 1b illustrates the variability of high frequency noise amplitudes over time.
No matter what Butterworth filter cutoff frequency
is used the effect of the outliers on the filtered signal
was damaging (Fig. 2). The benefits of applying the
Hampel filter are readily apparent in Fig. 3: deviations
in the filtered signal were an order of magnitude
smaller. The two sets of figures also show the impact of
the cutoff frequency in the filtered signal. At first sight,
the best cutoff frequency would seem to be 0.12 Hz
since extremely smooth signals were obtained. Nevertheless, with this cutoff frequency an undesirable delay
was introduced in the measured signal, worsening
control performance severely. With a cutoff frequency
of 1.2 Hz we achieved a compromise between
smoothing and delay that resulted in good control
performance.
Model predictive control
Model predictive controls are widely used in industry
and, although a number of useful heuristics have been
developed, no formal rules have been established to
tune them. By way of example, Ricker et al. [12] suggested that the sampling period should be around 1/30
to 1/20 times the settling time and that the prediction
horizon, P, should be long enough to cover the settling
time. The same authors also advise using a small

Bioprocess Biosyst Eng (2006) 29:399407

403

Fig. 3 Noise signal filtered

with Hampel and then with
second-order Butterworth
filter: a cutoff frequency of
0.12 Hz; b cutoff frequency of
1.2 Hz; c cutoff frequency of
12 Hz

control horizon, C, of three to five sample times.

Nevertheless, based on our experience controlling a
pilot bioreactor [3], we set a sample time, Ts, of 5 min,
a prediction horizon, P, of 12 and a control horizon, C,
of 3 or less. As manipulation of the inlet air temperature was privileged over inlet air humidity (to reduce
evaporation and prevent over-drying of the bed)
the following input weights were used: WTi
D = 0.1 and
WHgi
=
0.01.
D
To illustrate the detrimental effect of noise in the
control, we will first show the controlled system without noise (Fig. 4). In this idealized, non-representative
system good control was achieved; the controlled variable (Fig. 4a) responded smoothly and, aside from the
period from 25 to 50 h when control actions were saturated, remained close to the set point throughout the
fermentation. The observable peaks were due to the
effect pre-established pulses of water had on bed
temperature. Water was added to replenish that lost
through evaporation [3]. Control actions were smooth
for most of the fermentation (Fig. 4b) only hitting the
bounds once the set point changed, and became saturated when the maximum cooling capacity of the system was reached. Here, the small peaks are controller
attempts to compensate for disturbances caused by the
pulses of added water.
When measurements contained the noise pattern
depicted in Fig. 1, applying MPC control without any

signal filtering resulted in instability and unacceptable

performance (not shown). After common practice in
industry, we therefore used a low-pass filter in the
simulation shown in Fig. 5, in which measured bed
temperature was contaminated with the noise pattern
shown in Fig. 1. To facilitate comparison between
different controllers and filters, the measured signal
(bed temperature) is displayed in the figures without
noise, i.e., we show the measured signal directly from
the SSF model block, prior to adding noise. Comparing
Figs. 4 and 5, even with low-pass filtering the deleterious effects of the outliers and high frequency noise to
control performance are clear. These disturbances
caused large excursions in bed temperature at 15 and
80 h (Fig. 5a) and an offset of 0.5C after the set-point
change. Moreover, despite reducing the control horizon (C = 1) to achieve asymptotically stable performance, the control actions appear noisier (Fig. 5b) and
hit the bounds three times (15, 50 and 80 h).
The benefits of shaving the outliers (using the
Hampel algorithm) prior to low-pass filtering are
illustrated in Fig. 6. The large excursions in bed temperature seen in Fig. 5b are absent. However, the
control system was still sluggish, presenting an offset
after the set-point change and the same large deviation
between 25 and 50 h. Aside from the strong control
movements at 15 and 80 h, the control signals shown in
Figs. 5b and 6b appear similar.

Fig. 4 Bed temperature

control with MPC without
noise and set-point change
(P = 12; C = 3). a Controlled
variable; b manipulated
variables

123

404

Bioprocess Biosyst Eng (2006) 29:399407

Fig. 5 Bed temperature

control with noise and MPC
using Butterworth filter with
cutoff frequency of 1.2 Hz
(P = 12; C = 1; WTgi = 0.1;
WHgi = 0.01). a Controlled
variable; b manipulated
variables

The effect of controller tuning using both the

Hampel and Butterworth filters in tandem was then
explored. The prediction horizon, P, had little effect on
controlled temperature when varied between 6 and 20
(not shown), though the controller proved more sensitive to the control horizon, C. At C = 2 or 3, for
instance, bed temperature is highly oscillatory and the
manipulated variables shatter (not shown).
Altering the input weights can produce significant
changes in behaviour in the controlled and input variables. For example, increasing the temperature weight
to WTi
D = 1 resulted in strong control actions of the
inlet air humidity while inlet air temperature only
moved appreciably when inlet air humidity was saturated (not shown). In addition, poor bed temperature
control performance was reflected in an offset as well
as in low frequency oscillations. When, on the other
hand, a smaller value of this weight, say WTi
D = 0.01
was used, the inlet air temperature moved more often,
and displayed high frequency oscillations while the
inlet air humidity changed very little (not shown). At
this temperature weight, save for the low amplitude
high frequency oscillations, the controlled variable
looks like that in Fig. 6.
Adjusting the inlet air humidity weight to say
WHgi
D = 0.001, the controlled variable presented more
offset and low frequency oscillations; inlet air humidity
Fig. 6 Bed temperature
control with noise and MPC
using Hampel and
Butterworth filter with cutoff
frequency of 1.2 Hz (P = 12;
C = 1; WTgi = 0.1; WHgi =
0.01). a Controlled variable;
b manipulated variables

123

moved much more than the temperature (not shown).

Increasing the humidity weight to WHgi
D = 0.1 did not
alter control performance much (Fig. 7a), but produced stronger high frequency oscillations in the inlet
air while humidity moved very slowly (Fig. 7b), except
when the temperature was saturated. Here, the
behaviour of the inlet air humidity is worrying. The
input weights only define how much a given manipulated variable will be moved, but they do not tell the
MPC algorithm the direction of those movements.
Consequently, in this case Fig. 7b shows that after the
set point changed the humidity remained low and
therefore cooled the bed while, simultaneously, the
temperature took on a high value causing bed heating.
The net result of both competing actions, however, is
that the bed got cooler than desired (lower than the set
point).
Since control action strength is mainly dependent on
the relative values of the control weights, reducing the
weight of the inlet air humidity has a similar effect to
increasing the weight of the inlet air temperature, and
vice versa.
Internal model control
Just as in the MPC simulations, here we first show
noise-free control performance (Fig. 8).

Bioprocess Biosyst Eng (2006) 29:399407

405

Fig. 7 Bed temperature

control with MPC using
Hampel and Butterworth
second-order filter with cutoff
frequency of 1.2 Hz (P = 12;
C = 1; WHgi = 0.1; WHgi =
0.1). a Controlled variable;
b manipulated variables

Internal model control achieved even better control

than MPC (compare with Fig. 4). Under IMC, the
controlled variable (Fig. 8a) responded smoothly and
remained close to the set point throughout the fermentation, except when control actions were saturated
(3050 h). Peaks due to the effect of pre-set pulses of
added water are almost imperceptible. Control actions
were also smooth throughout the fermentation
(Fig. 8b) and were only saturated when the system
reached its maximum cooling capacity. IMC compensated far better than MPC for water addition pulseinduced disturbances.
Noise also affected IMC control performance, although less deleteriously than with MPC. All the same,
unfiltered noisy bed temperature measurement resulted in poor performance (not shown). Figure 9
shows the results obtained with IMC control using just
a low-pass filter. Comparing Figs. 8 and 9, we see a
larger deviation in bed temperature from 25 to 50 h
(Fig. 9a), an offset, albeit small, between 10 and 25 h,
and large deviations from 70 h onwards. Control actions were also noisier (Fig. 9b). We had to detune the
IMC controller by increasing the filter constant, e, to
0.8 to account for the delay introduced by the filter to
avert unstable control (not shown).
The best control for noisy measurements was
achieved upon applying full filtering, Hampel plus

Butterworth, (Fig. 10), which let us reduce the value of

e. Small deviations in bed temperature were coupled
with smooth control movements. Again, covering the
same conditions, we observe much better performance
than from MPC (Fig. 6). The improved performance of
IMC is due to the split-range algorithm that, depending
on the needs of the control system, ensured both
manipulated variables were always either cooling or
heating. At times, under MPC, as shown in Fig. 7, one
manipulated variable was heating while the other was
cooling thus reducing overall control capacity and
performance.

Discussion
The control strategy presented here should simplify,
though not entirely solve, the scaling up of SSF processes from laboratory experiments to viable commercial production. Since it is usually not possible to
generalize control and filtering performance results
when dealing with non-linear systems [13], it is advisable first to test and adapt the strategy to the specific
application. We recommend developing a realistic
model, similar to the one described here, even if
accurate model parameters and noise patterns are not
available. Then the engineer can assess and consider

Fig. 8 Bed temperature

control with IMC without
noise (e = 0.1). a Controlled
variable; b manipulated
variables

123

406

Bioprocess Biosyst Eng (2006) 29:399407

Fig. 9 Bed temperature

control with noise and IMC
using Butterworth filter with
cutoff frequency of 1.2 Hz
(e = 0.8). a Controlled
variable; b manipulated
variables

Fig. 10 Bed temperature

control with noise and IMC
using Hampel and secondorder Butterworth filter with
cutoff frequency of 1.2 Hz
(e = 0.5). a Controlled
variable; b manipulated
variables

the impact that input variable constraints and expected

disturbances have on the design.
We are confident that this strategy will improve the
performance of existing large-scale intermittently
mixed SSF bioreactors: better bed temperature control
will result in higher biomass and product yields. To
apply this control strategy in existing plants, process
linear models will have to be identified directly from
real process noisy data. Several textbooks describe the
procedure [10, 11]. In this regard, IMC is simpler to
implement than linear MPC, because transfer function
models are easier to identify from the real data than
the Finite Impulse Response (FIR) and Finite Step
Response (FSR) models that MPC algorithms use. A
further point is that IMC can be designed to ensure
robust performance even if the transfer function model
parameters vary [10].
It seems that this is the first paper to address
explicitly and simultaneously the main difficulties
encountered in the control of temperature of largescale SSF bioreactors, i.e., non-linear dynamics, interactions and constraints in manipulated variables, and
complicated noise patterns. Using a more realistic
model to represent the complex behaviour observed in
real measurements from large-scale intermittently
mixed SSF bioreactors, we showed that linear predictive controllers can control average bed temperature

123

reasonably well. In particular, we found that proper

noise handling, i.e., filtering high frequency noise and
outlier shaving, is fundamental to achieve good control
performance in large-scale SSF bioreactors. Nonetheless the extra delay the low-pass filter introduced did
require the control algorithms to be detuned.
Despite the advances in control techniques put forward in this study, good temperature control alone is
insufficient to achieve optimum and reproducible
operation of industrial-scale SSF bioreactors. The most
important of the remaining variables (beyond the
scope of this paper) that needs to be well controlled is
average bed water content. Until recently the main
obstacle for controlling this variable has been the lack
of reliable and affordable moisture sensors suitable for
SSF bioreactors. Nevertheless, Khanahmadi et al. [14]
have described a virtual humidity sensor applicable to
SSF bioreactors that is both reliable and affordable.
They also advanced a practical method for controlling
bed water content for systems in which bed water
activity is not overly sensitive to water content.
Otherwise, another control system would need to be
designed and assessed, based for example, on a realistic
model such as the one developed here.
Another issue our simulations overlook is the effect
of process noise disturbances arising from bed
channelling, bed mixing and gas occlusion. These

Bioprocess Biosyst Eng (2006) 29:399407

disturbances, associated with bed heterogeneity, can

be simulated well with an SSF-distributed parameter
model, such as the one developed by Von Meien and
Mitchell [15]. Nevertheless, further research is still
needed to design practical strategies to minimize the
impact these disturbances have on SSF bioreactor
control performance.
Finally, optimum bioreactor operation requires a
number of other variables such as biomass, nutrients
and product contents in the solid bed to be controlled,
too. This is difficult since these variables cannot be
measured reliably with affordable instruments. Estimation and control of these variables have been
implemented in a laboratory-scale SSF bioreactor [16]
using the Extended Kalman Filter, a widely used state
estimation technique. However, as yet these estimations have not been reported for large-scale SSF bioreactors, which involve far more complex noise
patterns. The realistic model adopted here may also
prove useful in the development of an estimation/
control system for these variables.
Acknowledgments This work has been supported by project
Fondecyt#1030325/2003. We very much appreciate the assistance
of Alex Crawford in improving the style of the text and the
valuable suggestions of the anonymous referees.

References
1. Raimbault M (1998) General and microbiological aspects of
solid substrate fermentation. Electron J Biotechnol 1:2627
2. Perez-Correa JR, Agosin E (1999). Solid state fermentation,
automation of solid substrate fermentation processes. In:
Flickinger M (ed) Encyclopedia of bioprocess technology:
fermentation, biocatalysis and bioseparation. Wiley, New
York, pp 24292446
3. Fernandez MA (2001) Control automatico de un bio-reactor
piloto para cultivos sobre substrato solido. PhD thesis, Universidad de Chile. 247 pp

407
4. Ryoo D, Murphy VG, Karim MN, Tengerdy RP (1991)
Evaporative temperature and moisture control in a rocking
reactor for solid substrate fermentation. Biotechnol Tech
5:1924
5. Pajan H, Perez-Correa R, Solar I, Agosin E (1997) Multivariable model predictive control of a solid substrate pilot
bioreactor: a simulation study. In: Wise D (ed) Global
environmental biotechnology. Kluwer, Boston, pp 221232
6. von Meien O, Luz LF, Mitchell DA, Perez-Correa JR, Agosin E, Fernandez-Fernandez M, Arcas JA (2004) Control
strategies for intermittently-mixed, forcefully-aerated solidstate fermentation bioreactors based on the analysis of a
distributed parameter model. Chem Eng Sci 59:44934504
7. Fernandez M, Perez-Correa JR, Solar I, Agosin E (1996)
Automation of a solid substrate cultivation pilot reactor.
Biopro Eng 16:14
8. Araya M (2006) Estimacion de parametros y reconciliacion
dinamica de datos en sistemas de cultivo sobre sustrato
solido. Masters thesis. Pontificia Universidad Catolica de
Chile. 190 pp
9. Bustamante, CR (2005) Diseno y evaluacion de una
estrategia de control predictivo basado en modelo no lineal
para un bio-reactor piloto de cultivo sobre sustrato solido.
Masters thesis, Universidad de Talca. 129 pp
10. Maciejowski JM (2002) Predictive Control with Constraints.
Prentice Hall, Englewood Cliffs, 352 pp
11. Brosilow C, Joseph B (2002) Techniques of model-based
control. Prentice Hall, Upper Saddle River, 704 pp
12. Ricker NL, Morari M, Bemporad A (1998). The MPC
graphical user interface. Available from http://depts.washingtong.edu/control/LARRY/GUI/
13. Pearson RK (2006) Nonlinear empirical modeling techniques. Comput Chem Eng (in press, available online)
14. Khanahmadi M, Roostaazad R, Mitchell DA, Miranzadeh
M, Bozorgmehri R, Safekordi A (2006) Bed moisture estimation by monitoring of air stream temperature rise in
packed-bed solid-state fermentation. Chem Eng Sci 61:5654
5663
15. Von Meien OF, Mitchell DA (2002) A two-phase model for
water and heat transfer within an intermittently-mixed solidstate fermentation bioreactor with forced aeration. Biotech
Bioeng 79:416428
16. Ryoo D (1990) On-line estimation and control in solid substrate fermentation. PhD thesis, Colorado State University

123