Professional Documents
Culture Documents
Bernard (2011) Dynamical Model Development and Parameter Identification For An Anaerobic Wastewater Treatment Process
Bernard (2011) Dynamical Model Development and Parameter Identification For An Anaerobic Wastewater Treatment Process
Abstract: This paper deals with the development and the model of the process is required for the design of such
parameter identi®cation of an anaerobic digestion pro- monitoring and control algorithms.
cess model. A two-step (acidogenesis±methanization)
mass-balance model has been considered. The model The dynamical modeling of anaerobic digestion has
incorporates electrochemical equilibria in order to in- been an active research area during the last three decades.
clude the alkalinity, which has to play a central role in the Andrews (1968) introduced the Haldane model to char-
related monitoring and control strategy of a treatment acterize growth inhibition that can emphasize the process
plant. The identi®cation is based on a set of dynamical instability (i.e., the biomass wash-out via the accumula-
experiments designed to cover a wide spectrum of op-
erating conditions that are likely to take place in the tion of acids). A model with a single bacterial population
practical operation of the plant. A step by step identi®- was then proposed (Graef and Andrews, 1974).
cation procedure to estimate the model parameters is The representation of the process was then improved
presented. The results of 70 days of experiments in a 1- by considering three stages: solubilization of organics,
m3 fermenter are then used to validate the model. ã 2001 acidogenesis, and methanogenesis (Hill and Barth,
John Wiley & Sons, Inc. Biotechnol Bioeng 75: 424-438, 2001.
Keywords: anaerobic digestion; dynamical modeling; 1977). Mosey (1983) introduced a four-population
model identi®cation model with two acidogenesis reactions and two meth-
anization reactions which also emphasizes the role of
hydrogen. These main modeling studies have then been
INTRODUCTION extended and detailed by other authors in order to get
The anaerobic wastewater treatment process presents closer to the complexity of the process (Moletta et al.,
very interesting advantages compared to the classical 1986; Batstone et al., 1997; Costello et al., 1991a; Cos-
aerobic treatment (Mata-Alvarez et al., 2000; Pavlosta- tello et al., 1991b; Fernandes et al., 1993; Kiley et al.,
this, 1994): It has a high capacity to degrade concentrated 1997). It results in detailed models of the process which
and dicult substrates (plant residues, animal wastes, include several bacterial populations and several sub-
food industry wastewater, and so forth), produces very strates. As a result, the number of parameters in these
few sludges, requires little energy, and, in some cases, it models can become very large.
can even recover energy using methane combustion. But As suggested in the above paragraph, there exists a
in spite of these advantages, the anaerobic treatment wide range of models dealing with anaerobic digestion.
plants are still very rare at the industrial scale, probably However, the models describing with detail all the pro-
because they are known to become easily unstable under cesses responsible for anaerobic digestion are generally
some circumstances, such as variations of the process dicult to use for control purposes (Bastin and Do-
operating conditions (Fripiat et al., 1984). Nevertheless, chain, 1990). In addition, the question of model identi-
this drawback can be overcome by associating a moni- ®cation and validation is rarely performed in a
toring procedure with a decision support system that suciently large range of operating conditions (typi-
allows enhancement of the stable performance of the cally, loading rates and retention times). Moreover, in
online wastewater treatment operation via a feedback all these models, the considered process is often assumed
control loop (Dochain et al., 1991; Perrier and Dochain, to behave like a continuously stirred tank reactor
1993; Steyer et al., 1999). Therefore, a reliable dynamic (CSTR). In practice, the technologies often aim at in-
creasing the contact surface between the biological
phase and the organic matter in order to improve the
Correspondence to: Olivier Bernard process eciency. As a consequence, the technology
m2/m3). The dilution of the in¯uent is performed by Measurement of the Chemical Oxygen Demand
adding water to 20 L of vinasses (measured with the (COD)
¯owmeter) in a 200-L tank. The feeding tanks are
The principle of chemical oxygen demand (COD) mea-
equipped with level sensors that allows obtaining of a
surement (NF T 90-101) is the oxidation of the organic
selected concentration in the in¯uent. The pH is mea-
matter by a potassium bichromate excess, in acid media
sured and controlled in the feeding tank. Figure 1 de-
(H2SO4) at boiling temperature. The oxidant excess is
scribes the pilot plant and the measurement devices.
titrated by a reducing solution of Mohr salt (ammonium
and ferrum sulfate).
The Online Measurements
The temperature inside the reactor is controlled at 35°C. Measurement of the Alkalinity
The temperature regulation is performed in the recycle Acid (HCl) is added to the sample in order to reach pH
loop via an electric heater using a PID controller. The = 5.75 (the volume titrated corresponds to partial al-
in¯uent ¯ow rate is measured by an electromagnetic kalinity). Then, acid is added again until the pH reaches
sensor (Khrone). the value of 4.3 and the total added acid volume is the
The gas analysis loop (see Fig. 1) consists of a dryer total alkalinity. The concentration of acetate and bi-
that eliminates the humidity by cooling the gas. The carbonate can be determined from partial and total al-
Ultramat 22P sensor (Siemens) measures the CO2/CH4 kalinity (Ripley et al., 1986).
percentage of the analyzed gas. The gas ¯owmeter is
located at the output of the loop. It uses an electro-
magnetic ¯oater to continuously measure the produced The Experimental Protocol
gas ¯ow rate.
The experimental protocol has been determined in order
to cover a wide range of organic loading rates and to
The Of¯ine Measurements obtain situations close to the destabilization of the fer-
menter. This is performed via consecutive step variation
The samples used to determine the concentrations in the of both the dilution rate and the in¯uent COD. They are
inlet are taken in the pipe just after the feeding pump. maintained constant for a suciently long period of
The samples for the outlet concentrations are taken just time in order to reach a steady state. The in¯uent time
before their rejection to the sewer. The samples are evolutions are presented in Fig. 2. Note that some fail-
stored at 4°C. The dissolved part is obtained after cen- ures (leaks, pump failures, tube clogging, and so forth)
trifugation for 15 min at 15,000 rpm. have slightly disturbed the initial protocol.
426 BIOTECHNOLOGY AND BIOENGINEERING, VOL. 75, NO. 4, NOVEMBER 20, 2001
speci®c growth rates of acidogenesis and methanization,
respectively.
Chemical Species
The Inorganic Carbon
Let us consider the chemical reactions involving the
inorganic carbon mainly composed of dissolved CO2,
3 bicarbonate (B), and carbonate in line with Rozzi
(1984). In normal operating conditions, the pH range is
between 6 to 8, and the temperature is between 35 and
38°C. In those conditions, the anity constant for car-
bonate/bicarbonate (Kc = 4.7 ´ 10)11 mol/L) indicates
that the carbonate concentration will remain negligible
compared to the bicarbonate. The total inorganic car-
bon C in the considered pH range is then approximately
Figure 2. In¯uent pro®les during the experiment. The values of COD, equal to:
VFA, and pH are extrapolated for any time from oine measure-
ments. The values of the dilution rate result from online measurements. C CO2 B
3
The time periods used for the parameter identi®cation are represented
with a thick line on the time axis. and the bicarbonate and dissolved CO2 concentrations
are determined by the following chemical reaction (H
directly linked to the model complexity. As one of our are the protons):
objectives is to obtain a model that would be able to
represent the destabilization phenomenon while being B H CO2 H2 O
4
identi®able, we assume that the bacterial populations we denote Kb the anity constant of this reaction (Kb =
can be divided into two main groups of homogeneous 6.5 ´ 10)7 mol/L):
characteristics, and that the anaerobic digestion can be
described by a two-stage process. In the ®rst step (aci- H B
Kb
5
dogenesis), the acidogenic bacteria (X1) consume the CO2
organic substrate (S1) and produce CO2 and volatile
fatty acids (S2). The population of methanogenic bac-
teria (X2) uses, in a second step, the volatile fatty acids The Volatile Fatty Acids
as substrate for growth, and produce CO2 and methane. The total concentration of VFA is composed of ions S
On the basis of hydrodynamical tests, we assume that (mainly acetate) and un-ionized SH (mainly acetic acid):
the reactor behaves like a perfectly mixed tank, and that
the biomass is uniformly distributed within the reactor. S2 SH S
6
The corresponding anity constant is equal to:
Biological Reaction Pathways H S
Ka
7
The acidogenic and methanogenic bacteria intervene in SH
the two following biological reactions:
The numerical value of Ka in the considered pH range
· Acidogenesis (with reaction rate r1 = l1X1): (Ka = 1.5 ´ 10 5 mol/L) shows that [SH] is negligible
r1
and therefore:
k1 S1 ! X1 k2 S2 k4 CO2
1
S2 ' S
8
· Methanization (with reaction rate r2 = l2X2):
r2
k3 S2 ! X2 k5 CO2 k6 CH4
2 The Ion Balance
S1 represents the organic substrate (and its concentra- The total alkalinity Z is de®ned as the sum of dissoci-
tion) characterized by its COD (g/L). The total con- ated acids in the medium:
centration of VFA is denoted S2 (mmole/L). In the
Z B S
9
sequel, we assume that S2, which is mainly composed of
acetate, propionate, and butyrate, basically behaves like From Eq. (8), we have in the considered pH range:
pure acetate. It is important to note that the total COD
is composed of S1 and S2. l1 and l2 (d)1) represent the Z ' B S2
10
428 BIOTECHNOLOGY AND BIOENGINEERING, VOL. 75, NO. 4, NOVEMBER 20, 2001
dZ 2 3 2 3
D
Zin Z
22 0 0
dt 6 7 6 7
6 7 6 7
6 0 7 6 0 7
dS1 6 7 6 7
D
S1in S1 k1 l1
nX1
23 6 7 6 7
6 7 6 7
dt 6 DZin 7 6 0 7
6 7 6 7
F6
6
7;
7 Q6
6
7;
7
dS2 6 DS1in 7 6 0 7
D
S2in S2 k2 l1
nX1 k3 l2
nX2
24 6 7 6 7
dt 6 7 6 7
6 7 6 7
6 DS2in 7 6 0 7
6 7 6 7
4 5 4 5
dC
D
Cin C qC
n k4 l1
nX1 k5 l2
nX2 DCin qC
n
dt
25 2 3
aD 0 0 0 0 0
6 7
6 7
with 6 0 aD 0 0 0 0 7
6 7
6 7
6 7
qC
n kL aC S2 Z KH PC
n
26 6 0 0 D 0 0 0 7
6 7
D6
6
7
7
32
6 0 0 0 D 0 0 7
where PC(n) is computed from Eqs. (12), (16), (18), and 6 7
6 7
(19) as follows: 6 7
6 0 0 0 0 D 0 7
q 6 7
4 5
/ /2 4KH PT
C S2 Z 0 0 0 0 0 D
PC
n
27
2KH
The model given by Eq. (30) will serve as a basis for the
with design of online monitoring and control strategies of the
anaerobic digestion process (Bastin and Dochain, 1990).
k6 For this purpose, the modeling of the growth rates l1(n)
/ C S2 Z KH PT l
nX2 and l2(n) is not required. Yet, for numerical simula-
kL a 2
tions, analytical expressions for the growth rates are
S1in (g COD/L), S2in (mmole/L), Cin (mmole/L), and Zin needed. In the following section, expressions for l1(n)
(mmole/L) are the in¯uent concentrations of S1, S2, C, and l2(n) are proposed.
and Z, respectively.
Moreover, we have the following model equations for Modeling of the Bacterial Kinetics
the methane gas ¯ow rate and for the pH from Eq. (12)
and Eqs. (7), (10), and (19): The modeling of biological kinetics is a dicult task for
which a systematic methodology is still lacking. For the
qM
n k6 l2
nX2
28 sake of model simplicity, and in line with other works on
anaerobic digestion modeling, we shall consider the
following models for bacterial kinetics.
C Z S2
pH
n log10 Kb
29
Z S2
Acidogenic Bacteria
The model can then be rewritten in a more general We consider Monod-type kinetics for the growth of
matrix form: acidogenic bacteria; that is:
dn S1
Kr
n Dn QF
30 l1
S1 l1max
33
dt S1 KS1
where
where l1max is the maximum bacterial growth rate, and
2 3 2 3 KS1 is the half-saturation constant associated with the
X1 1 0
6 X2 7 6 0 1 7 substrate S1.
6 7 6 7
6Z7 l1
nX1 6 0 0 7
n6 7
6 S1 7; r
n ; K6
6
7
6 7 l2
nX2 6 k1 0 7 7 Methanogenic Bacteria
4 S2 5 4 k2 k3 5
In order to emphasize the possible VFA accumulation,
C k4 k5
we have considered Haldane kinetics for the methano-
31 genesis:
430 BIOTECHNOLOGY AND BIOENGINEERING, VOL. 75, NO. 4, NOVEMBER 20, 2001
that, when no biomass measurements are available, the Then, the possible steady states for S2 are solutions of
yield coecients are not identi®able and one can obtain the following equation, deduced from Eqs. (36) and (34):
only ratios of yield coecients.
S22 l2 max
1 S2 KS2 0
39
KI2 aD
Principles for Identi®cation
We have split the data set into one set for parameter We denote S2 and Sy2 the lowest and the largest solutions
calibration and one set for parameter validation. One of of Eq. (39), respectively. We also denote E and Ey the
our primary goals was to have a model able to predict corresponding equilibria. Note that E corresponds to
properly the process steady states. Therefore, we have a steady state in the inhibition phase of the methano-
selected a set of steady-state values for parameter cali- genesis.
bration. We have then used the data corresponding to The computation of the equilibrium for Z is
the other steady states and to the transients for model straightforward from Eq. (22):
validation.
Z Zin
40
Note that this approach is consistent and perfectly
valid from an identi®cation point of view. The model
structure is typically composed of the combination of
hydrodynamics terms, liquid±gas terms, and conversion Steady State of the Biomasses
(kinetic + yields) terms. The conversion and liquid±gas Using Eqs. (23) and (35), we get:
transfer terms contain all the parameters to be cali-
brated, while the terms related to the hydrodynamics are 1
X1
S1in S1
41
typically characterized by the (known) values of the in- ak1
¯uent and euent ¯ow rates.
In the next section, the steady-state values of the From Eqs. (24), (35), (36), and (41), we have two pos-
model variables are computed with respect to the pa- sible values for X2:
rameters, in order to be used later on in the model
1 k2
identi®cation procedure. X2 S2in S2
S1in S1
42
ak3 k1
p with
l2max KI2
D p p
38
a KI2 2 KS2 w Cin Z S2 k4 aX1 k5 aX2
KH P2
C x PC PT w 0
47 1 a a 1
KS1
49
D l1max l1max S1
where
This relationship can be used with the measurements of
kL a D
the equilibrium values of S1, S1 , to estimate the para-
x KH PT w k6 aX2
kL a meters a/l1max and KS1 via a linear regression. Un-
fortunately, the parameters a and l1max cannot be
We know from Eq. (17) that only the lower root is distinguished from this relationship. We chose therefore
physically admissible. Thus: to select values of l1max from classical bibliographical
p results (Ghosh and Pohland, 1974).
x x2 4KH PT w
PC
48 Equation (36) provides the following relationship:
2KH
1 a a 1 1 a
The steady-state values CO2 and qC can then be directly KS2 S2
50
D l2max l2max S2 KI2 l2max
derived from Eqs. (46) and (14), respectively.
The steady-state values associated with E can be Linear regression then gives the values of the following
computed by using a similar procedure, and simply by parameters: a/l2max, KS2, and KI2. Using the estimated
replacing the symbol * by in Eqs. (44) to (48). value of a obtained in the previous step, we then get l2max.
432 BIOTECHNOLOGY AND BIOENGINEERING, VOL. 75, NO. 4, NOVEMBER 20, 2001
k6 k2 k5 k4 This means that all the variables but the biomasses
; ; ;
56
k3 k1 k3 k1 depend only on the ratio of yield coecients. The values
of the yield coecients themselves (and not of their ra-
are identi®ed. Then, in a second step, we use the mea- tio) are then needed only if we want to have an estimate
surements of the VSS to obtain an approximation of of the biomasses in the fermenter.
each yield coecient. For that purpose, we need additional information and
We ®rst consider the methane gas ¯ow rate, which we measurements related to the biomasses. We propose ®rst
compute by combining Eqs. (42) and (44): to use the ratio m of acidogenic and methanogenic bac-
teria. This information is quite qualitative and can be
k6 k2
qM D S2in S2
S1in S1
57 determined despite the heterogeneity between the liquid
k3 k1
and the solid phases. We propose also to use the VSS
From this regression, we get the ratio of yield coe- concentration as a (rough) indicator of the total biomass
cients k6/k3 and k2/k1. X1 + X2.
From the consideration of the CO2 ¯ow rate (45), From Eqs. (41) and (42) we have:
using Eqs. (41) and (42), we obtain:
X1 1 S1in S1
m '
64
k4 k5 k2
X1 X2 ak1 VSS
qC D Cin C
S1in S1
k1 k3 k1
If we assume that m remains approximately constant, we
®nally have an estimate of k1:
k5
S2in S2
58
k3 1
1 S1in S
k1
65
We rewrite this equation as follows: am VSS
The value of m has been taken from the literature (m =
qC k4
S1in k5 qM
Cin C S1
59 0.2) (Sanchez et al., 1994).
D k1 k6 D Now, if we consider Eqs. (41) and (42), we have:
This regression gives the values of k4/k1 and k5/k6.
X1 k3 S1in S1
m
X1 X2 k1
S2in S2
kk21 kk31
S1in S1
Determination of the Yield Coef®cients
66
In this second step, we show how to estimate each yield and thus:
coecient. It turns out that the yield coecients are not
identi®able if we do not measure the biomasses. Indeed, m S2in S2 k2
k3 k1
67
it can be veri®ed that if we rescale the biomass by the 1 m S1in S1 k1
factors k1 and k2,
An estimate of k3 can then be found from Eq. (67). The
X01 k1 X1
60 ratios identi®ed in the previous step can now be used to
derive estimates of k2, k4, k5, and k6.
X02 k2 X2
61 Note that these values have to be considered with care
because of the uncertainty of the VSS measurements, the
then the biomass rescaling can be compensated by the uncertainty of the ratio of methanogenic and acidogenic
following parameter rescaling: bacteria, and, last but not least, the uncertainty of the
correlation between the total biomass and the VSS.
k1 k2 k4
k01 ; k02 ; k04
62
k1 k1 k1
Sensitivity Analysis
k3 k5 k6
k03 ; k05 ; k06
63 In this section, we study the sensitivity of the model to
k2 k2 k2
its parameters. Note that there does not exist any
The numerical simulation of the model with the yield methodology to discuss the parameter sensitivity in
coecients ki and k0i will give the same values for all the general: The usual methods refer to parameter sensitiv-
model variables (except for the variables X1 and X2 that ity for a given system trajectory (i.e., in reference to a
are not measured). The yield coefficients are not iden- given set of parameters, initial conditions, and in¯uent
tifiable if no measurements of the biomass is available; ¯ow rates). Therefore, it is of great importance to cor-
this is consistent with the study of Chappell and God- rectly choose the reference simulation from which the
frey (1992), who proved a similar result when only the sensitivity analysis is performed. This results in fact
biomass is measured. from an iterative approach, where the ``best parameter
Z
Dp 1 tf y
p Dp; x0 ; u; s y
p; x0 ; u; s
ry ds
tf 0 y
p; x0 ; u; s
the considered periods are presented in Table II. The
where y(p, x0, u, s) denotes the simulated value at time s
estimated standard deviation for some parameters is
of the variable y associated with the parameter p, the
quite high. However, this value is probably overesti-
initial condition x0, and the input u.
mated if we keep in mind the relatively small number of
In the sequel, we have focused the discussion on the
equilibria (seven points) from a statistical point of view.
sensitivity of the four following quantities to parameter
Note also that the deviations are particularly high for
variations: S1, S2, qC, and qM.
two classes of parameters. First, the estimates of the
The results are presented in Fig. 3. Note ®rst that the
kinetic coecients suer from the already mentioned
cascade model structure has a strong in¯uence on the
lack of reliability of the kinetic expression used. Indeed,
parameter sensitivity. Indeed, the only parameters in-
the fact that the expressions retained for the biological
¯uencing S1 are l1max, KS1, k1, and a. The parameters
kinetics are only rough approximations results in high
in¯uencing S2 are those in¯uencing S1 plus l2max, KS2,
variability of the corresponding parameter values. The
KI2, k2, and k3. These parameters (plus k6) also in¯uence
other group of parameters for which the estimates seem
qM. Finally, all the parameters act on qC. Neverthe-
to be less precise are the ratio of parameter related to k1
less, the in¯uence of the parameters only related to S1
(k2/k1 and k4/k1). This is probably due to the fact that
is much lower on S2 and on the gaseous ¯ow rates.
the composition of the substrate S1 is changing
Similarly, the parameters in¯uencing S2 have less eect
throughout the experiment. As a consequence, the yield
on qC.
coecient associated to its degradation (i.e., k1) may
From this study, it results that the parameters that
¯uctuate during the considered period. As we shall see in
played the main role are a (because it modi®es the dy-
the sequel, in spite of this apparent uncertainty, the
namics of the whole model) and k3 (which has a strong
model correctly ®ts the data.
in¯uence on the gaseous ¯ow rates). Note that the small
Tables III and IV summarize the obtained kinetic
16 values of l2max and KI2 also strongly change the model
parameters and yield coecient ratio values. Table V
predictions: with low value the equilibrium S2 becomes
then gives the values obtained for all the yield coe-
less and less stable.
cients.
The parameters k2, k4, kLa, KS1 and KS2 have little
in¯uence on the model, and therefore they will be less
precisely estimated.
Finally, it can be noted that sensitivity analysis (in %) Table III. Estimates of the kinetic parameters.
for the ratio of yield parameters ki/kj is the same as that
of ki (in %). Parameter Meaning Unit Value SDa
434 BIOTECHNOLOGY AND BIOENGINEERING, VOL. 75, NO. 4, NOVEMBER 20, 2001
Figure 3. Sensitivity for the model parameters. The mean changes of S1, S2, qC, and qM are represented with respect to the deviation of the
nominal value of the considered parameter.
436 BIOTECHNOLOGY AND BIOENGINEERING, VOL. 75, NO. 4, NOVEMBER 20, 2001
1198). It also presents research results of the Belgian Pro-
gramme on Inter-University Poles of Attraction, initiated by
the Belgian State, Prime Minister's oce for Science, Tech-
nology, and Culture. The scienti®c responsibility rests with its
authors.
NOMENCLATURE
B bicarbonate concentration (mmol/L)
C, Cin total inorganic carbon concentration (mmol/L)
D dilution rate (d 1 )
d/dt time derivative
k1 yield for substrate degradation
k2 yield for VFA production (mmol/g)
k3 yield for VFA consumption (mmol/g)
k4 yield for CO2 production (mmol/g)
k5 yield for CO2 production (mmol/g)
k6 yield for CH4 production (mmol/g)
Ka, Kb equilibrium constants (mol/L)
Figure 6. Comparison between measured VSS and simulated total KH Henry's constant (mmol/L per atm)
biomass (X1 +X2). The periods considered for the calibration step are kLa liquid±gas transfer constant (d 1 )
underlined. KI2 inhibition constant (mmol/L)
KS1 half-saturation constant (g/L)
KS2 half-saturation constant (mmol/L)
the main modeling uncertainties and that avoids the PC CO2 partial pressure (atm)
plant destabilization (Bernard et al., 2001). PT total pressure (atm)
qC carbon dioxide ¯ow rate (mmol/L per d)
qM methane ¯ow rate (mmol/L per d)
r, r1, r2 reaction rates (d 1 )
CONCLUSION
S1, S1in organic substrate concentration (g/L)
S2, S2in volatile fatty acids concentration (mmol/L)
In this paper, we have built, identi®ed, and validated a
X1 concentration of acidogenic bacteria (g/L)
model for an anaerobic treatment plant. The four fol- X2 concentration of methanogenic bacteria (g/L)
lowing points are important, because they guarantee Z, Zin total alkalinity (mmol/L)
that our model can be useful to monitor and control the Z0 anion concentration (mmol/L)
anaerobic process. a fraction of bacteria in the liquid phase
m mean fraction of acidogenic bacteria
l1 speci®c growth rate of acidogenic bacteria (d 1 )
1. It is based on mass-balance considerations. The l1max maximum acidogenic bacteria growth rate (d 1 )
modeling uncertainty due to the variability of the l2 speci®c growth rate of mathanogenic bacteria (d 1 )
biological kinetics is concentrated in the reaction l2max maximum methanogenic bacteria growth rate (d 1 )
rates terms. n vector of the process variables
2. An identi®cation procedure privileging the steady-
state predictions has been developed which allows
identi®cation of all the parameters of the model and References
understanding of the role played by the parameters Andrews J. 1968. A mathematical model for the continuous culture of
in the process dynamics. microorganisms utilizing inhibitory substrates. Biotechnol Bioeng
3. Experiments have been designed, covering a wide 10:707±723.
range of experimental conditions, in order to develop Bastin G, Dochain D. 1990. On-line estimation and adaptive control of
bioreactors. Amsterdam: Elsevier.
and validate the model. This diversity is obtained via
Batstone D, Keller J, Newell B, Newland M. 1997. Model development
various organic loading rates (given by various dilu- and full scale validation for anaerobic treatment of protein and fat
tion rates and various in¯uent COD concentrations), based wastewater. Water Sci Technol 36:423±431.
but it also results from a wide range of substrate Bernard O, Hadj-Sadok Z, Dochain D. 1999. Dynamical modelling
compositions, because the vinasses used during the and state estimation of anaerobic wastewater treatment plants. In:
Proceedings of ECC99 (CD-ROM). Germany: Karlsruhe.
experiment do not all have the same origin.
Bernard O, Polit M, Hadj-Sadok Z, Pengov M, Dochain D, Estaben
4. The validation of the model has been performed for M, Labat P. 2001. Advanced monitoring and control of anaerobic
a broad set of transient conditions. The model that wastewater treatment plants: Software sensors and controllers for
was identi®ed during steady states proves to be e- an anaerobic digestor. Water Sci Technol 43(7): 175±182.
cient in dynamical conditions, in particular during Chappell M, Godfrey KR 1992. Structural identi®ability of the pa-
the destabilization phases. rameters of a nonlinear batch reactor model. Math Biosci 108:241±
251.
This paper includes results of the project AMOCO, which is Chouakri N, Fonteix C, Marc I, Corriou J. 1994. Parameter estimation
supported by the Agriculture & Fisheries program (FAIR) of of a Monod-type model. Part I: Theoretical identi®ability and
the European Community (Contract ERB-FAIR-CT96- sensitivity analysis. Biotech Tech 8:683±688.
438 BIOTECHNOLOGY AND BIOENGINEERING, VOL. 75, NO. 4, NOVEMBER 20, 2001