Parameter Estimation in Biokinetic Degradation Models in

Wastewater Treatment - A Novel Approach Relevant for Micro-

pollutant Removal

Monika Schnerklee*, Momtchil Peev*

* Austrian Research Centers GmbH - ARC, A 2444 Seibersdorf, Austria (email:

monika.schoenerklee@arcs.ac.at, momchil.peev@arcs.ac.at ), Tel.: +43 50 550 3450 (-3452 Fax)

Abstract

In this paper we address a general parameter estimation methodology for an extended biokinetic

degradation model [1] for poorly degradable micropollutants. In particular we concentrate on

parameter estimation of the micropollutant degradation sub-model by specialised microorganisms. In

this case we focus on the case when only substrate degradation data are available and prove the

structural identifiability of the model. Further we consider the problem of practical identifiability and

propose experimental and related numerical methods for unambiguous parameter estimation based

on multiple substrate degradation curves with different initial concentrations. Finally by means of

simulated pseudo-experiments we have found convincing indications that the proposed algorithm is

stable and yields appropriate parameter estimates even in unfavourable regimes.

Keywords: biokinetic model, micropollutants, parameter estimation, model identifiability

Nomenclature

Abbreviation
X biomass (g COD l-1)
S substrate (g COD l-1)
growth rate (d-1)
max maximum growth rate (d-1)
Ks saturation constant (g COD l-1)
Y growth yield
b biomass decay rate (d-1)
t time (d)

Indices
mp micropollutant
0 Initial condition at t=0

1/1
Introduction

Quantitative mathematical modelling of wastewater treatment (as e.g. the well known

IAWQ-ASM models) is a powerful tool allowing optimal design and operation of

municipal treatment systems. However, standard models operate with integral

system state-variables (such as COD) and do not enter into further details concerning

specific (low-concentration) organic compounds with particular environmental impact.

To achieve improved understanding and optimization of biological degradation of

specific micropollutants, quantitative modelling going beyond the integral approach is

required.

In this paper we address the parameter estimation for specific micropollutants to be

included in an extended biokinetic degradation model [1]. It builds upon standard

biokinetic model approaches, as e.g. ASM [2] and aims to describe the degradation

of low-concentration organic substances by specialised minority bacterial sub-

populations. To address such pollutants we concentrate on the extension of basic

biokinetics with respect to a single process type heterotrophic, aerobic

biodegradation alone.

A major challenge for application of such extended models is parameter estimation,

especially with respect to parameters related to the specialised biomass. The

problem is to indirectly estimate its growth and decay rates, as to our best knowledge

there are no easily available experimental methods for direct quantification of the

concentration of specialised, active biomass. In this situation it has first of all to be

clarified whether all parameters can be uniquely determined from substrate

concentration measurements alone. To this end structural and practical identifiability

analyses have to be performed.

2/2
State of the art and motivation

In a previous publication [1] we have addressed an extended biokinetic model

describing the simultaneous biodegradation of two types of substrate, one of them

being easily biodegradable (normal) and a second one consisting of low-

concentration micropollutants. The second substrate is assumed to be hardly

biodegradable by a hypothetic specialised microorganism population only, which in

reality can consist of many different species. This model contains three types of

parameters:

Parameters describing the normal biokinetic sub-system

Parameters describing the specialised biokinetic sub-system

Parameters describing the interactions of both sub-systems

This theoretical model extension has practical relevance only in case if the additional

biokinetic model parameters are identifiable and can be determined by feasible

experiments. The first type of parameters can easily be identified following standard

parameter estimation methodologies [3, 4, 5]. The determination of the interaction

parameters is currently an open question.

In this paper we focus specifically on determining the biokinetic parameters of the

specialised sub-system. To this end we address the following experimental situation:

Consider a biodegradation system where the full model described above is assumed

to be relevant. By extracting a sludge sample and isolating it from sources of easily

biodegradable (normal) substrate, the specialised biokinetic sub-system can be

considered as closed (see Equation 2 in [1]) and reduces to Equation (1). This

reduced biokinetic sub-model describes biodegradation of micropollutant substrate by

a specialised population only. It does not focus on particular substances, therefore in

3/3
the model equations we have changed the indices P3 (referring to persistent polar

pollutants) in [1] to micropollutant (mp) compounds.

dXmp
= mp Xmp bmp Xmp
dt (1)
dSmp mp
= Xmp
dt Ymp

Herebelow we restrict ourselves to standard Monod kinetics for the functional

dependence of the specific growth rate mp:

Smp
mp = max,mp (2)
K s,mp + Smp

We assume as usual that the kinetic parameters are constant and do not have a

complex functional form.

As mentioned above in the situation at hand, due to the difficulty of direct

measurements of active biomass Xmp, the parameter estimation methodology we

have developed relies on substrate concentration measurements alone. Note that

from an abstract point of view parameter estimation for biokinetic systems described

by Equation (1) has long been studied under the assumption that both biomass and

substrate concentrations (or other quantities which are functions of these variables)

can be experimentally assessed [6]. The major difference of the approach we

propose is based therefore on the assumption that only Smp measurements are

available. (It should be noted that in case of micropollutants standard respirometric

measurements are not feasible due to the extremely low substrate concentrations

and the respective immeasurably weak respirometric signals.) As a consequence the

methodology proposed here is not restricted to the case of micropollutants and can

always be applied whenever Equation (1) is valid and only substrate concentration

measurements are available.

4/4
There are two different questions to be answered [7]:

1. Is it possible at all to determine uniquely the biokinetic parameters even from

ideal (error-free) experimental data? And if yes which parameters? This

problem is known in literature as structural identifiability [8].

2. Is it possible to achieve unambiguous parameter estimation from real (noisy)

experimental data? This problem is usually called practical identifiability [9].

Structural Identifiability

In this section we omit the subscripts mp. This is done on the one hand for simplicity,

but also because our results have a general mathematical validity (not restricted to

micropollutants but applying to any biokinetic system described by equations of type

(1, 2)).

From a mathematical point of view we have to solve the following problem: Given the

(closed) system of equations (1, 2) and provided that the function S(t) is exactly

known for all times (t 0), is it in principle possible to uniquely determine the set of

biokinetic parameters max, Ks, b, Y and the unknown initial condition (biomass at t=0)

X(0)=X0?

As a first step we note that by formal integration of the biomass equation one can

transform the system of differential equations (1,2) into the following equivalent

integro-differential equation.

dS X0 S t
S
= max expmax d bt (3)
dt Y Ks + S 0
Ks + S

This equation has one initial condition S(0)=S0 and the other initial condition X(0)=X0

appears as a new parameter. Already at this point it is clear that it is only the ratio

5/5
X0/Y that can be structurally identifiable as any independent values of X0 and Y

yielding the same ratio have an identical impact on equation (3). (This can be

independently seen from equation (23) below.) Therefore the question is whether the

set of parameters max, Ks, b and X0/Y of equation (3) can be uniquely identified

provided that the function S(t) is known for all times.

To proceed further we take the logarithm of both sides of equation (3) multiplied by

minus one.

dS X0 S t
S
ln = ln(max ) + ln
+ ln
+ max d bt
Ks + S Ks + S
dt Y (4)
0

Please note that this operation is well defined as both the right and the left hand side

of equation (3) is a negative quantity for all t 0.

Let us now consider a fixed solution of equation (3) S(t), t 0, S(0)=S0. Then

equation (4) is an algebraic relation between the set of four parameters max, Ks, b

and X0/Y. Let us further assume that there exists a different set of parameters

~ ,K ~ ~ ~ ~
max s , b , X 0 /Y which satisfies the following relation

~
dS X0 ~ t S
= ln(max ) + ln ~
~ + ln ~ S + ~
K + S max K
ln d b t (5)
dt ~
Y s 0 s +S

for the same function S(t), t 0.

Subtracting the right hand sides of relations (4) and (5) we obtain the function (t)

which has to be identically equal to zero for all times t 0 due to the fact that the left

hand sides are identical.

X
0 ~
S( ) S( ) K s + S(t)
max Y
( )
t t
~ ~
+ bb (6)
(t) = ln ~ + ln ~ + max K + S() d max ~ d + ln

t 0, t 0
max X 0 0 s 0 K s + S( ) K s + S(t)
~
Y

6/6
From equation (6) it immediately follows that

~ ~
X 0 max X 0 K s + S 0
(0) = 0 ~ =~ (7)
Y max Y K s + S 0

~ ~
Thus we have expressed the parameter X 0 /Y as a function of the remaining ones.

Taking into account that the first derivative of equation (6) has also to be identically

equal to zero for all times t 0 and taking the limit t we find also that

) ( )
d S(t) ~ S(t) X(t) S(t) Ks Ks ~
+ b b 0, t 0
= max
dt K s + S(t)
max ~
K s + S(t)
max
Y K s + S(t) (K s + S(t)) K s + S(t)
~
(

(8)
d ~ ~
limt =bb b = b
dt

From the last result, the first derivative can be rewritten as follows:

d
(max ~max )S(t )3 + (max (K s + K~ s ) 2K s~max )S(t )2 + K s (maxK~ s ~maxK s )S(t ) max (K s K~ s )X(t )S(t ) (t )
= Y = 0
dt (K s + S(t ))2 (K~ s + S(t )) (t )
(9)
where (t ) = (max
~ )S(t )3 + K + K
max max s( (
~ ~ ) 2
) ~ ~
(
s 2K s max S(t ) + K s maxK s maxK s S(t )
max
Y
) ~
(
K s K s X(t )S(t ) )
2 ~
(
and (t ) = (K s + S(t )) K s + S(t ) )

The function (t) is positive at all times and therefore the function (t) has to be

identically equal to zero. However, this function is a polynomial in the variables S(t)

and X(t) (the span of which is a compact segment of the set of real positive numbers)

and it can be identically equal to zero, if and only if the coefficients of the different

powers are also identically equal to zero. This requirement automatically leads to

~ ~
KS = Ks and max = max (10)

These equalities are simultaneously necessary and sufficient conditions for the

function (t) to be identically equal to zero. Further from equations (10) and (7) it

follows that:

~
X0 X0
~ = (11)
Y Y

7/7
Equations (8), (10) and (11) demonstrate the unique structural identifiability of the

parameters max, Ks, b and X0/Y of equation (3).

Practical Identifiability

Practical identifiability, i.e. determination of kinetic parameters from experimental

data, is however a major challenge. In spite of the fact that it is possible to uniquely

determine these parameters from error-free single degradation curves, as

demonstrated above, this can hardly be done from experimental measurement data,

the problems being mathematically ultimately related to the well known high

sensitivity of biokinetic parameters encountered in standard microbial growth models

[10, 11]. Therefore the potentially feasible methods to determine the biokinetic

parameters would have to rely on additional information. One way to this end is to try

to employ the information contained not in one but in series of degradation curves,

for which it is on the one hand guaranteed that all relevant parameters Ks,mp, max,mp,

the ratio X0,mp/Ymp and bmp are identical while the initial values of the substrate

concentrations S0,mp are different but known.

In this approach to solve the problem of practical identifiability there are two implicit

tasks:

To find a constructive method for determining the biokinetic parameters from

ideal (error-free) experimental data of multiple degradation curves.

To find a modus of application of this method to the case when the available

experimental data are noisy and scarce.

The first of these two tasks can be solved explicitly. To this end we have developed

an experimental design and protocol which is feasible for the case of micropollutants

aiming at the estimation of the basic biokinetic parameters Ks,mp, max,mp, the ratio

X0,mp/Ymp and the decay rate bmp of the specialised biomass. We have designed two
8/8
basic types of short-term batch experiments to determine these quantities. These

should be thought of from two perspectives:

First the experiments are sort of ideal Gedankenexperimente where it is

assumed that they deliver sufficient data which are entirely free of noise. This

allows to obtain a precise approximation by means of curve fitting of the decay

and degradation curves.

Second, the experiments are also suggested as real experiments by means of

which the noisy and scarce experimental data are to be used to obtain

reasonable approximation as described in detail below.

Experimental design and constructive parameter estimation methodology in

the ideal case

1. Decay rate experiment - Determination of the decay rate of specialised biomass

bmp in mixed cultures

The endogenous decay rate bmp is obtained by monitoring the decay of active cell

mass in the absence of growth substrate. The major challenge in this case is finding

a suitable method for evaluating the quantity of active biomass. A method that can be

applied for mixed cultures, based on previous results by [12] can be described as

follows:

The decay experiment is performed as a series of subsequent batch experiments

using activated sludge in which effective biodegradation of the target micropollutant

substance occurs. Initially a sludge sample is left to decay without substrate. From

this decaying culture sub-samples are periodically withdrawn at time instances ti.

Each sub-sample is then spiked with a defined fixed concentration of micropollutant

target substrate and the initial slope of each degradation curve is estimated. The

procedure is based on the following relations:

9/9
 The initial slope of each of the resulting degradation curves ki (for t=ti) is

proportional to the concentration of active microorganisms Xmp(ti).

dSmp (S )
= mp mp Xmp (ti ) = k i (12)
dt Ymp

The specialised (active) biomass fraction Xmp(ti)/Xmp(t0) remaining at time ti for all i

can be computed by dividing the initial slopes ki for each sub-sample by the initial

slope k0 (t0 =0) at the beginning of the decay period.

b t
ki Xmp (t i ) X 0, mp e mp i b t
= = = e mp i (13)
k 0 Xmp (t 0 ) X 0, mp

In Equation (12) the endogenous biomass decay process is described as well

known by the function

bmp t i
X mp (t i ) = X 0, mp e (14)

which is a solution of Equation (1) in the case of absence of any substrate

(=mp=0).

The semilog plot of the specialised biomass fraction vs. time yields a straight line

with slope of -bmp.

Figure 1 shows the resulting idealised curve for the determination of the decay

rate bmp.

k0/k0=1 Semilog plot of active biomass

fraction vs. time
Active
biomass k1/k0
fraction
Xmp(t)/Xmp(t0) k2/k0
Slope -bmp
k3/k0

Time

Figure 1: Determination of the specialised biomass decay rate

10/10
2. Degradation experiments for determination of the biokinetic constants Ks,mp,

max,mp, X0,mp/Ymp and bmp

This experiment is performed as a series of parallel batch experiments using identical

activated sludge samples in which effective biodegradation of the target

micropollutant substance occurs. In each sample a chosen concentration of the

target micropollutant substrate is spiked, whereby the concentrations for the separate

samples are different and form a monotonous series. The basic procedure includes

the following steps:

Monitoring the degradation of selected micropollutant substrate, caused by

specialised biomass. For different initial substrate concentrations the resulting

degradation curves are recorded for m+1 different time instances (0, t1, , tm).

For n different initial substrate concentrations S1,mp (0 ),K, S n,mp (0 ) these curves

can be seen as solutions of the following differential equations:

dS1,mp max,mp S1,mp (t)

= X1,mp (t),
dt Ymp K s,mp + S1,mp (t)
K (15)
dS n,mp max,mp S n,mp (t)
= X n,mp (t).
dt Ymp K s,mp + S n,mp (t)

Determination of Ks,mp, max,mp and X0,mp/Ymp

For t=0 these differential equations can be seen as a set of algebraic equations of

the form:

dS1,mp max,mp S1,mp (0)

(0) = X1,mp (0)
dt Ymp K s,mp + S1,mp (0)
K (16)
dS n,mp max,mp S n,mp (0)
(0) = X n,mp (0)
dt Ymp K s,mp + S n,mp (0)

11/11
Note that due to the experiment design X1,mp(0)=.= Xn,mp(0). In what follows we

denote this parameter with X0,mp. Equations (16) are linear with respect to the two

unknown variables (x1=Ks,mp and x2=max,mp*X0,mp/Ymp).

dS1,mp
(0) S1,mp(0)
dS
- S1,mp(0) 1,mp (0)
dt dt
K K x1 = Ks,mp K
A xv = yr, A
= K v r
(17)
K , x = x = X 0,mp , y = K

Ymp
2 max,
mp
K K K
dSn,mp dSn,mp
(0) Sn,mp(0) - Sn,mp(0) (0)
dt dt

As it is well known [13], provided that the columns of are linearly independent

then exists

~v + r
x = A y, where (
A + = A T A )
1
A T is the pseudoinve rse matrix of A . (18)

The pseudoinverse provides a least squares solution to a system of linear

~v
equations. Thus x is equal to the least square solution of the system given in

equation (17).

Abstractly, the linear independence of the columns of is always guaranteed.

Indeed, as

dS1,mp S1,mp(0)
(0) S1,mp(0) S1,mp(0)
dt Ks,mp + S1,mp(0)
K K K K
= K X
A K = max.mp (19)
0,mp
K K
Ymp
K K K K
dSn,mp
(0) Sn,mp(0)
Sn,mp(0)
Sn,mp(0)
dt K + S (0)
s,mp n,mp

it is only in the regime when Ks,mp >> Si,mp(0) (i=1,,n) that has effectively rank

equal to one and taking the pseudoinverse is practically impossible.

12/12
 The determination of max,mp plus independent estimations of max,mp*X0,mp/Ymp and

bmp is based on equation (4) applied to all experimental data for all degradation

curves and is carried out as follows:

zv = vv
B (20)

t1
S1,mp dS1,mp S1,mp (t1 )
1 t1
K d ln
(t1 ) ln
+ S dt K s,mp + S1, mp ( t 1 )

L L
0 s,mp 1,mp
L
L X0,mp
t1
Sn,mp z = ln dS
ln n,mp (t ) ln S ( t ) (21)
1 t1 max,mp Y
0 K s,mp + Sn,mp v
n,mp 1
d 1

s,mp + Sn,mp (t1 )
mp
dt
1 K
B = , vv =
, z = z 2 = bmp
S1,mp dS1,mp S1,mp (t 2 )
t2
1 t
2 0 K s,mp + S1,mp d z3 = max,mp

ln
dt
(t 2 ) ln
K + S (t )


L L s,mp 1,mp 2
L L

Sn,mp dSn,mp Sn,mp (t m )
tm
1 t
0 K s,mp + Sn,mp d ln (t m ) ln
K s,mp + Sn,mp (t m )
m
dt

~v + v 1
z = B v, where B + = B T B B T ( ) is the pseudoinve rse matrix of B . (22)

Once again, the linear independence of the columns of B is abstractly always

given, but this time in the regime when Ks,mp << Si,mp(0) (i=1,,n) B has

effectively rank equal to two and taking the pseudoinverse is practically

impossible.

In summary, we have identified methods to constructively determine Ks,mp, max,mp,

X0,mp/Ymp and bmp from ideal decay and multiple degradation experiments. Two

quantities, namely bmp and max,mp*X0,mp/Ymp are independently determined in two

different ways. It should be stressed however that the second of these quantities, i.e.

max,mp,*X0,mp/Ymp is determined using two different algorithms from the same

experimental data. In contrast the decay rate is determined from two independent

experiments. Thus abstractly Experiment 1, above can be skipped. On the other

hand it increases the confidence of the results of Experiment 2.

13/13
Parameter estimation from realistic experimental data

To go to realistic (noisy) experiments, one has to be able to obtain good curve fits for

each of the degradation curves and at least the initial slope of the decay curves. A

straightforward way to do so is to use some standard curve-fit methodology.

Unfortunately for few experimental points and some degree of noise the result can be

strongly deviating if some standard family of interpolation functions as e.g. splines is

applied. A feasible approach can be defined as follows:

Take some reasonable, but imprecise approximation of the degradation curves by

some family of fitting functions which most appropriately have to monotonously

decrease as a function of time as do the degradation curves.

This allows the initial determination of the biokinetic parameters: first of all bmp,

then Ks,mp and the product max,mp*X0,mp/Ymp and finally max,mp plus an

independent estimation of bmp and max,mp*X0,mp/Ymp.

Now one can attempt a renewed approximation of the experimental data, this time

by numerical solution of the system of differential equations (1) and (2) by using

the biokinetic parameters estimated in the previous step and least square fitting.

Additionally this approach takes advantage of the fact that

dX mp dx mp
= mp X mp b mp X mp = mp x mp b mp x mp
dt dt

dS mp mp dS mp
= X mp = mp x mp (23)
dt Ymp dt
X 0,mp
S mp (0) = S 0,mp , X mp (0) = X 0,mp S mp (0) = S 0,mp , x mp (0) =
Ymp

This allows to find solutions for Smp although we cannot determine Ymp separately.

Finally one can once again determine the biokinetic parameters making use of

these new approximations of the degradation curves and iterate all algorithms

until stable results are achieved.

14/14
The above discussion does not yet allow estimating the confidence region of the

evaluated biokinetic parameters in case of noisy data and scarce experimental

points. While a full statistical analysis of the proposed method is still missing, below

we give an initial experimental estimate of these confidence regions by means of

simulating pseudo-experimental data.

Noise simulation

The pseudo-experimental data are obtained by taking calculated ideal values and

adding statistically independent noise. We have deliberately chosen to have a small

number of measurement points as in real experiments only a limited number of

analysed samples is usually available. This is due to a number of reasons: For low-

concentration micropollutants the analysis procedure is extremely complex, time-

consuming and costly. Additionally sample collection under identical experimental

conditions is prerequisite for sound parameter estimation. While this is still possible

for a small number of samples, it becomes increasingly difficult with larger sample

series.

In particular the pseudo-experimental data are obtained as follows:

We simulate equation (23) for the following parameter values: max,mp =1 d-1,

Ks,mp= 22 g/l, X0,mp/Ymp=330 g/l, bmp=0,3 d-1 for four fixed initial values of the

micropollutant substrate concentration at t=0.

We add normally distributed noise (four values of standard deviations stabw= 0,

0.02, 0.05, 0.1) for each pseudo-experimental value. Adding noise technically

means that we produce an array of normally distributed random numbers with a

mean value of 1 (the length of the array being equal to the length of the pseudo-

experimental series) and multiply in a point-by-point way the random number

array and the pseudo-experimental data array.

15/15
 For each value of standard deviation of the noise we produce the data for a

number of noisy pseudo-experiments as described above.

From each of these pseudo-experimental noisy curves we take out a sub-set

consisting of 7 experimental points at fixed time instances.

7 14
Simulated ideal degradation curve Simulated ideal degradation curve
Pseudo-experimental points Pseudo-experimental points
6 12
Degradation curve with added noise stabw=0.1 Degradation curve with added noise stabw=0.1

5 10
Substrate [g COD/l]

Substrate [g COD/l]
4 8

3 6

2 4

1 2

0 0
0 0.05 0.1 0.15 0.2 0.25 0 0.05 0.1 0.15 0.2 0.25
Time [days] Time [days]

30 60
Simulated ideal degradation curve Simulated ideal degradation curve
Pseudo-experimental points Pseudo-experimental points
25 Degradation curve with added noise stabw=0.1 50 Degradation curve with added noise stabw=0.1
Substrate [g COD/l]

Substrate [g COD/l]

20 40

15 30

10 20

5 10

0 0
0 0.05 0.1 0.15 0.2 0.25 0 0.05 0.1 0.15 0.2 0.25
Time [days] Time [days]

Figure 2: Plot of a set of simulated ideal degradation curves, these degradation curves with added

noise stabw=0.1 and experimental points at seven fixed time instances

One should note that these final pseudo-experimental series are really unfavourable

for parameter estimation. In addition to the mentioned scarcity of data points, the

noise we have introduced is proportional to the true pseudo-experimental values.

This needs not actually be the case in a real measurement error as the distribution

could be constrained in a fixed range and not necessarily proportional to the

experimental value. With the approach we have chosen we particularly increase the

noisiness of degradation curves with higher initial substrate concentrations and in

general the error in the initial part of every degradation curve. This affects

particularly these parts of our estimation algorithm which are based on initial slopes

and substrate concentrations.

16/16
For each of the pseudo-experimental series we apply the algorithmic approach for

parameter estimation as described above using 100 iteration steps. The results are

arrays of parameter estimates indexed by the iteration step of the algorithm. An

example of such type of results for standard deviation 0.02 is presented in Figure 3.

max,mp Calculation - stabw 0.02 Ks,mp Calculation - stabw 0.02

3,5 30

3
25

2,5
20

Ks,mp [g COD/l]
2
max,mp [d ]
-1

15
1,5

10
1

0,5 5

0 0
1 5 9 13 17 21 25 29 33 37 41 45 49 53 57 61 65 69 73 77 81 85 89 93 97 101 1 5 9 13 17 21 25 29 33 37 41 45 49 53 57 61 65 69 73 77 81 85 89 93 97 101
Number of iterations Number of iterations

max,mpX0,mp/Ymp Calculation - stabw 0.02 bmp Calculation - stabw 0.02

1,4
400

350 1,2

300 1
max,mpX0,mp/Ymp [g l-1 d-1]

250
0,8
bmp [d-1]

200
0,6
150

0,4
100

50 0,2

0 0
1 5 9 13 17 21 25 29 33 37 41 45 49 53 57 61 65 69 73 77 81 85 89 93 97 101 1 5 9 13 17 21 25 29 33 37 41 45 49 53 57 61 65 69 73 77 81 85 89 93 97 101
Number of iterations Number of iterations

Figure 3: Calculated values for max,mp, Ks,mp, max,mp *X0,mp/Ymp and bmp using noise standard deviation

stabw=0.02 for 22 sets of pseudo-experimental data using 100 iterations in each parameter

estimation

The results of all parameter estimations based on pseudo-experimental data are

given in Figure 4, whereby the straight line corresponds to the values of the

parameters used in the simulation before adding noise. (Note that these values

coincide with the algorithm output in the case of no noise standard deviation of

noise equal to zero). For each value of the input noise standard deviation a mean

value of all algorithm estimates and the corresponding output standard deviation

are given.

17/17
 Calculation of m ax,m p for different standard deviations Calculation of Ks,m p for different standard deviations
(means and standard deviation range) (means and standard deviation range)
3,5
30
3
25
2,5
max,mp [d ]
-1

20

Ks,mp [g/l]
2

1,5 15

1 10

0,5 5

0 0
0 0,02 0,04 0,06 0,08 0,1 0 0,02 0,04 0,06 0,08 0,1
Noise (stabw) Noise (stabw)

Calculation of m ax,m pX0,m p/Ym p for different standard deviations Calculation of bm p for different standard deviations
(means and standard deviation range) (means and standard deviation range)
400 0,8

350 0,7
max,mp X0,mp /Ymp [g l d ]
-1

0,6
300
-1

0,5
250
bmp [d ]
-1

0,4
200
0,3
150
0,2
100
0,1
50
0
0
0 0,02 0,04 0,06 0,08 0,1
0 0,02 0,04 0,06 0,08 0,1
Noise (stabw)
Noise (stabw)

Figure 4: Diagramme for max, mp, Ks,mp, max,mp *X0,mp/Ymp and bmp for different standard deviations

(mean plus confidence range) for 22 sets of pseudo-experimental data using 100 iterations in each

parameter estimation

This approach gives satisfactory indications of the standard deviation range of the

evaluated biokinetic parameters as a function of the degree of noisiness

characterised by e.g. the standard deviation of the applied noise function.

Simultaneously it should be noted that it cannot account for any modelling errors

i.e. possible deviations of the experimental data to the form of the model assumed.

Conclusions

In this paper we have addressed the structural and practical identifiability of

equations (1) and (2) in the case when only substrate measurement data are

available from a methodological point of view. We have proven the structural

18/18
identifiability of the problem and have developed a new approach allowing practical

identifiability on the basis of multiple substrate degradation curves with different initial

concentrations. While a full statistical sensitivity analysis is still missing, we have

shown by means of simulated pseudo-experimental noisy data that the proposed

parameter estimation algorithm is stable and gives appropriate parameter estimates

even in a strongly unfavourable regime chosen above.

The approach described above is the basis for parameter estimation for a number of

micropollutant substances investigated in activated sludge and MBR systems. In

particular, specific experiments were carried out in order to determine biokinetic

degradation parameters for NDSA and BTSA for which the presented methodology

was additionally tested and verified. The experimental results and the corresponding

parameter estimation will be presented in a subsequent publication.

Acknowledgement

The presented work was carried out within the framework of the project PTHREE

(www.pthree.de) financed under contract number EVK1-CT-2002-00116 by the

European Commission (FP5).

References

[1] M. Peev, M. Schnerklee, H. De Wever (2004). Modelling the degradation of low

concentration pollutants in membrane bioreactors. Wat. Sci. Tech. 50(5), 209218.

[2] M. Henze, C. P. L. J. Grady, W. Gujer, G. V. R. Marais, T. Matsuo (1987). Activated

sludge model no.1. Scientific and Technical Report 2, IWAQ, London.

[3] B. Petersen (2000). Calibration, identifiability and optimal experimental design of activated

sludge models. PhD Thesis. BIOMATH Department, Ghent University, Belgium.

19/19
[4] B. Petersen, K. Gernaey, M. Henze and P.A. Vanrolleghem (2003). Calibration of

activated sludge models: A critical review of experimental designs. In: Biotechnology for

the Environment: Wastewater Treatment and Modeling, Waste Gas Handling. Eds.

Agathos S.N. and Reineke W., Kluwer Academic Publishers, Dordrecht, The

Netherlands. 101-186.

[5] P.A. Vanrolleghem, G. Insel, B. Petersen, G. Sin, D. De Pauw, I. Nopens, H. Doverman,

S. Weijers and K. Gernaey (2003). A comprehensive model calibration procedure for

activated sludge models. In: Proceedings 76th Annual WEF Conference and Exposition.

Los Angeles, USA, October 11-15, 2003. (on CD-ROM).

[6] K. Gernaey, B. Petersen, D. Dochain, P.A. Vanrolleghem (2002). Modeling aerobic

carbon source degradation processes using titrimetric data and combined respirometric-

titrimetric data: structural and practical identifiability. Biotechnol Bioeng 2002;79:75467.

[7] D. Dochain and P. A. Vanrolleghem (2001). Dynamical Modeling and Estimation in

Wastewater Treatment Processes, IWA Publishing, London, UK.

[8] Dochain, D., Vanrolleghem, P.A., Van Daele, M. (1995). Structural identifiability of

biokinetic models of activated sludge respiration. Water Res 1995;29:25719.

[9] P.A. Vanrolleghem, M. V. Daele, D. Dochain (1995). Practical identifiability of a biokinetic

model of activated sludge respiration. Water Research, 29(11):25612570.

[10] A. Holmberg (1982). On the practical identifiability of microbial growth models

incorporating Michaelis-Menten type nonlinearities. Math. Biosci. 62 (1982), no.1, 23--43,

MathSciNet.

[11] B. Petersen, K. Gernaey, P.A. Vanrolleghem (2001). Practical identifiability of model

parameters by combined respirometric-titrimetric measurements. Water Science and

Technology Vol 43 No 7 pp 347355, 2001.

[12] C.S. Criddle (1993). The kinetics of co-metabolism. Biotechnology and Bioengineering

Vol. 41, Number 11, 1048-1056.

[13] R. Penrose (1956). On best approximate solution of linear matrix equations.

Proceedings of the Cambridge Philosophical Society 52, 17-19.

20/20