State estimation for large-scale wastewater

treatment plants
Jan Busch
a,1
, David Elixmann
a
, Peter Ku hl
b,2
, Carine Gerkens
c
,
Johannes P. Schloder
b
, Hans G. Bock
b
, Wolfgang Marquardt
a,
*
a
AVT Process Systems Engineering, RWTH Aachen University, Germany
b
IWR, Ruprecht-Karls-Universita t Heidelberg, Germany
c
LASSC, Universite de Lie`ge, Belgium
a r t i c l e i n f o
Article history:
Received 27 June 2012
Received in revised form
2 April 2013
Accepted 5 April 2013
Available online 15 April 2013
Keywords:
State estimation
Moving horizon estimation
EKF
Wastewater treatment
ASM
BSM1
a b s t r a c t
Many relevant process states in wastewater treatment are not measurable, or their mea-
surements are subject to considerable uncertainty. This poses a serious problem for pro-
cess monitoring and control. Model-based state estimation can provide estimates of the
unknown states and increase the reliability of measurements. In this paper, an integrated
approach is presented for the optimization-based sensor network design and the estima-
tion problem. Using the ASM1 model in the reference scenario BSM1, a cost-optimal sensor
network is designed and the prominent estimators EKF and MHE are evaluated. Very good
estimation results for the system comprising 78 states are found requiring sensor networks
of only moderate complexity.
2013 Elsevier Ltd. All rights reserved.
1. Introduction
One of the challenges in applying activated sludge models to
the monitoring and control of activated sludge wastewater
treatment plants (WWTP) is the lack of available measure-
ment for many of the variables in these models. For example,
the concentrations of heterotrophic and autotrophic biomass
and of soluble substrate are not directly measurable, but they
considerably inuence process behavior. Some states such
as the concentrations of total suspended solids and soluble
nutrients (ammonia, nitrate) are measurable, but their mea-
surement by automatic sensors may involve signicant error
(Vanrolleghem and Lee, 2003). Reliable access to all these
states is of great value for different operational tasks such as
process monitoring, simulation-based decision support, and
advanced multi-variable control. They are a necessity for
model-based control approaches based on dynamic process
models (e.g., Stare et al., 2007). For a given process model,
the success of state estimation depends on the availability
of a suitable sensor network, of a sufciently accurate
process model, and of an appropriate estimation method
(Vanrolleghem, 2003; Brdys et al., 2008).
The intention of this paper is to present an integrated
approach to the state estimation problem for large-scale
* Corresponding author. Tel.: 49 241 80 94668; fax: 49 241 80 92326.
E-mail addresses: schloeder@iwr.uni-heidelberg.de (J.P. Schlo der), bock@iwr.uni-heidelberg.de (H.G. Bock), wolfgang.marquardt@avt.
rwth-aachen.de (W. Marquardt).
1
Present address: Bayer MaterialScience AG, Leverkusen, Germany.
2
Present address: BASF SE, Ludwigshafen, Germany.
Available online at www.sciencedirect.com
j ournal homepage: www. el sevi er. com/ l ocat e/ wat res
wa t e r r e s e a r c h 4 7 ( 2 0 1 3 ) 4 7 7 4 e4 7 8 7
0043-1354/$ e see front matter 2013 Elsevier Ltd. All rights reserved.
http://dx.doi.org/10.1016/j.watres.2013.04.007
WWTP and to investigate two distinct state estimation ap-
proaches from the practitioners point of view. First, an
optimization-based approach determines the cheapest sensor
network that is required for the state estimation task. Second,
Extended Kalman Filtering (EKF) and Moving Horizon Esti-
mation (MHE) are alternatively employed to estimate the un-
known model states of the large-scale WWTP model ASM1 as
suggested in the BSM1 reference scenario (Copp, 2002). Large
measurement errors, plant/model-mismatch, and unknown
feed concentrations are considered.
State estimation aims at statistically optimal predictions of
measurable and unmeasurable process states. Different state
estimation approaches include different kinds of observers
(Dochain, 2003), the well-known Kalman Filter (KF ) and its non-
linear extension, the Extended Kalman Filter (EKF ) (Becerra et al.,
2001), the Unscented Kalman Filter (UKF ) (Julier and Uhlmann,
2004), the Particle Filter (PF ) (Andrieu et al., 2004), and the fully
non-linear Moving Horizon Estimator (MHE ) (Rao et al., 2003).
Though it is beyond the scope of this paper to elaborate on all
of these approaches (Daum, 2005), their application to biolog-
ical degradation processes is sketched in the following. The
EKF and the MHE are discussed in more detail in Section 4.
An introduction to the design of observers and Kalman
lters, written from the point of view of a wastewater treat-
ment specialist, has been presented by Bernard et al. (2006).
Dochain (2003) provides a more general overview of (com-
bined) state and parameter estimation for chemical and
biochemical processes focusing on unknown inputs and
small-scale models. The application of state estimation to
wastewater treatment processes using activated sludge
models has rst been investigated by Ayesa et al. (1991) and
Larrea et al. (1992) who performed numerical experiments on
state and parameter estimation using an EKF. Oswald et al.
(1998) report the application of a Kalman Filter for the moni-
toring of a wastewater treatment plant using a reduced
version of the ASM1 model. In recent times, Kalman-based
estimators for wastewater applications have been presented
by Beltra n et al. (2009), who apply an EKF to estimate the solids
concentrations inside a pilot reactor using a model with four
states, and Chai et al. (2007), who compare a KF, EKF and UKF
for a small plant. Rutkowski and Brdys (2007) report the
application of a modied EKF to the model of a full-scale
wastewater treatment plant with over 100 states, based on
the ASM2d model (Henze et al., 1999). Elixmann et al. (2010)
apply the EKF as part of a model-predictive control algo-
rithm for a treatment plant model with 42 states. Successful
formulations of observers for wastewater treatment processes
have been presented by Lubenova et al. (2003), Lopez et al.
(2003), Goffaux and Vande Wouwer (2005), and Nagy Kiss
et al. (2011), who apply different observer types to estimate
the states of bioprocess models with up to ten state variables.
Ja uregui-Medina et al. (2009) design an observer to estimate
states and unknown inputs of an aerobic digestion reactor
using a reduced model with four unknown states.
MHE has seldomly been employed for WWTP, even though
its properties are well-suited for the problem (cf. Section 4).
Arnold and Dietze (2001) report an MHE implementation for
WWTP, but only preliminary results are given.
A few conclusions can be drawn from the cited literature.
Observers are straightforward to design and implement for
small-scale models with some 10 states. For larger models,
observer design becomes challenging. An exception are
asymptotic observers, which do not require the kinetic
models, but offer slow convergence. The EKF has often been
used for large-scale state estimation. It is easy to implement
and much experience is available concerning its design and
tuning. It is not clear yet whether the related estimators UKF
and PF can signicantly outperform the EKF in practical
implementations. The MHE is a promising option, but it is
neither clear whether its increased implementation effort is
justied by better estimation results in WWTP applications.
Large-scale simulation case studies for WWTP state esti-
mation are rare, and real-life case studies are e to the best of
the authors knowledge e not available. Also, while the
properties of the sensor network are decisive for the success
of any state estimation approach, this aspect has not been
treated much with respect to WWTP applications. Singh and
Hahn (2005) present an overview on methods for optimal
sensor network design, which have been developed in the last
decades. Rutkowski and Brdys (2007) discuss the sensor
problem for wastewater treatment, but they do not consider
the placement of sensors within the plant as a part of their
estimator design.
Ideally, the choice of a sensor network, of a process model,
and of the estimator should be considered as one integral
problem. This is beyond our possibilities today, but remains
the long-term research aim. In this contribution, we will
present a simple approach for the systematic design of a
minimal sensor network, which leads to an observable system
for a given large-scale plant model involving 78 differential
states. An EKF and an MHE are employed as state estimators.
Their performance is evaluated in face of considerable mea-
surement noise, plant/model-mismatch, and unknown feed
concentrations to simulate realistic conditions.
2. Plant layout, scenario, and process model
The process and the process model are based on the BSM1
benchmark (Copp, 2002), which has been developed as a
reference scenario for the evaluation and comparison of
different control approaches for WWTP. The plant layout is
depicted in Fig. 1. Q
0
and Z
0
are the feed rate and concentra-
tions. This stream is mixed with two recycle streams before
entering the plant. Two denitrication basins (each 1000 m
3
)
are followed by three aerated nitrication basins (each
1333 m
3
). The rst recycle, characterized by the ow rate Q
a
and the vector of concentrations Z
a
, is withdrawn from the
last nitrication basin. The settler used in BSM1 is replaced by
a membrane ltration unit, which is located in a separate
basin (250 m
3
without biological activity, to simulate a small
external tank) and which is modeled as an ideal splitter in this
Fig. 1 e Modied BSM1 plant layout.
wa t e r r e s e a r c h 4 7 ( 2 0 1 3 ) 4 7 7 4 e4 7 8 7 4775
work. The puried water leaves the plant with owrate Q
e
and
concentrations Z
e
. A second recycle and a waste stream are
dened by Q
r
and Q
w
as well as the vectors of concentrations
Z
r
and Z
w
, respectively. All basins are assumed to be well-
mixed.
The degradation processes in the ve biological basins are
described by the ASM1 model (Henze et al., 1987) with pa-
rameters taken from Copp (2002). The ASM1 model captures 8
reactions and the concentrations of inert soluble matter S
I
,
soluble substrate S
S
, inert particulate matter X
I
, particulate
substrate X
S
, heterotrophic biomass X
B,H
, autotrophic biomass
X
B,A
, particulate inert decay products X
P
, dissolved oxygen S
O
,
nitrate S
NO
, ammonia S
NH
, soluble organic nitrogen S
ND
, par-
ticulate organic nitrogen X
ND
, and the alkalinity S
ALK
. The
resulting model comprising mass balances and the kinetic
model contains 78 differential states representing the ASM1
concentrations and 240 algebraic variables representing re-
action rates and auxiliary variables. It is formulated as a semi-
explicit differential-algebraic model of differential index 1
given by
_ x fx; z; u; p; (1)
0 gx; z; u; p: (2)
x are the differential and z the algebraic states, u are the
manipulated variables, and p are the parameters. Addition-
ally, measurement equations are formulated according to
y C$x; (3)
where y are the measurable outputs and C is the measure-
ment matrix. This matrix can include a single unit entry per
row, selecting a particular state as measurement, or more
entries per row in case of lumped measurements such as the
chemical oxygen demand. The Eqs. (1) and (2) have to be
complemented by consistent initial conditions
_
xt
0

zt
0

_
x
0
z
0
_
: (4)
The BSM1 benchmark describes a dry weather scenario for a
period of 100 days with constant manipulated variables, feed
rates, and feed concentrations to reach a steady state. This is
followed by a period of 14 days with dynamic feed conditions.
One of the three different dynamic scenarios, the storm sce-
nario, is used in this paper. It is characterized by dry weather
feed superposed by two storm events. The corresponding feed
rate and ammonia feed concentration are depicted in Fig. 2.
The concentrations of dissolved oxygen (DO), autotrophic
biomass, decay products, and nitrate are zero, and the
alkalinity is 7. The remaining feed concentrations are depicted
in the results section (Fig. 7). The values of the manipulated
variables are constant and summarized in Table 1.
3. Sensor network design
The success of a state estimation approach depends on the
process model, on the state estimation algorithm, and on the
sensor network which supplies the measurements. In this
section, the design of the sensor network is discussed.
Optimal sensor network design aims at a sensor network
which leads to optimal state estimates at limited cost, or
similarly, reliable state estimates at minimumcost (Singh and
Hahn, 2005). So far, systematic approaches to this important
aspect of state estimation have been neglected in the litera-
ture on WWTP applications. Rather, the sensor network is
chosen based on experience and intuition.
However, systematic sensor network design should be
taken very seriously. To make the industrial implementation of
advanced monitoring and control approaches economically
attractive, they have to be realized at reasonable investment
costs for the measurement equipment. Optimal sensor
network design is a key methodology to fulll this requirement.
In the following, a simple hands-on approach is proposed
to quickly obtain an optimal sensor network for WWTP pro-
cesses. The sensor network is fully dened by the measure-
ment matrix C (Eq. (3)), which relates process states x to
measurements y. By assigning a price tag to the measurement
devices, a cost function
f fC (5)
is obtained, which describes the cost of the sensor network in
dependence of the sensor network dened by the measure-
ment matrix C. We want to nd the sensor network or
equivalently the measurement matrix C which gives mini-
mum cost f but still yields reliable state estimates. The latter
restriction is realized by demanding that the resulting system
needs to be observable.
3.1. Observability test
Observability is a system property which guarantees that,
given the inputs u and measurements y on a nite time ho-
rizon [t
0
,t], it is possible to estimate the state values x
0
at t
0
(and consequently, at any other point in time in [t
0
,t] predicted
by the process model). For linear, deterministic systems of the
form
0 5 10
0
2
4
6
x 10
4
time [d]
f
l
o
w

r
a
t
e

[
m
3
/
d
]
Q
0
0 5 10
0
20
40
time [d]
c
o
n
c
e
n
t
r
a
t
i
o
n

[
g
/
m
3
]
S
NH
Fig. 2 e Feed rate Q
0
and ammonia feed concentration Z
0;SNH
according to storm scenario suggested by Copp (2002).
Table 1 e Manipulated variables in the storm
scenario.
Variable Value
kl
a,basin 3
240 1/d
kl
a,basin 4
240 1/d
kl
a,basin 5
84 1/d
Q
r
18,446 m
3
/d
Q
a
55,338 m
3
/d
Q
w
385 m
3
/d
wa t e r r e s e a r c h 4 7 ( 2 0 1 3 ) 4 7 7 4 e4 7 8 7 4776
_ x A$x B$u; (6)
y C$x; (7)
observabilitydepends onthe matrices AandC. If it holds, thenit
is possible to design an observer that, given y, converges arbi-
trarily fast to the true values of x. There exist extensions of the
observability denition to nonlinear and stochastic systems.
However, nopracticallyusableobservability tests exist for large,
nonlinear systems such as ASM1. While this means that the
success of a state estimation approach for such systems can
only be guaranteed for very small or special cases, observability
tests can still be appliedto check if linear approximations of the
system are observable. This is neither a necessary nor a suf-
cient property for the success of a state estimation approach,
but it is a useful indicator (Chen et al., 2004).
We linearize the non-linear WWTP model at I points in
time during a typical transient and check these linearized
models for observability. This approximation gives at least an
indication of observability.
The linearized process models are denoted by
_ x A
i
$x B
i
$u ; x t
0
x
0
; (8)
y C$x ; (9)
where i 1,.,I is the index of the points in time where the
original model is linearized. It is omitted in the following for
brevity. The PopoveBelevicheHautus (PBH) rank test can now
be employed to check the observability of the systems given
by Eqs. (8) and (9) (Kailath et al., 2000).
Observability can be analyzed by means of the observ-
ability matrix
O
PBH;i

l
j
$I A
i
C
_ _
; (10)
where l
j
is the jth eigenvalue of A
i
and where I is the identity
matrix. Observability of the linear system(8)e(9) is guaranteed
if and only if O
PBH,i
has full column rank, i. e.
rank O
PBH;i
_ _
N cl
j
; j 1; .; N; (11)
where N is the number of differential states in Eq. (8). Eq. (11)
can equivalently be evaluated according to
s
min;i
min
l
j
s
min
O
PBH;i
l
j
_ _ _ _ _ _
> 0 ; j 1; .; N: (12)
s
min
is the minimum singular value of O
PBH,i
(l
j
), and s
min;i
is
the smallest s
min
for all j. Hence, all singular values of O
PBH,i
for
all eigenvalues of A
i
need to be positive.
The PBH test assumes that all input variables u (including
feed rate and feed concentrations) are known. In order to
consider unknown and time-varying inputs in the observ-
ability matrix, they can be introduced as additional differen-
tial states (Jazwinski, 1970). These states need to follow the
original feed concentration trajectories. Hence, the following
differential equations are introduced to the simulation model
(1)e(2):
x
_
in;i
Z
0;i
x
in;i
; i S
I
; S
S
; X
I
; X
S
; X
B;H
; S
ND
; X
ND
_ _
: (13)
x
in,i
are the newly introduced differential states, and Z
0,i
are
the original feed concentrations, where i is the index of the
components. In the estimation model (8)e(9), Eq. (13) cannot
be used because the values of Z
0,i
are unknown. Hence, the
following equations are introduced alternatively in the esti-
mation model for x
in,i
:
_ x
in;i
0; i S
I
; S
S
; X
I
; X
S
; X
B;H
; S
ND
; X
ND
_ _
: (14)
These simple equations, quite common in unknown input
estimation (Kurtz and Henson, 1998) reect the fact that the
feed concentrations are assumed to remain relatively con-
stant on the time scale of the update of the state estimator.
3.2. Design of the sensor network under observability
constraints
We now have a means to assess e in an approximate sense e
whether the WWTP model is observable for a given sensor
network. Hence, an optimization problem can be formulated
to nd the sensor network C which gives minimum cost, but
still yields observability with respect to the linearized model:
min
C
fC (15)
s:t: s
min;i
; i 1; .; I; (16)
where is a small positive error tolerance. The approach is
applied to the simulation scenario described in Section 2. It is
assumed that measurement devices are available for the
concentrations of DO S
O
, alkalinity S
ALK
, ammonia S
NH
, nitrate
S
NO
, chemical oxygen demand COD, ltered chemical oxygen
demand COD
f
, biological oxygendemand BOD, and suspended
solids X
TS
. Considering 8 possible measurements in 6 basins
gives a total of 2
6.8
w 2.8$10
14
measurement congurations,
from which the optimal one needs to be found. Table 2 lists
the denitions of all considered measurement devices and
their relative costs. The numbers chosen roughly reect ap-
proximations of the relative measurement equipment cost
rather than exact equipment prices.
It should be noted that some assumptions are made con-
cerning the measurement equations for BOD and X
TS
in Table
2: Z
XS
in the BOD measurement equation is assumed to be
completely particulate. In practice some part of Z
XS
may be
lterable, which would necessitate a modication of the
measurement equation. Furthermore, it is also assumed that
the measurement value registered by the BOD sensor is
identical with long-term BOD or that a relationship between
the measurement value and long-term BOD is known (from
Table 2 e Denition of measurement devices and their
relative cost.
Measurement Denition Rel. cost
S
O
Z
SO
10
S
ALK
Z
SALK
10
S
NH
Z
SNH
20
S
NO
Z
SNO
20
COD Z
SI
Z
SS
Z
XI
Z
XS
Z
XB;H
Z
XB;A
40
COD
f
Z
SI
Z
SS
40
BOD Z
SS
Z
XS
40
X
TS
Z
XI
Z
XS
Z
XB;H
Z
XB;A
Z
XP
Z
XND
40
wa t e r r e s e a r c h 4 7 ( 2 0 1 3 ) 4 7 7 4 e4 7 8 7 4777
laboratory analysis or otherwise). If no such relationship is
available, short-term BOD with BOD
st
Z
SS
would have to be
substituted. Lastly, the X
TS
measurement is assumed to
correspond to the concentration of suspended biomass. This
assumption would have to be validated by daily or weekly
laboratory measurements of the volatile suspended solids
(VSS) concentration. If the VSS concentration is found to be
lower than X
TS
(indicating that the sludge contains signicant
amounts of inorganic suspended solids), the measurement
equation would have to be modied by the addition of a
corrective term corresponding to a constant or slowly
changing inorganic solids concentration.
The non-linear process is simulated for the rst 10 days of
the storm scenario, rst with completely known inputs and
then with unknown feed rate and concentrations. Lineariza-
tion is performed after each day to obtain ten systems ac-
cording to Eqs. (8) and (9), respectively. A genetic algorithm is
employed to solve problem (15) (Goldberg, 1989; Gerkens and
Heyen, 2008), where the observability criterion needs to be
fullled for each of the ten linearized systems. The following
sensor network is found to give observability at minimum
cost:
With known feed rate and concentrations:
basin 1: COD, S
ALK
,
basin 2: S
O
,
basin 5: X
TS
.
With unknown feed rate and concentrations:
basin 1: X
TS
, S
ALK
,
basin 2: BOD,
basin 3: BOD, S
O
, S
ALK
,
basin 4: S
ALK
,
basin 5: COD,
membrane basin: COD.
The rst result is quite surprising, as it implies that only
four sensor devices sufce to estimate all 78 model states if
feed rate and concentrations are known. It is a theoretical
rather than practical result (the measurement of dissolved
oxygen in tank 2 may in practice be difcult due to the unfa-
vorable signal-to-noise ratio at the very low oxygen concen-
trations typically encountered in anoxic tanks) but it
illustrates that rigorous mathematical analysis can lead to
novel insights. Clearly, this is not an intuitive result which
could have been obtained by process insight or experience
alone.
The second sensor network for the estimation of plant
states and feed states is more complex, but technically and
economically feasible. The results indicate that, at least in
theory, measurement of oxygen, alkalinity, organic substrate
and total suspended solids sufce to estimate all suspended
and particulate concentrations in all tanks. Since this sensor
network does not include sensors for the nitrate or ammonia
concentrations, it can be hypothesized that the concentration
of the autotrophic biomass is inferred from the measured
alkalinity and substrate dynamics. However, establishing
clear relationships between specic sensors and individual
variables is not possible due to the large size of the model and
the high correlation between the different modes of the
observability matrix.
3
As mentioned in Section 3.1, observability is a useful in-
dicator, but it does not guarantee good convergence in the
presence of measurement and process noise. Therefore some
standard sensor devices, which are commonly available in
municipal and industrial WWTP, are added to the sensor
network. These are DOsensors in the aerated basins as well as
nitrate, ammonia, alkalinity, and COD measurements in the
efuent (membrane basin).
4. State estimators
State estimation generally refers to retrieving all states of a
dynamic system in real-time by utilizing available measure-
ments and a mathematical model of the system. For linear
systems, this task is largely solved and powerful tools such as
the KF and the optimization-based MHE exist. The rst be-
longs to the class of recursive estimators that only use the
most recent set of measurements to compute an estimate of
the current states. The latter works on a horizon or window
covering a limited number of past measurements. As the
process time evolves, this window moves on, hence the name
MHE. The state estimate is the result of a least-squares opti-
mization problem involving the measurements within the
window and a linear model of the process (Robertson et al.,
1996).
The state estimation problem becomes signicantly more
difcult for non-linear systems. Here, most methods are ex-
tensions of linear state estimators, such as the EKF as
described e.g., by Becerra et al. (2001) for the class of DAE
systems treated here. For the non-linear MHE, the underlying
dynamic optimization problem has to be solved for a non-
linear process model. Excellent overviews of the general
state-of-the-art in non-linear state estimation are given by
Daum (2005) and by Rawlings and Bakshi (2006).
Two non-linear state estimators e the EKF and the non-
linear MHE e are implemented and evaluated in the case
study presented in Section 5. In the following, the funda-
mental equations and implementation details are presented.
4.1. Extended Kalman Filter
The Extended Kalman Filter is a recursive method for state
estimation with non-linear process models. It consists of a
prediction step (time update) and a measurement update. Past
data are processed and propagated by means of a suitable
statistics. For the discrete time non-linear process model
x
k
f
k
x
k1
; u
k1
; p m
k
; (17)
y
k
C
d
$x
k
y
k
; (18)
with initial condition x
0
, measurement noise y
k
wN0;
~
V and
process noise m
k
wN0;
~
W where k denotes the sampling
instant and d emphasizes that the system is described in
3
The process model, the linearization matrices and a Matlab
program for the observability analysis are contained in the sup-
plementary data.
wa t e r r e s e a r c h 4 7 ( 2 0 1 3 ) 4 7 7 4 e4 7 8 7 4778
discrete time, the respective lter equations for an innite
horizon are:
x

k
f
k
x
k1
; u
k1
; p; (19a)
P

k

vf
k
vx
k1

x
k1
$P
k1
$
vf
k
vx
k1

T
x
k1
W; (19b)
K
k
P

k
$C
T
d
$C
d
$P

k
$C
T
d
V
1
; (19c)
P
k
I K
k
$C
d
$P

k
; (19d)
x
k
x

k
K
k
$
_
y
k
C
d
$ x

k
_
: (19e)
The matrix P
k
is the covariance matrix associated with the
state estimates x
k
at sampling time k. It reects the con-
dence one can have in this estimate. The matrices V and W
describe the assumed covariances of measurement and pro-
cess noise,
4
respectively. The initial conditions x
0
comple-
ment the state Eq. (19a).
As they indicate how much one can trust in the measure-
ments and the process model, respectively, they can be seen
as the tuning knobs of the Kalman Filter, though systematic
methods for their choice are only available for the linear case
(Odelson et al., 2006). The Kalman Filter gain K
k
then reects
the trade-off between the measurements and the process
model.
More implementational details can be found elsewhere
(Becerra et al., 2001), where the lter equations for the class of
differential-algebraic equations are thoroughly derived. In
case of the WWTP model, only the state equations had to be
linearized, but not the (linear) measurement equations.
4.2. Moving horizon estimation
A drawback of the EKF and most other estimation methods is
that they cannot deal with constraints on the estimated states
as they are often required, e.g., that concentration estimates
cannot be lower than zero. The moving horizon estimator
(MHE) that has been suggested rst for linear (Robertson et al.,
1996) and then for non-linear systems (Rao et al., 2003) over-
comes this problem. Another advantage is that the formula-
tion allows to additionally estimate process parameters or
unknown inputs without rst reformulating them as dummy
states. Since it is an optimization-based estimation approach,
it heavily depends on the underlying numerical optimization
schemes. This is particularly true for non-linear systems as
described by Eqs. (1) and (2).
Unlike the EKF, the MHE uses more than just the most
recent measurements. Instead, at a certain time t
j
a number of
M 1 measurements (y
j M
,.,y
j
) associated to the time in-
stants t
j M
<.<t
j
in the past are explicitly used for estima-
tion. The period t
j
t
j M
is called the horizon length and one
denes L : j M. It is assumed that measurement and pro-
cess noise are normally distributed with zero mean and
covariance matrices V and W. Additionally it is assumed that
there are a-priori distributions for the states x(t
L
) and pa-
rameters p that is Gaussian with expectation value x
L
; p
L
and
covariance matrices P
L
x
and P
L
p
.
The state estimation problemto be solved at time t
k
egiven
the measurements y
j
for j L,L 1,.,k, the known input u(t)
for t[t
L
,t
k
] and given x
L
; p
L
and P
L
e has the following form:
min
x $ ;p
kx t
L
x
L
k
2
P
x
L
kpp
L
k
2
P
p
L

k
jL
ky
j
C$x t
j
_ _
k
2
V
_
_
_
_
(20a)
s:t: _ xt fxt; zt; ut; p; tt
L
; t
k
; (20b)
0 gxt; zt; ut; p; (20c)
x
min
xt x
max
; (20d)
p
min
p p
max
; (20e)
where the applied norm is dened as kxk
2
V
: x
T
V
1
x.
The name moving horizon estimation becomes evident from
this problem formulation: At each new sampling time t
k
, one
new measurement vector y
k
enters the set of measurements,
while the last one y
L
becomes y
L 1
and drops out of the
horizon.
The initial weight terms kxt
L
x
L
k
2
P
x
L
and kp p
L
k
2
P
p
L
are
necessary to aggregate information in the MHE problem prior
to the horizon beginning at time t
L
. These terms are often
referred to as arrival cost. From a theoretical point of
view it would be ideal to dene the arrival cost as

L1
kN
ky
k
C$xt
k
k
2
V
, which, however, can only be computed
exactly in case of linear models. A simple remedy is to at least
approximately calculate the arrival cost. Typically, this is
done in a recursive way by applying the extended Kalman
Filter approach to update x
L
and p
L
(Ungarala, 2009). In this
case, an MHE with M 0, i.e., with only the most recent
measurements, is equivalent to the EKF if no constraints are
active. The horizon length L and the weighting matrices V
1
,
P
1
form the tuning parameters of the MHE. While V
1
is the
inverse of the measurement covariance matrix, P can be
interpreted as the process noise covariance W extended by
additional entries for the additionally estimated parameters.
Acrucial point for a working MHE implementationis to use a
fast and reliable numerical scheme for the constrained non-
linear dynamic optimization problem (20). The implementa-
tion used in this work is built around the optimization package
MUSCOD-II (Leineweber et al., 2003), which is based on the
direct multiple shooting approach to reformulate the dynamic
problem into a non-linear program. The estimation algorithm
employs so-called real-time iterations as described elsewhere
in detail (Diehl et al., 2002; Zavala et al., 2008). Integration of the
system equations has been done with the DAE solver DAESOL
(Bauer et al., 1999), which also provides all required sensitiv-
ities, both for the EKF and the MHE optimization problem.
5. State estimation results
The EKF and the MHE are used as state estimators for the
process model and scenario described in Section 2 and the
4
The covariances
~
V and
~
W that actually drive the noise pro-
cesses are, of course, unknown to the user.
wa t e r r e s e a r c h 4 7 ( 2 0 1 3 ) 4 7 7 4 e4 7 8 7 4779
sensor network calculated in Section 3. The difculty of the
estimation task is successively increased to evaluate the
estimation performance under realistic conditions.
Intentionally, only little effort has been devoted to the ne-
tuning of the estimators, since the aim of the study is not to
report on the best possible performance for a specic scenario
but to investigate the applicability and general performance of
the two estimation methods. In all simulation runs, the tuning
matrices V of the EKF and the MHE are diagonal matrices,
where the entries on the main diagonal have been chosen
pragmatically as V
jj
0:05$y
0;j

2
. j is the output index, and
y
0,j
y
j
(t 0). This choice corresponds to an expected white
measurement noise with a standard deviation of 5% of the
initial value. (For some sensors, 5% may be an optimistic es-
timate of the measurement standard deviation. Under real-
istic conditions, where standard deviations are known, these
values should be used instead.)
Analogously, the main diagonal of the tuning matrix W
describing the expected process noise is given by
W
ii
0:05$x
0;i

2
, where i is the state index. The initial guess P
0
for P is equal to W. The MHE uses an estimation horizon of
L 5 measurement samples. As mentioned in Section 4, this
parameter can be subject to further ne-tuning. Typically, a
larger horizon yields smoother estimates while often missing
fast changes in estimated process parameters. The MHE
constrains the estimated states and unknown inputs to values
larger or equal to zero, as all of them relate to component
concentrations.
The initial guess for the state values x
0
is set to 1:3$x
0
to
introduce an initial offset, thus imitating the poor insight one
typically has into the true initial process state. White mea-
surement noise y with a standard deviation of 5% of the initial
measurements is introduced in all experiments according to
y
k
C
d
$x
k
y
k
: (21)
The sampling interval for the measurements and the esti-
mation procedure is uniformly set to 15 min. In the following
presentation of the results, the state estimates in the third
basin are shown exemplarily, as there are no measurements
in this basin except for the DO concentration, and the esti-
mation performance is the worst compared to the remaining
basins. The DO concentration is not visualized as the esti-
mates always closely follow the true values. The states shown
are soluble inert matter S
I
, soluble substrate S
S
, particulate
inert matter X
I
, particulate substrate X
S
, heterotrophic
biomass X
B,H
, autotrophic biomass X
B,A
, inert decay products
0 5 10
0
20
40
S
I
0 5 10
0
5
S
S
0 5 10
1000
2000
3000
X
I
0 5 10
0
200
X
S
0 5 10
2000
4000
X
B,H
0 5 10
0
100
200
X
B,A
0 5 10
0
500
1000
X
P
0 5 10
0
5
10
S
NO
0 5 10
0
20
S
NH
0 5 10
0
0.5
1
1.5
S
ND
0 5 10
0
10
20
X
ND
0 5 10
4
6
8
S
ALK
Fig. 3 e EKF estimates, third basin, with known feed concentrations. The x-axis shows the time in days, and the y-axis
shows the concentrations in [g/m
3
]. The unit of S
ALK
is [mol/m
3
]. The graphs show the true process state values (light gray)
and their estimates (black).
wa t e r r e s e a r c h 4 7 ( 2 0 1 3 ) 4 7 7 4 e4 7 8 7 4780
X
P
, nitrate S
NO
, ammonia S
NH
, soluble organic nitrogen S
ND
,
particulate organic nitrogen X
ND
, and alkalinity S
ALK
.
The averaged residual error (ARE) of all J samplings is used
in the following as a measure to compare the overall estima-
tion performance. It is dened as
ARE
1
J

J
j1

N
i1
_
x
i;j
x
i;j
x
i;j
_
2

_
; (22)
where J is the number of samples x
i,j
and x
i;j
, i 1,.,N. The
ARE values for the different simulation case studies, including
a nominal scenario with known feed concentrations and no
process noise, are stated in Table 3.
5.1. Known feed concentrations
In the rst scenario, full knowledge of the feed rate and con-
centrations is assumed for reference. In order to simulate a
mismatch between the plant and the model, random noise is
added to the differential equations of the model. The process
noise m has zero mean and a standard deviation of 5% of the
initial states and enters the state Eq. (1) of the continuous
simulation model (1)e(3) according to
_ x fx; z; u; p m: (23)
Note that any negative concentrations resulting from the
addition of process noise are set to zero to ensure that the
equations remain physically feasible. Fig. 3 shows the results
for the third basin using the EKF. The estimated state values
(black lines) quickly converge from their initial offset to the
true values (gray lines). Only the inert concentrations X
P
, X
I
and S
I
show some occasional offset, which may be due to the
difculty in estimating these compounds which do not
participate in the biodegradation processes. For the EKF with
known inputs, an ARE of 0.7 is obtained.
Fig. 4 depicts the corresponding results for the MHE. Good
performance is obtained, althoughthe ARE is 0.8 and therefore
slightly higher than for the EKF (Table 3). Comparing the re-
sults for the third basin, the MHE shows higher oscillations
than the EKF of the concentration estimates for inert matter X
I
and heterotrophic biomass X
B,H
(Figs. 3 and 4).
5.2. Unknown feed concentrations
Up to now it has been assumed that the feed rate and con-
centrations are perfectly known. This assumption is not
0 5 10
0
20
40
S
I
0 5 10
0
5
S
S
0 5 10
1000
2000
3000
X
I
0 5 10
0
200
X
S
0 5 10
2000
4000
X
B,H
0 5 10
0
100
200
X
B,A
0 5 10
0
500
1000
X
P
0 5 10
0
5
10
S
NO
0 5 10
0
20
S
NH
0 5 10
0
0.5
1
1.5
S
ND
0 5 10
0
10
20
X
ND
0 5 10
4
6
8
S
ALK
Fig. 4 e MHE estimates, third basin, with known feed concentrations. The x-axis shows the time in days, and the y-axis
shows the concentrations in [g/m
3
]. The unit of S
ALK
is [mol/m
3
]. The graphs show the true process state values (light gray)
and their estimates (black).
wa t e r r e s e a r c h 4 7 ( 2 0 1 3 ) 4 7 7 4 e4 7 8 7 4781
realistic. While the feed rate is indeed well-known, at least
part of the feed concentrations are not. Typically historic data
are employed to obtain daily, weekly, and yearly trends and
patterns of, e.g., the concentrations or the composition of the
COD. However, a substantial bias between these predictions
and the real feed concentrations must be expected. To
acknowledge this aspect, a worst case situation is considered
in the following. The feed concentrations of soluble inert
matter S
I
, soluble substrate S
S
, particulate inert matter X
I
,
particulate substrate X
S
, heterotrophic biomass X
B,H
, and of
soluble S
ND
and particulate organic nitrogen X
ND
are treated as
unknown model inputs. The feed concentrations of DO S
O
,
autotrophic biomass X
B,A
, decay products X
P
, and nitrate X
NO
are set to zero, and the alkalinity S
ALK
is set to 7, which cor-
responds to typical feed characteristics. The feed rate Q
0
and
the feed ammonia concentration S
NH
are measurable. The
unknown inputs need to be estimated together with the un-
known states.
The sensor network applied in the following has been
determined in Section 3. The EKF nowestimates the appended
state vector x
T
; x
T
in

T
. The covariance matrix W is adapted to
the new state vector. The expected process noise standard
deviation of the unknown inputs is specied as 5% of their
nominal values, in analogy to the model state process noise.
The initial guess for the feed concentrations is also disturbed
by 30%. Fig. 5 depicts the estimated states for the third basin.
The estimation performance is again satisfactorily except for
two states. The estimationof inert particulate matter X
I
shows
considerable offset from the true values. This is, however, not
severe, as X
I
(like X
P
) is inert matter which does not affect the
reaction rates and is therefore irrelevant for process predic-
tion. The second state to exhibit a signicant offset is the
concentration of heterotrophic biomass X
B,H
. This is more
serious as heterotrophic biomass is responsible for the
degradation of substrate and nitrate. The offset in these two
biomass concentrations may be correlated due to the difculty
in differentiating between active and inert biomass using only
X
TS
and the concentrations of organic substrate (BOD and
COD) as measurements.
The estimation is qualitatively correct and certainly good
enough for monitoring purposes. Whether the offset is critical
e.g., in model-based control approaches needs to be evaluated
in future research. Fine-tuning of the estimator might further
minimize the deviation. The addition of nitrate sensors in the
anoxic zone could improve the accuracy of X
B,H
by enabling
a better estimation of biomass activity, since nitrate
0 5 10
0
20
40
S
I
0 5 10
0
2
4
6
S
S
0 5 10
1000
2000
3000
X
I
0 5 10
0
100
200
300
X
S
0 5 10
2000
3000
4000
5000
X
B,H
0 5 10
0
100
200
X
B,A
0 5 10
0
500
1000
X
P
0 5 10
0
5
10
S
NO
0 5 10
0
10
20
30
S
NH
0 5 10
0
0.5
1
1.5
S
ND
0 5 10
0
10
20
X
ND
0 5 10
4
6
8
S
ALK
Fig. 5 e EKF estimates, third basin, with unknown feed concentrations. The x-axis shows the time in days, and the y-axis
shows the concentrations in [g/m
3
]. The unit of S
ALK
is [mol/m
3
]. The graphs show the true process state values (light gray)
and their estimates (black).
wa t e r r e s e a r c h 4 7 ( 2 0 1 3 ) 4 7 7 4 e4 7 8 7 4782
consumption under anoxic conditions is an indicator of het-
erotrophic biomass activity. The overall ARE is 1.8 for the
states (Table 3) and 2.0 for the inputs.
Fig. 6 shows the estimation results for the states in the
third basin as obtained by the MHE. The results do not differ
much from those of the EKF. Again, the two states inert par-
ticulate matter X
I
and heterotrophic biomass X
B,H
show the
largest deviations. From visual inspection, the rst seems to
stay closer to the true value but then exhibits a sudden and
sharp drop which is not present in the real trend. The overall
ARE for this case is 1.6 (Table 3) and hence slightly better than
for the EKF. The estimated inputs achieve an ARE of 2.1.
The estimated feed concentrations are depicted in Fig. 7.
All estimates exhibit high-frequent oscillations, for the MHE
even more signicantly than for the EKF, which could prob-
ably also be improved by ne-tuning of the estimators. The
estimation of the concentration of heterotrophic biomass X
B,H
again shows a stronger offset during days 11e13 but returns to
the true value. The graphs of inert particulate matter X
I
and
particulate organic nitrogen X
ND
show that the estimates are
not able to follow sudden concentration peaks (day 9).
The main trends in the feed data are captured well, but it is
not clear especially with respect to the concentrations of inert
particulate matter X
I
and heterotrophic biomass X
B,H
whether
these inputs are actually observable. To clarify the issue, the
scenario is calculated again, but this time with a large jump in
these two feed concentrations after about 6 days (Fig. 8). The
results show that the estimated concentrations follow the
main trend, however, again they are not able to follow the
faster ups and downs with periods in the order of hours. Our
conclusion is that these unknown inputs are indeed (weakly)
observable e but they inuence the process only by very slow
dynamics (dilution in the tanks) that are not inuenced by
short perturbations.
Table 3 e ARE for the estimated states of different
simulation scenarios and estimators.
Estimator Known feed
conc., no
process noise
Known feed
conc., process
noise
Unknown feed
conc., process
noise
EKF 0.3 0.7 1.8
MHE 0.4 0.8 1.6
0 5 10
0
20
40
S
I
0 5 10
0
5
S
S
0 5 10
1000
2000
3000
X
I
0 5 10
0
200
X
S
0 5 10
2000
4000
X
B,H
0 5 10
0
100
200
X
B,A
0 5 10
0
500
1000
X
P
0 5 10
0
5
10
S
NO
0 5 10
0
20
S
NH
0 5 10
0
0.5
1
1.5
S
ND
0 5 10
0
10
20
X
ND
0 5 10
4
6
8
S
ALK
Fig. 6 e MHE estimates, third basin, with unknown feed concentrations. The x-axis shows the time in days, and the y-axis
shows the concentrations in [g/m
3
]. The unit of S
ALK
is [mol/m
3
]. The graphs show the true process state values (light gray)
and their estimates (black).
wa t e r r e s e a r c h 4 7 ( 2 0 1 3 ) 4 7 7 4 e4 7 8 7 4783
0 5 10
150
300
X
I
0 5 10
45
90
X
B,H
0 5 10
5
10
S
ND
0 5 10
15
30
X
ND
0 5 10
50
100
S
S
0 5 10
200
400
X
S
0 5 10
150
300
X
I
0 5 10
45
90
X
B,H
0 5 10
5
10
S
ND
0 5 10
15
30
X
ND
0 5 10
50
100
150
S
S
0 5 10
200
400
X
S
EKF
MHE
Fig. 7 e EKF estimates (upper half) and MHE estimates (lower half) of unknown feed concentrations. The x-axis shows the
time in days, and the y-axis shows the feed concentrations in [g/m
3
]. The graphs show the true input concentrations (light
gray) and their estimates (black). Note that the estimate of S
I
is omitted here, because S
I
is constant throughout most of the
simulation.
0 5 10
1500
3000
X
I
0 5 10
450
900
X
B,H
0 5 10
5
10
S
ND
0 5 10
15
30
X
ND
0 5 10
50
100
S
S
0 5 10
200
400
X
S
Fig. 8 e MHE estimates of unknown feed concentrations, with large jumps in the feed concentrations of X
I
and X
B,H
. The x-
axis shows the time in days, and the y-axis shows the feed concentrations in [g/m
3
]. The graphs show the true feed
concentrations (light gray) and their estimates (black).
wa t e r r e s e a r c h 4 7 ( 2 0 1 3 ) 4 7 7 4 e4 7 8 7 4784
The results of all estimation scenarios and the model
implementation are available as data les in the supplemen-
tary data.
5.3. Computational times
A general belief that can often be found is that optimization-
based estimation methods such as MHE are impractical
because of the excessive computation times to be expected.
Indeed the time required to solve a constrained optimization
problem to full convergence will necessarily be larger than
solving the same but unconstrained problem in a recursive
fashion. The numerical approach sketched in Section 4 can
signicantly reduce the computational times. Table 4 lists the
maximum and average CPU times measured for the different
simulation scenarios. All computations were performed on
the same personal computer, a Pentium 4 machine with
2.8 GHz, 1024 kB L2 cache, 1 GB RAM under Suse Linux 9.3.
Any similar numerical operations, e.g., linearizations, were
executed with the same algorithms. The maximum CPU
times must be interpreted with care because the CPU times
were measured without taking internal system tasks into
account. The computational times for the EKF are clearly
smaller than those needed to solve the MHE problems.
However, despite solving a nonlinear constrained dynamic
optimization problem, CPU times are in the range of a few
seconds only. This is by far sufciently fast for the state and
unknown inputs estimation task for waste-water treatment
plants.
5.4. On measurement sampling
A comment should be made on the issue of measurement
frequencies and measurement delays. In the simulations, all
measurements were assumed to be available with the same
frequency, which is certainly not the case for all of them in
reality, since delays of up to 30 min are described in BSM1 for
some automatic sensor devices and laboratory assays may
involve delays of up to 2e3 h. Sensor fusion or multi-rate
estimation methods would have to be applied to allow the
use of measurements with multiple frequencies and delays
(cf. Gopalakrishnan et al., 2011).
However, such an extension would not affect the pre-
sented observability analysis: Observability would still hold
on time points where measurements from all sensor are
available. The actual performance of the state estimation,
i.e., the speed of convergence of estimated states toward
their true values, would depend on how much the mea-
surements between those time points contribute to the
convergence. The slowest measurement frequency suggested
in BSM1 (2 h
1
for automatic BOD, COD, TOC and total-N
sensors) is half the frequency used in this study for all
sensors (4 h
1
), including those that would actually be avail-
able much more frequently. So a worst-case estimate of the
speed of convergence would be half of that shown in Figs. 4e8,
and potentially better if a multi-rate estimation scheme ex-
ploits those measurements that are available at higher fre-
quencies (oxygen and pH sensors as well as ion-sensitive
sensors).
6. Conclusions
In this paper an integrated approach to the state estimation
problem for large-scale wastewater treatment plants is pre-
sented. The process is based on the reference scenario BSM1
and employs the dynamic, non-linear process model ASM1.
The approach includes the optimization-based design of the
sensor network, which yields an observable system at lowest
cost. Based on this result, the two prominent state estimators
EKF and MHE are evaluated in simulation case studies. Large
process and measurement disturbances as well as unknown
inuent conditions have been investigated.
The results show that it is possible to obtain a fully
observable system with only few additional sensors. The un-
conventional design of the sensor network underlines the
necessity for the optimization-based design approach. The
results are theoretical in nature, but provide a rigorous
framework that should be considered in future applications of
state estimation to wastewater treatment.
Both the EKF and the MHE display good estimation per-
formance even under difcult conditions. The EKF shows a
marginally better performance for the scenarios with known
feed concentrations. In the case of unknown feed concentra-
tions, the MHE delivered slightly better state estimates. These
do not fully justify the higher implementation effort for the
MHE. However, its simple and straightforward handling of
unknown feed conditions and parameters is an advantage
over the EKF.
Future research directions are more detailed models of
wastewater sensors, the tuning of the estimators and their
integration inmodel-based control strategies for municipal and
industrial wastewater treatment plants (Elixmann et al., 2010).
Acknowledgments
We thank the German research foundation (DFG) for the
nancial support in the projects Optimization-based process
control of chemical processes (grant MA 1188/27-1) and Nu-
merical Methods for optimization-based control: Coupling of
online estimationand robust optimization (grant BO864/13-1).
Table 4 e CPU times (in seconds) for the different simulation scenarios and estimators.
Known feed conc., no process noise Known feed conc., process noise Unknown feed conc., process noise
Estimator Mean Max Mean Max Mean Max
EKF 0.52 2.34 0.47 1.21 0.88 3.47
MHE 2.29 4.45 2.29 9.43 2.34 5.07
wa t e r r e s e a r c h 4 7 ( 2 0 1 3 ) 4 7 7 4 e4 7 8 7 4785
Appendix A. Supplementary data
Supplementary data related to this article can be found at
http://dx.doi.org/10.1016/j.watres.2013.04.007.
r e f e r e n c e s
Andrieu, C., Doucet, A., Singh, S.S., Tadc, V.B., 2004. Particle
methods for change detection, system identication, and
control. Proceedings of the IEEE 92 (3), 423e438.
Arnold, E., Dietze, S., 2001. Nonlinear moving horizon state
estimation of an activated sludge model. In: Filip, F.,
Dumitrache, I., Iliescu, S. (Eds.), 9th IFAC Symposium on Large
Scale Systems: Theory and Applications, pp. 554e559.
Bucharest, Romania.
Ayesa, E., Florez, J., Garciaheras, J.L., Larrea, L., 1991. State and
coefcients estimation for the activated-sludge process using
a modied Kalman lter algorithm. Water Science Technology
24 (6), 235e247.
Bauer, I., Bock, H.G., Schlo der, J.P., 1999. DAESOL e a BDF-code for
the Numerical Solution of Differential Algebraic Equations.
Internal report, IWR, SFB 359. Universita t Heidelberg.
Becerra, V.M., Roberts, P.D., Grifths, G.W., 2001. Applying the
extended Kalman Filter to systems described by nonlinear
differential-algebraic equations. Control Engineering Practice
9, 267e281.
Beltra n, S., Irizar, I., Monclu s, H., Rodrguez-Roda, I., Ayesa, E.,
2009. On-line estimation of suspended solids in biological
reactors of WWTPs using a Kalman observer. Water Science
and Technology 60 (3), 567e574.
Bernard, O., Chachuat, B., Steyer, J.-P., 2006. Wastewater Quality
Monitoring and Treatment. John Wiley & Sons Ltd.,
pp. 247e263. Ch. State estimation for wastewater treatment
processes.
Brdys, M.A., Grochowski, M., Gminski, T., Konarczak, K.,
Drewa, M., 2008. Hierarchical predictive control of integrated
wastewater treatment systems. Control Engineering Practice
16 (6), 751e767.
Chai, Q., Furenes, B., Lie, B., 2007. Comparison of state estimation
techniques, appliedtoabiological wastewater treatment process.
In: Proceedings of the 10th IFAC Symposium on Computer
Applications in Biotechnology, pp. 353e358. Cancun, Mexico.
Chen, B.M., Lin, Z., Shamash, Y., 2004. Linear Systems Theory.
Birkha user, Boston, Basel, Berlin.
Copp, J.B. (Ed.), 2002. The COST Simulation Benchmark.
Description and Simulator Manual. Ofce for Ofcial
Publications of the European Communities, Luxembourg.
Daum, F., 2005. Nonlinear lters: beyond the Kalman lter. IEEE A
& E Systems Magazine 20 (8), 57e69.
Diehl, M., Findeisen, R., Schwarzkopf, S., Uslu, I., Allgo wer, F.,
Bock, H., Gilles, E., Schlo der, J., 2002. An efcient algorithm for
nonlinear model predictive control of large-scale systems.
Part I: description of the method. Automatisierungstechnik 50
(12), 557e567.
Dochain, D., 2003. State and parameter estimation in chemical
and biochemical processes: a tutorial. Journal of Process
Control 13, 801e818.
Elixmann, D., Busch, J., Marquardt, W., 2010. Integration of
model-predictive scheduling, dynamic real-time optimization
and output tracking for a wastewater treatment process. In:
Banga, J.R., et al. (Eds.), Proceedings of 11th IFAC Symposium
on Computer Applications in Biotechnology. (CAB2010).
Gerkens, C., Heyen, G., 2008. Sensor placement for fault detection
and localisation. In: Braunschweig, B., Joulia, X. (Eds.),
Proceedings of 18th European Symposium on Computer Aided
Process Engineering e ESCAPE 18. Elsevier B.V.
Goffaux, G., Vande Wouwer, A., 2005. Bioprocess state estimation:
some classical and less classical approaches. In: Meurer, T.
(Ed.), Control and Observer Design. Springer, Berlin,
Heidelberg, pp. 111e128.
Goldberg, D.E., 1989. Genetic Algorithms in Search, Optimisation
and Machine Learning. Addison-Wesley Publishing Company,
Reading, Massachusetts.
Gopalakrishnan, A., Kaisare, N.S., Narasimhan, S., 2011.
Incorporating delayed and infrequent measurements in
Extended Kalman Filter based nonlinear state estimation.
Journal of Process Control 21 (1), 119e129.
Henze, M., Grady Jr., C.P.L., Gujer, W., Marais, G.V.R., Matsuo, T.,
1987. A general model for single-sludge wastewater treatment
systems. Water Research 21 (5), 505e515.
Henze, M., Gujer, W., Mino, T., Matsuo, T., Wentzel, M.C., v. R.
Marais, G., Van Loosdrecht, M.C.M., 1999. Activated sludge
model No. 2d, ASM2D. Water Science and Technology 39 (1),
165e182.
Ja uregui-Medina, E.A., Alcaraz-Gonza lez, V., Me ndez-
Acosta, H.O., Gonza lez-A

lvarez, V., 2009. Observer-based

input estimation in continuous anaerobic wastewater
treatment processes. Water Science and Technology 60 (3),
805e812.
Jazwinski, A.H., 1970. Stochastic Processes and Filtering Theory.
Academic Press, New York.
Julier, S.J., Uhlmann, J.K., 2004. Unscented ltering and nonlinear
estimation. Proceedings of the IEEE 92 (3), 401e422.
Kailath, T., Sayed, A.H., Hassibi, B., 2000. Linear Estimation.
Prentice Hall, New Jersey.
Kurtz, M.J., Henson, M.A., 1998. State and disturbance estimation
for nonlinear systems afne in the unmeasured variables.
Computers & Chemical Engineering 22 (10), 1441e1459.
Larrea, L., Garciaheras, J.L., Ayesa, E., Florez, J., 1992. Designing
experiments to determine the coefcients of activated-sludge
models by identication algorithms. Water Science and
Technology 25 (6), 149e165.
Leineweber, D.B., Scha fer, A., Bock, H.G., Schlo der, J.P., 2003. An
efcient multiple shooting based reduced SQP strategy for
large-scale dynamic process optimization. Computers &
Chemical Engineering 27, 167e174.
Lopez, T., Pulis, A., Mulas, M.M., Baratti, R., 2003. A software
sensor for a wastewater treatment plant. In: Proceedings of
ADCHEM 2003, pp. 267e272.
Lubenova, V., Rocha, I., Ferreira, E.C., 2003. Estimation of multiple
biomass growth rates and biomass concentration in a class of
bioprocesses. Bioprocess and Biosystems Engineering 25,
395e406.
Nagy Kiss, A.M., Marx, B., Mourot, G., Schutz, G., Ragot, J., 2011.
Observers design for uncertain Takagi-Sugeno systems with
unmeasurable premise variables and unknown inputs.
Application to a wastewater treatment plant. Journal of
Process Control 21 (7), 1105e1114.
Odelson, B.J., Rajamani, M.R., Rawlings, J.B., 2006. A new
autocovariance least-squares method for estimating noise
covariances. Automatica 42 (2), 303e308.
Oswald, G., Mather, M., Spa th, W., Thomann, W., Gilles, E., 1998.
Model-based monitoring and automatic control of an
industrial biological wastewater treatment plant. at e
Automatisierungstechnik 46, 257e266 (in German).
Rao, C.V., Rawlings, J.B., Mayne, D.Q., 2003. Constrained state
estimation for nonlinear discrete-time systems: stability and
moving horizon approximations. IEEE Transactions on
Automatic Control 48 (2), 246e258.
Rawlings, J.B., Bakshi, B.R., 2006. Particle ltering and moving
horizon estimation. Computers & Chemical Engineering 30,
1529e1541.
wa t e r r e s e a r c h 4 7 ( 2 0 1 3 ) 4 7 7 4 e4 7 8 7 4786
Robertson, D.G., Lee, J.H., Rawlings, J.B., 1996. A moving horizon-
based approach for least-squares estimation. AIChE Journal 42
(8), 2209e2224.
Rutkowski, T., Brdys, M.A., 2007. The hybrid estimation algorithm
for wastewater treatment plant. In: 11th International IFAC
Symposium on Large Scale Complex Systems: Theory and
Applications. IFAC, pp. 1e6.
Singh, A.K., Hahn, J., 2005. Determining optimal sensor locations
for state and parameter estimation for stable nonlinear
systems. Industrial & Engineering Chemistry Research 44 (15),
5645e5659.
Stare, A., Vre cko, D., Hvala, N., Strm cnik, S., 2007. Comparison of
control strategies for nitrogen removal in an activated sludge
process in terms of operating costs: a simulation study. Water
Research 41, 2004e2014.
Ungarala, S., 2009. Computing arrival cost parameters in moving
horizon estimation using sampling based lters. Journal of
Process Control 19 (9), 1576e1588.
Vanrolleghem, P.A., 2003. Models in advanced wastewater
treatment plant control. In: Proceedings Colloque Automatique
et Agronomie, Montpellier 22e24 January 2003, p. 26.
Vanrolleghem, P.A., Lee, D.S., 2003. On-line monitoring
equipment for wastewater treatment processes: state of the
art. Water Science Technology 47 (2), 1e34.
Zavala, V., Laird, C., Biegler, L., 2008. A fast Moving Horizon
Estimation algorithm based on nonlinear programming
sensitivity. Journal of Process Control 18 (9), 876e884.
List of abbreviations
ARE: Average residual error
ASM: activated sludge model
BOD: biological oxygen demand
BSM: benchmark simulation model
COD: chemical oxygen demand
DAE: differential-algebraic equations
DO: dissolved oxygen
EKF: extended Kalman lter
KF: Kalman lter
MHE: moving horizon estimator
MLSS: mixed-liquor suspended solids (identical to TSS)
MLVSS: mixed-liquor volatile suspended solids
PBH test: PopoveBelevicheHautus rank test
PF: particle lter
TSS: total suspended solids (identical to MLSS)
UKF: unscented Kalman lter
WWTP: wastewater treatment plant
wa t e r r e s e a r c h 4 7 ( 2 0 1 3 ) 4 7 7 4 e4 7 8 7 4787