Vous êtes sur la page 1sur 10

w a t e r r e s e a r c h 4 6 ( 2 0 1 2 ) 2 1 0 1 e2 1 1 0

Available online at www.sciencedirect.com

journal homepage: www.elsevier.com/locate/watres

Optimal aeration control in a nitrifying activated sludge


process
L. Amand a,b,*, B. Carlsson b
a
b

IVL Swedish Environmental Research Institute, P.O. Box 210 60, 100 31 Stockholm, Sweden
Department of Information Technology, Uppsala University, P.O. Box 337, 751 05, Uppsala, Sweden

article info

abstract

Article history:

An important tool to minimise energy consumption in activated sludge processes is to

Received 3 November 2011

control the aeration system. Aeration is a costly process and the dissolved oxygen level will

Received in revised form

determine the efficiency of the operation as well as the treatment results. What aeration

18 January 2012

control should achieve is closely linked to how the effluent criteria are defined. This paper

Accepted 20 January 2012

explores how the aeration process should be controlled to meet the effluent discharge

Available online 28 January 2012

limits in an energy efficient manner in countries where the effluent nitrogen criterion is
defined as average values over long time frames, such as months or years. Simulations have

Keywords:

been performed using a simplified Benchmark Simulation Model No. 1 to investigate the

Activated sludge process

effect of different levels of suppressing the variations of the effluent ammonium concen-

Aeration control

tration. Optimisation is performed where the manipulated variable for aeration (the oxygen

Optimisation

transfer coefficient, KLa) is minimised with the constraint that the average daily flow-

Supervisory ammonium control

proportional ammonium concentration in the effluent should reach a desired level. The
optimisation results are compared with constant dissolved oxygen concentrations and
supervisory ammonium control with different controller settings. The results demonstrate
and explain how and why energy consumption can be optimised by tolerating the
ammonium concentration to vary around a given average value. In these simulations, the
optimal oxygen peak-to-peak amplitude range between 0.7 and 1.8 mg/l depending on the
influent variation and ammonium level in the effluent. These variations can be achieved
with a slow ammonium feedback controller. The air flow requirements can be reduced by
1e4% compared to constant dissolved oxygen set-points. Optimal control of aeration
requires up to 14% less energy than needed for fast feedback control of effluent ammonium.
2012 Elsevier Ltd. All rights reserved.

1.

Introduction

Since the introduction of more reliable dissolved oxygen (DO)


sensors in the 1970s (Olsson et al. 2005), research within
control of aeration in activated sludge processes has span
from classical control methods such as feedback
and feedforward control (Lindberg and Carlsson 1996;
Ingildsen et al., 2002; Vrecko et al., 2003) to advanced model

based and multivariable control (Steffens and Lant 1999; Shen


et al. 2009). It is well known that unnecessary high air flow
rates and oxygen concentrations should be avoided, due to the
decreased aeration efficiency and oxygen transfer this will
cause (Olsson et al. 2005; Thunberg et al., 2009).
The possibilities of full-scale control have been boosted by
the improvement of online sensor technology and the introduction of more advanced SCADA (Supervisory Control And

* Corresponding author: IVL Swedish Environmental Research Institute, P.O. Box 210 60, 100 31 Stockholm, Sweden. Tel.: 46 8 598 564 19;
fax: 46 8 598 563 90.
mand).
E-mail address: linda.amand@ivl.se (L. A
0043-1354/$ e see front matter 2012 Elsevier Ltd. All rights reserved.
doi:10.1016/j.watres.2012.01.023

2102

w a t e r r e s e a r c h 4 6 ( 2 0 1 2 ) 2 1 0 1 e2 1 1 0

Data Acquisition) systems at wastewater treatment plants


(WWTPs). On the other hand, development of the Activated
Sludge Models (ASM) (Henze et al., 2000) and the IWA/COST
Benchmark Models (BSM) (Copp 2001) have improved the
possibilities for a fair comparison between different control
strategies by simulations.
Wastewater treatment processes are often described as
complex with non-linear dynamics and strong interactions
within the multivariable system (Samuelsson et al., 2005).
Despite this, well performing decentralised controllers have
been demonstrated to minimise the coupling between variables and independently control specific unit processes within
the system (Machado et al., 2009; de Arau`jo et al., 2011). Using
decentralised controllers could be supported by the fact that
interactions within the process have a large range of time
constants (Olsson and Jeppsson 1994) thereby allowing for
separate control of these processes. For instance, Olsson and
Jeppsson (1994) report that sludge properties change over
a period of days while oxygen transfer occurs within minutes.
Earlier work investigating unit process controllers include
i.e. Vrecko et al. (2002). When controlling the suspended solids
concentration, the dissolved oxygen concentration and the
nitrate recirculation flow rate, the energy consumption is
lower but with higher effluent concentrations compared to
other published work using advanced ammonium feedback
and feedforward control. The tested controllers also experienced a lower fewer effluent violations. Stare et al. (2007)
compare constant set-points and supervisory feedback and
feedforward controllers with an ideal multivariable Model
Predictive Controller (MPC) for control of nitrogen removal,
concluding that feedback and feedforward control can
compete with MPC with respect to operating costs and treatment results.
Several authors have discussed the importance of control
goal formulation and control structure design in relation to
WWTP process control (Weijers 2000; Ingildsen 2002). Of
particular interest for control goal formulation are the regulations governing the compliance to the effluent criteria.
Authorities measure compliance to effluent criteria based on
long or short time frames. In countries such as Sweden and
Spain compliance of the nitrogen permit is assessed over long
time frames. In Sweden annual averages are often used. As
discussed by Ingildsen (2002), the regulations have an impact
on the type of control goal to strive for in the control of
aeration systems. A repetitive load variation entering the
plant can be considered either as a disturbance (which there is
a wish to suppress) or as a variation from the mean (tolerated
as long as the average concentration is below the effluent
limits). Having a control goal that allows the output signal to
vary around an average is also referred to as averaging control
strom and Hagglund, 1995).
(A
There have been earlier efforts to optimise the activated
sludge process which are related to aeration control. With
respect to dynamic optimisation of aeration it is common to
consider one single objective, as in this work. Earlier work has
been focused on optimising the length of the aerated and nonaerated periods in an intermittently aerated process (Fikar
et al., 2005; Balku and Berber, 2006; Holenda et al., 2007). As
pointed out by Beraud et al. (2009), the limitation of singleobjective optimisation is that the long-term effects on the

process and the trade-off between other operation objectives


are
not
taken
into
account.
Several
authors
have included optimisation in hierarchical control
strategies (Yamanaka et al., 2006; Brdys et al., 2008; Beraud
et al., 2009). In such set-ups, supervisory control layers
decide set-points for lower control layers, such as DO
controllers, based on results from optimisation of a process
model. Note also that MPC (Shen et al., 2009) is based on
a receding horizon optimisation.
This study investigates by simulations how aeration can be
controlled in an optimal manner using simple decentralised
controllers to control aeration in a nitrifying activated sludge
process. A discussion is provided about how fast a control
action in a feedback loop should react to the influent variations for different loads and influent profiles, given that the
permit for nitrogen is assessed over long time frames. The
purpose of this work is to provide an insight into how to set-up
the control to achieve minimum energy consumption while
still complying with the effluent criteria. The paper includes
a section about the process non-linear dynamics, a method
section
describing
the
simulation
model
and
performed simulations, and the result section is finally followed by a discussion of the results and conclusions.

2.
Non-linear aeration and growth rate
dynamics
The manipulated variable in a diffused aeration system is the
valve opening, affecting the air flow rate to the basin. There
are several steps taking place from the point of deciding the
air flow rate until the growth of aerobic bacteria takes place
with resulting nitrogen removal and removal of carbonaceous
material. The dynamics of the process is non-linear due to the
following reasons:
1. The air flow rate causes oxygen to transfer from gas phase
to liquid phase. The transfer efficiency is decided by the
oxygen transfer efficiency (OTE). The OTE is decreasing at
increasing air flow rates.
2. The oxygen transfer will affect the dissolved oxygen
concentration in the liquid. When the dissolved oxygen
concentration increase the oxygen transfer rate is reduced
since diffusion is slower closer to saturation.
3. The growth rate of aerobic bacteria is non-linear with
respect to dissolved oxygen concentration and ammonium
concentration.
In the list above, the last two non-linearities are covered in
the benchmark models. The first non-linear relationship arises
from the reduced increase of the oxygen transfer coefficient
(KLa) at high air flow rates. This is caused by higher rise times
and lower surface areas of the bubbles when they are more
prone to aggregate at higher air flow rates. In the benchmark
models, the KLa is used to model air flow rate directly.
The second relationship can be described by the standard
model for oxygen transfer under process conditions (eq. (1)).
dC
KL aCS  Ct  rM
dt

(1)

w a t e r r e s e a r c h 4 6 ( 2 0 1 2 ) 2 1 0 1 e2 1 1 0

KLa, Oxygen transfer coefficient (d1); Cs, DO saturation


concentration (g/m3); Ct, DO concentration at time t (g/m3); rM,
respiration by microorganisms (g/m3, d).
The final non-linearity is captured by Monod kinetics
(Monod, 1942). In the Monod function the growth rate is, with
respect to a certain substrate limiting for growth, a monotonically increasing non-linear function which eventually
approaches its maximum specific growth rate at high
substrate concentrations. In wastewater treatment modelling
and analysis, the Monod function is widely used to describe
the dynamics of the growth of bacteria in the activated sludge
process, and Monod-type expressions are reported to provide
a reasonable description of this behaviour (Jeppsson, 1996). In
ASM1 a combination of two Monod functions is used to
describe growth of autotrophs, shown in eq. (2) (Henze et al.,
2000).
dXB;A
dt


mmax;A

SNH
SO
KNH SNH KO;A SO


 bA XB;A

(2)

XB,A, concentration of autotrophic biomass (g/m3); mmax,A,


maximum specific growth rate for autotrophs (d1); SNH,
ammonium concentration (g/m3); KNH, half-saturation
constant for ammonium (g/m3); SO, dissolved oxygen
concentration (g/m3); KO,A, autotrophic oxygen half-saturation
constant (g/m3); bA, decay rate of autotrophs (d1).
Contour lines of growth rate as a function of DO and
ammonium concentration are found in Fig. 1. The growth rate
is constant along each line. A change in DO or ammonium
concentration has a larger effect on the nitrification rate at
lower concentrations. In an activated sludge system, an
increase in DO concentration that increases the nitrification
rate will cause a decrease in ammonium concentration,
which would partly counteract the effect on the nitrification
rate.

3.

Material and methods

The work constitutes two different parts. First, different


ammonium controller settings were evaluated that all result

Fig. 1 e Contour lines with different autotrophic growth


rates. mmax,A [ 0.5 g new cells/g cell, d, KNH [ 1 mg/l,
KO,A [ 0.4 mg/l.

2103

in the same average effluent ammonium concentration. A


flow-proportional average was used, meaning an average
concentration based on a load weighted with the average flow
rate. The different controllers were tested to investigate which
controller resulted in the lowest energy consumption,
expressed as average daily KLa. Second, a mathematical optimisation was performed where the KLa-vector, constituting
KLa values for 24 h of operation, was minimised given the
constraint that the daily average ammonium concentration
should be below a certain value. The simulations with
different ammonium controllers and the results from the
optimisation was compared with the reference scenario of
constant dissolved oxygen set-point and evaluated for a range
of influent loads to the model.

3.1.

Simulation model

Simulations were performed in a simplified version of BSM1.


MATLAB version R2011a with Simulink (MathWorks) was
used. The MATLAB implementation of BSM1 was provided
by Lund University (courtesy of Dr. Ulf Jeppsson). The model
has one large aerobic compartment for nitrification. For
simplification, denitrification was not included in the model
to minimise the simulation time during the optimisation
procedure. Three versions of the model with different
levels of control were used. One where the ammonium
concentration in the aerobic compartment is controlled by
supervisory ammonium control, one with only dissolved
oxygen control and one where a KLa-vector is specified in
advance. The model including supervisory ammonium
control is depicted in Fig. 2. For all models, the internal
recycle, return activated sludge rate and waste activated
sludge rate were the same in all simulations. The internal
recycle was set to approximately three times the influent
flow. The return activated sludge rate and waste activated
sludge rate were the same as in the default BSM1 steadystate settings.
All the days had an identical influent during simulation.
The time varying parts of the influent were sinusoids with
a period time of 24 h. Different influents were tested with
respect to average and amplitude of the sinusoids. All the
variables were treated the same when amplitude or averages
was changed. The different influents used in the simulations
are described in Table 1 and compared with the default BSM1
dry influent. Two different levels of the average influent and
the amplitude of the influent are used and combined in
Influent 1e4. Since the return and waste activated sludge rates
were not manipulated between the simulations, the sludge
age varies slightly when using different influents. The aerobic
sludge age was 7.8 d for Influent 1 and 2 and 7.2 d for Influent 3
and 4, with respect to the biomass concentration (XB,A and
XB,H). This can be compared with the standard Benchmark setup with default conditions which has a total sludge age of
approximately 9 days.
An influent similar to the BSM1 dry influent in amplitude
was not used since the large variations would cause the dissolved oxygen controller to saturate at 4 mg/l even with
moderately fast ammonium controllers. The control authority
is lost during these periods. In reality, plants may have larger
variations in incoming load depending on local conditions.

2104

w a t e r r e s e a r c h 4 6 ( 2 0 1 2 ) 2 1 0 1 e2 1 1 0

Fig. 2 e Benchmark plant layout. Simplification of BSM1 in MATLAB, including supervisory ammonium feedback control.

Since process control is in focus in this work, very large


influent variations were not considered. Apart from the wish
to have control authority in the model for ammonium feedback control, the specific volumetric ammonium loads were
chosen within the range of what can be found in large
Swedish wastewater treatment plants.
A simulation proceeded as follows. The influent file to be
tested and relevant controller parameters were initialised.
Before the dynamic influent was run, the model was run in
steady state for 150 days, the constant input being
the average of the influent to the dynamic model. The
dynamic model was subsequently run during 8 weeks and
evaluation was performed on the last 24 h. The long simulation time was chosen to be sure that the initial transient
had died out.

3.2.

Supervisory ammonium control

Simulations were performed with different controller settings


and compared with the reference scenario of constant DO setpoint. Two levels of oxygen set-points were used for each
influent: 1.5 mg/l and 2.5 mg/l resulting in different mean flowproportional effluent ammonium concentrations (a). Supervisory proportional-integral (PI) control was used, with feedback
from the effluent ammonium concentration. The proportional
constant (K ) ranged from 0.1 to 5 and the integral time
constant (Ti) from 0.01 to 3 d in the ammonium controller. The
settings in the dissolved oxygen controller were the default
benchmark settings (K 100, Ti 0.01 d).
For each influent and DO level, all simulations yielded the
same value on a as the scenario with constant dissolved

Table 1 e Influents used in the simulations, compared with the standard BSM1 dry influent. Amplitude refers to the peakto-peak amplitude.
State variable

Abr.
3

Soluble inert organic matter (gCOD/m )

SI

Readily biodegradable substrate (gCOD/m3)

SS

Particulate inert organic matter (gCOD/m3)

XI

Slowly biodegradable substrate (gCOD/m3)

XS

Active heterotrophic biomass (gCOD/m3)

XBH

Ammonium and ammonium nitrogen (gN/m3)

SNH

Soluble biodegradable organic nitrogen (gN/m3)

SND

Particulate biodegradable organic nitrogen (gN/m3)

XND

Alkalinity (mol/m3)

SALK

Influent flow (m3/d)

Qi

Influent

BSM1 dry

Mean
Amplitude
Mean
Amplitude
Mean
Amplitude
Mean
Amplitude
Mean
Amplitude
Mean
Amplitude
Mean
Amplitude
Mean
Amplitude
Mean
Amplitude
Mean
Amplitude

30
0
65
80
51
89
201
156
28
26
30
30
7
8
11
10
7
0
19 764
20 647

30
0
68
21
53
16
215
65
30
9
32
10
7
2
11
3
7
0
11 592
3 477

30
0
68
41
53
32
215
129
30
18
32
19
7
4
11
7
7
0
11 592
6 955

30
0
86
21
67
16
269
65
37
9
40
10
9
2
14
3
7
0
14 489
3 477

30
0
86
41
67
32
269
129
37
18
40
19
9
4
14
7
7
0
14 489
6 955

w a t e r r e s e a r c h 4 6 ( 2 0 1 2 ) 2 1 0 1 e2 1 1 0

oxygen concentration. The precision of the ammonium


concentration was set at 0.0001 mg/l. The goal a is
not reached when using ammonium feedback since the flow is
not accounted for in the controller and the average flowproportional concentration will be too high. This also occurs
when the DO controller saturates. Therefore, a lower
controller set-point is required. In the simulations, an iterative
strategy in MATLAB was used to find the appropriate set-point
for each combination of controller settings. The dissolved
oxygen set-point was limited between 0.5 and 4 mg/l.

3.3.

Optimisation of the oxygen transfer coefficient

The results from the simulations with constant dissolved


oxygen set-point and supervisory ammonium feedback
control were compared to the optimal solution to the
following optimisation problem:
Zt2
KL atdt

min

(3)

t1

subject to

1
t2  t1

Zt2
t1

NH4 tQt
dt  a  0
Qmean

(4)

where KLa is the oxygen transfer coefficient (d1), NH4 is the


ammonium concentration in the aerobic compartment, Q is
the flow to the plant, Qmean is the average daily flow and a is
the desired flow-proportional ammonium concentration. The
integrals span over 24 h and are discretised in N steps, where
N 36. N is a balance between optimisation time and the
resolution of the results. The optimisation problem was
solved in MATLAB with fmincon, a function for non-linear
constrained optimisation. Optimisation was performed in the
simplified BSM1 model where the aeration is not controlled
and KLa was decided before the start of the simulation. The KLa
values at each of the 36 instances make up a vector with KLa
values representing one day of operation. The constraint
function in fmincon calls the Simulink model which is run with
the pre-determined KLa-vector and then evaluates eq. (4). The
KLa-vector is iteratively changed towards the optimal solution
by the optimiser, given the objective function eq. (3). The
initial guess of the KLa-vector was a constant vector given by
the minimum average KLa obtained in the simulations with
ammonium feedback controllers. The influents and values on
a was the same as in the simulations with ammonium feedback control. The KLa-vector was unbounded in the
optimisation. For comparison, optimisations were also performed using an arithmetic mean in eq. (4), compared to
a flow-proportional mean.

4.

2105

Influent 1 (a 1.29 mg/l) are depicted as a surface in Fig. 3.


The level of suppression of ammonium by the controllers is
expressed by the daily variance of dissolved oxygen
concentration. The result from the reference simulation
with constant dissolved oxygen concentration and the result
from the optimisation are also included in the figure. The
daily KLa, oxygen and ammonium profiles for the reference
with constant oxygen concentration, the best performing
ammonium feedback controller, a faster feedback controller
and the simulation with optimal KLa-vector are found in
Fig. 4.
Figs. 3 and 4 suggest there is a minimum required energy
consumption. The minimum is found by a small variation of
the dissolved oxygen concentration. In Fig. 3, it is shown that
the controller with lowest energy consumption for each
setting on K is a controller with as low integral action as
possible, i.e. with a high value on Ti. The difference of using
a Ti of 2 instead of 3 d is small. Compared to the optimal
solution in Fig. 4, the KLa and ammonium concentration in the
most efficient feedback controller (K 0.75, Ti 3 d) is almost
identical while the DO concentration is slightly lagged. Two
different controller settings might cause the dissolved oxygen
concentration to vary to the same extent, expressed in variance, but the controller with the highest value on Ti will have
lower energy consumption for the same treatment effect. In
this respect, a purely proportional controller would be the
most efficient. Similar to the average KLa in Fig. 3, the removal
efficiency, expressed as the sum of KLa divided by the total
amount removed ammonium, also has its minima with the
controller settings K 0.75 and Ti 3 d.

Results

The results from the test of the different feedback controllers yielded data over what combination of controller setting
that achieved the lowest energy consumption (expressed in
KLa) for the same average effluent ammonium concentration
(expressed as a flow-proportional average). The results from
simulations with different ammonium controllers with

Fig. 3 e Average oxygen transfer coefficient (KLa) as


a function of variance of dissolved oxygen (DO) for different
values of the ammonium feedback controller settings. K
[ L0.1 to L5, Ti [ 0.03e3 d. The results from the constant
DO reference simulation (-) and from the optimisation
(C) are included. All simulations have an average effluent
flow-proportional ammonium concentration of 1.29 mg/l.

2106

w a t e r r e s e a r c h 4 6 ( 2 0 1 2 ) 2 1 0 1 e2 1 1 0

When the goal is to reach a daily average ammonium


concentration instead of a flow-proportional concentration,
the difference between constant dissolved oxygen concentration and the most efficient feedback controller is smaller
(Table 2).

5.

Fig. 4 e Daily variations in KLa, dissolved oxygen


concentration and effluent ammonium for different
simulations with Influent 1.

In Fig. 5, different ammonium controllers are compared for


Influent 1e4 and with different values on the flow-proportional ammonium concentration a. The integral time constant
is 3 d in all the simulations. All filled markers represent the
reference simulation with constant dissolved oxygen.
The plots with grey lines and markers represent the simulations with a higher value on a.
In Fig. 5, all influents show similar patterns but a lower
variance in the influent (Influent 1 and 3) makes the energy
requirement during constant DO control closer to the optimal
energy requirement than when the variance of the influent is
high (Influent 2 and 4). An approximate doubling of the variance of the influent ammonium doubles the optimal variance
of the oxygen concentration. This can be seen in Table 2,
which includes a summary of all the simulations, showing the
average and amplitude of the DO and ammonium concentrations and energy consumption expressed as KLa. The
reference simulation with constant oxygen concentration is
compared with the optimal solution and different levels of
feedback.
The ammonium controller settings were not decided by
any tuning method since a range of settings was used.
However, the settings could be compared with the controller
strom and Hagglund,
settings from a fast lambda tuning (A
1995) where K 5.2 and Ti 0.05 d. Using K 5 and
Ti 0.05 d, the dissolved oxygen controller saturates at the
maximum level of 4 mg/l during large parts of the time. With
fast ammonium feedback the average KLa is 10e14% higher
than for constant DO set-points depending on the influent
load and DO concentration, as can be seen in Table 2.
At higher dissolved oxygen concentrations (representing
the lower value on a for each influent) the constant dissolved oxygen strategy becomes relatively less efficient.
This effect is larger when the amplitude of the influent is
higher.

Discussion

The results in this paper demonstrate how the variations in


the influent load should be handled by means of control of the
dissolved oxygen concentration, when the control goal is to
achieve an average effluent ammonium concentration over
24 h. The discussion will focus on explaining the outcome of
the optimisation of the aeration based on the system
dynamics and discuss practical aspects and limitations when
the optimal solution might be difficult to use. The results will
also be related to previously published work.
The outcome of the optimisation is a result of the systems
non-linear dynamics in the aeration process and in the
biomass growth kinetics. For nitrification, DO and ammonium
is modelled as growth-limiting substrates. The two Monod
functions in eq. (2) can be considered as switching functions
which vary between zero and unity and that works complementary to each other. As could be seen in Fig. 1, the impact
on the ammonium removal rate is reduced at higher substrate
concentrations when the switching function approaches
unity. This implies that when high DO concentrations are
used to counteract an increased load, the increased energy
input cannot be fully compensated for during periods with
lower load. It also means that even though an increased
ammonium concentration can yield a higher removal rate, the
effect of a concentration change over certain ammonium
concentrations is limited. If only considering each Monod
function in eq. (2) separately, it is not beneficial to permit the
ammonium or oxygen concentrations to vary when the
influent load is changing. However, during dynamic conditions when the incoming load is varying, either the oxygen or
the ammonium concentration or both will vary. The control
task will be to balance these variations to meet the effluent
requirements, while also considering the energy demand.
According to the simulations in this study, the nitrification
rate can be optimised by increasing the variance of the DO
concentration while lowering the variance of the effluent
ammonium concentration just enough to have an optimal
removal rate profile along the day. The amplitudes will
depend on the operating range of DO and ammonium and the
half-saturation constants in the Monod functions. Also, at
higher DO concentration the DO approaches saturation. When
the DO concentration is high irrespective of influent load and
control action the control authority is limited.
When there are larger variations in the influent, additional
energy savings can be achieved by slow ammonium feedback
control compared to constant DO concentration. The ammonium concentration is reaching low levels more often at high
variations of the influent leading to a higher energy savings
potential. Keeping the DO levels constant when the ammonium concentration approaches zero could cause unnecessary aeration. This explains why there is most to gain by
feedback of ammonium at high influent variations and low

w a t e r r e s e a r c h 4 6 ( 2 0 1 2 ) 2 1 0 1 e2 1 1 0

2107

Fig. 5 e Average daily KLa as a function of the variance of dissolved oxygen for (a) Influent 1 (a [ 1.29 mg/l) and 2
(a [ 1.71 mg/l) (b) Influent 1 (a [ 1.06 mg/l) and 2 (a [ 1.36 mg/l), (c) Influent 3 (a [ 1.52 mg/l) and 4 (a [ 2.01 mg/l)
and (d) Influent 3 (a [ 1.23 mg/l) and 4 (a [ 1.55 mg/l). Ti [ 3 d.

average ammonium concentrations (Table 2). Theoretically,


larger variations in the influent load should make both
a constant DO control strategy and the optimal strategy relatively more efficient than fast supervisory ammonium
control, since suppressing the ammonium load variation
requires higher air flow rates and DO concentrations. This
effect cannot be seen in Table 2 since there is a maximum DO
limit which will be reached faster with larger influent variations or when operating at higher DO concentrations.
From an energy perspective it is close to optimal to meet
a time varying load with near constant oxygen concentration
given that treatment takes place in a completely stirred tank.
This is a convenient result, since this is what many process
engineers have used for many years. However, it is important
to recognise that finding a suitable constant DO concentration
for the process is only possible if the disturbance to the system
is repetitive. Adding a slow feedback loop to keep the effluent
ammonium on a specified level adds robustness to the process
operation, since in reality the load is never completely repetitive. Slow feedback makes sure that the effluent concentrations will stay on the specified control goal over longer time
periods. However, when using ammonium measurements for
process control the ammonium sensor will need sufficient
maintenance to provide a credible output signal.
A positive aspect of using slow ammonium feedback
compared to fast feedback is that the ammonium sensor can
be placed after the post sedimentation without affecting the
performance of the controller. This could result in better

measurements and less maintenance since the water is


cleaner. When a fast response is required the time lag created
by the hydraulic residence time in the sedimentation can
make it difficult to reach good results. Faster feedback will
provide more robustness to fast disturbances, but would also
use
more
energy
to
reach
the
same
effluent
concentrations. Slow feedback of effluent ammonium proving
to be most efficient in these simulations could fit well into
more advanced and multi-level control systems (Machado
et al., 2009; Yamanaka et al., 2006) as a slave controller fulfilling the ammonium set-point. Previous examples of slow or
averaging ammonium feedback in the literature include the
full-scale results with an ammonium floating controller (slow
integral controller) in Ingildsen (2002) and the supervisory
ammonium controller using a moving average to smooth the
daily variations implemented in full-scale in Ayesa et al.
(2006).
There are earlier examples in the literature that have
pointed towards higher energy consumption for faster and
more proactive controllers. In the comparison performed by
Stare et al. (2007) the aeration costs were lowest for oxygen PI
control, while more aggressive control strategies using
ammonium feedback or MPC had about 5% higher aeration
costs. A strategy with constant DO set-points has been
demonstrated to yield a 15% reduction in air flow rate
compared to fast supervisory ammonium control in full-scale
trials by Nordenborg et al. (2011). The effluent ammonium
concentration was kept on the same level.

2108

w a t e r r e s e a r c h 4 6 ( 2 0 1 2 ) 2 1 0 1 e2 1 1 0

Table 2 e Data from different runs with Influent 1e4 and for two different levels of ammonium reduction comparable to
a DO of 1.5 mg/l and 2.5 mg/l. Symbols are for comparison with Fig. 5.
a-level (mg/l)

1.29 B,
Fig. 5a

Influent
Constant DO

Best feedback (Ti 3)


Fast feedback
(K 5, Ti 0.05)

Optimal solution
flow prop. mean

Optimal solution
arithmetic mean
Change in average
KLa compared to
constant DO

Average KLa (d1)


Average DO (mg/l)
Average NH4 (mg/l)
Minimum NH4 (mg/l)
Maximum NH4 (mg/l)
Average KLa (d1)
K in controller
Average KLa (d1)
Average DO (mg/l)
Minimum DO (mg/l)
Max. DO (mg/l)
Average NH4 (mg/l)
Minimum NH4 (mg/l)
Maximum NH4 (mg/l)
Average KLa (d1)
Average DO (mg/l)
Minimum DO (mg/l)
Maximum DO (mg/l)
Average NH4 (mg/l)
Minimum NH4 (mg/l)
Maximum NH4 (mg/l)
Average KLa (d1)
Average DO (mg/l)
Optimal solution
(flow-proportional NH4) (%)
Optimal solution
(arithmetic NH4) (%)
Fast feedback:
(flow-proportional NH4) (%)

1.06 B,
Fig. 5b

1.71 ,,
Fig. 5a

1.36 ,,
Fig. 5b

1.52 V,
Fig. 5c

1.23 V,
Fig. 5d

2.01 D,
Fig. 5c

1.55 D,
Fig. 5d
4

141.7
1.5
1.25
0.72
1.94
140.4
0.75
165.3
1.89
0.53
4.01
1.29
1.09
1.44
140.3
1.41
1.04
1.79
1.26
0.83
1.80
140.7
1.43
1.0

171.6
2.5
1.03
0.61
1.56
169.6
1.0
194.6
2.70
0.86
4.03
1.04
0.76
1.40
169.7
2.38
1.88
2.90
1.03
0.66
1.50
170.6
2.42
1.1

145.7
1.5
1.51
0.46
3.24
141.8
0.50
165.3
1.58
0.48
4.02
1.61
0.81
2.37
141.3
1.23
0.60
1.92
1.55
0.69
2.91
142.8
1.29
3.1

176.7
2.50
1.21
0.40
2.53
170.8
0.75
198.3
2.38
0.49
4.06
1.26
0.73
2.24
170.2
2.09
1.23
3.06
1.22
0.47
2.40
172.6
2.21
3.8

217.2
1.50
1.48
0.88
2.27
215.2
0.75
246.9
1.84
0.60
4.01
1.52
1.35
1.65
215.2
1.42
1.07
1.80
1.49
1.03
2.08
215.8
1.44
1.0

263.0
2.50
1.20
0.75
1.79
260.3
1.0
295.3
2.69
0.95
4.03
1.21
0.93
1.58
260.1
2.39
1.92
2.88
1.20
0.80
1.71
261.2
2.42
1.1

220.9
1.50
1.82
0.59
3.87
214.5
0.50
248.1
1.61
0.48
4.02
1.93
1.10
2.69
213.9
1.24
0.64
1.93
1.86
0.89
3.41
215.8
1.29
3.2

267.8
2.5
1.42
0.50
2.92
257.9
0.75
298.5
2.43
0.50
4.07
1.47
0.89
2.53
257.6
2.12
1.29
3.07
1.43
0.61
2.73
260.6
2.21
4.0

0.7

0.6

2.0

2.4

0.7

0.7

2.3

2.8

14.3

11.8

11.9

10.9

12.0

10.9

11.0

10.3

There are examples of full-scale tests that have been in


favour of faster ammonium feedback and feedforward
control. The delicate balance between energy consumption
and treatment results has been demonstrated in full-scale by
Thornton et al. (2010). Compared to fixed set-points of dissolved oxygen, the aeration costs are significantly reduced but
the ammonium concentration was tolerated to be higher with
ammonium feedback control, making it difficult to distinguish
between the change in energy consumption resulting from the
shift in control strategy and the shift in treatment result. Good
control performance was achieved in pilot-scale with both
ammonium feedback and ammonium feedforward-feedback
control in Vrecko et al. (2006). The DO ranged from approximately 4.5 to 8 mg/l. Both strategies outperformed constant
DO concentrations with respect to energy consumption.
There is an interesting connection between the results in
this study which considers temporal variations of load over
the day, to the decision of a spatial DO profile which considers
variations along the basin for a given instant in time. Given
the same prerequisites as for the case with temporal variation
at a certain position, the results in this paper suggest a near
constant spatial DO profile ought to be optimal too. That
a constant DO profile along the activated sludge basin is the
best option has earlier been demonstrated by simulations in
the thesis by Ingildsen (2002), although this was not demonstrated by an optimisation procedure.

In this study it is assumed that the control goal is to reach


an average effluent concentration. There are situations when
the level of variance reduction of ammonium suggested in this
paper is not enough to guarantee the compliance to the
effluent permit. In some countries grab sampling is used
in compliance monitoring. Weekly averages and maximum
daily loads are other examples of short time frames in monitoring. In these situations the plant needs to be below
a desired effluent concentration all or most of the time andcontrollers more devoted to keeping an exact set-point
mand et al., 2011) or more advanced control algorithms
(i.e. A
as explained in i.e. Steffens and Lant (1999) could be considered. However, given that the minimum energy consumption
is reached with relatively slow control of the effluent ammonium, the results in this paper support longer time frames in
the effluent permits for treatment plants. For instance, annual
or monthly loadings should be premiered for waste load
allocations of nitrogen in the U.S. if the water quality of the
receiving water body can allow it. Similarly, the monitoring
within the river basin management plans in the EU Water
Framework Directive should allow for long-term averages in
the monitoring programmes of WWTPs when applicable.
Despite that the compliance to effluent criteria is estimated over long time periods, there might be special events in
operation or reasons related to the process configuration and
design which would motivate the use of a control strategy that

w a t e r r e s e a r c h 4 6 ( 2 0 1 2 ) 2 1 0 1 e2 1 1 0

move away from what is most energy efficient. One example


could be predenitrification, where low dissolved oxygen
concentrations in the last aerobic compartment could
improve denitrification.
Limitations caused by small aerobic volumes in relation to
the influent load could motivate a more varied oxygen
concentration. This study suggests that it is in these situations
that you gain the most from using ammonium feedback. If the
variance of the incoming ammonium is high, perhaps
combined with limited nitrification capacity, keeping
constant or near constant dissolved oxygen concentrations
might not provide the required average effluent ammonium
concentration. The removal rate will be too low during high
loaded periods and there is a risk for unnecessary aeration
during low loaded period. In the case of periodic unnecessary
aeration, one could consider control of the aerobic volume as
an option, as described by i.e. Samuelsson and Carlsson (2001)
or Svardal et al. (2003).
This study has been considering simulations of the activated sludge process. It is important to remember that the
dynamics that the results originate from are simplifications of
the actual full-scale process and that the results should be
verified in full-scale WWTPs. One simplification in the BSM1
model is that KLa is modelled instead of the air flow rate. If the
air flow rate dynamics was added to the model this would
render fast feedback even less efficient since high air flow
rates yield inefficient oxygen transfer. Another simplification
is the description of the biomass. Only one state variable is
used to describe nitrifying biomass (XB,A). Among ammonium
oxidising bacteria (AOB), there is a range of different halfsaturation values of ammonium (Koops and PommereningRoser, 2001) that in a real plant would determine the total
dynamics of the nitrifying microbial population. For this
reason, a topic for future research is to perform a sensitivity
analysis on the results in this paper with respect to the biokinetic parameters.
The discussion so far has focused on ammonium oxidation. It is not only growth of autotrophic bacteria that can be
described by Monod functions. In the ASM1 model, heterotrophic growth during aerobic conditions depend on easily
degradable carbon and oxygen in the same way as autotrophic
bacteria depend on ammonium and oxygen but with other
kinetic parameters. Nitrification has been in focus in this
work, but it can be noted that a parallel reasoning can be done
with respect to carbon removal. It is also important to recognise that the oxygen demand not only depend on oxidation of
ammonium, but also on oxidation of carbonaceous material.
The model implementation with only one aerated zone
provided the possibility to optimise one single KLa-vector, but
this will also lead to a lower average removal rate of ammonium compared to a solution with tanks-in-series. The influents used in the simulations were not only arranged to be
simple to interpret and to give good control authority of the
ammonium concentration, but end up unrealistic compared
to real influents. Real plant loads often changes faster than in
these simulations.
On full-scale sites, technology needs to mirror the demands
and resources available. In control one should always keep the
control strategy as simple as possible. This paper confirms in
simulations how a relatively simple control strategy of effluent

2109

ammonium can be energy efficient given long time frames on


the effluent criteria monitoring.

6.

Conclusions

If the compliance to the effluent nitrogen criterion is defined


over long time frames the effluent ammonium concentration
should be allowed to vary around a desired average value. The
control goal for the treatment could be translated to finding
a suitable amplitude reduction of ammonium in the effluent
that minimises the air flow rate. The following conclusions
can be drawn from optimisation and simulation in a nitrifying
single-zone BSM1 model:
 Near optimal aeration control can be achieved by a slow
ammonium feedback controller.
 Employing a constant dissolved oxygen concentration is
close to optimal. In this study, a constant oxygen concentration results in a 1e4% higher energy consumption than
the optimal solution depending on the influent variations
and effluent ammonium level.
 At higher variations in the influent, it is more favourable to
change the dissolved oxygen concentration by means of
ammonium feedback control, especially if the effluent
ammonium concentration is close to zero during parts of
the day.
 Fast feedback of effluent ammonium, with limitations on
the maximum and minimum dissolved oxygen concentrations of 4 and 0.5 mg/l respectively, required up to 14% more
energy than the optimal aeration strategy.

Acknowledgement
The authors would like to acknowledge the support provided
by Stockholm Vatten, Kappalaforbundet and Syvab which all
operate wastewater treatment plants in the Stockholm region.
Co-funding is provided from the Swedish Water and Wastewater Association (SWWA) (project number 29-116) and the
Foundation for IVL Swedish Environmental Research Institute
(SIVL) via grants from the Swedish Environmental Protection
Agency and the Swedish Research Council for Environment,
Agricultural Sciences and Spatial Planning (Formas). The
second author also acknowledges financial support from Formas (project number 211-2010-141) and SWWA (project
number 10-106). The MATLAB implementation of the BSM1
model has been provided by Lund University, courtesy of Dr.
Ulf Jeppsson.

references

mand, L., Nygren, J., Carlsson, B., 2011. Applications of repetitive


A
control in activated sludge processes. In: 8th IWA Symposium
on Systems Analysis and Integrated Assessment, 20e22 June,
San Sebastian, Spain.
strom, K.J., Hagglund, T., 1995. PID Controllers: Theory, Design
A
and Tuning, second ed. Instrument Society of America.

2110

w a t e r r e s e a r c h 4 6 ( 2 0 1 2 ) 2 1 0 1 e2 1 1 0

Ayesa, E., De la Sota, A., Grau, P., Sagarna, J.M., Salterain, A.,
Suescun, J., 2006. Supervisory control strategies for the new
WWTP of Galindo-Bilbao: the long run from the conceptual
design to the full-scale experimental validation. Water
Science and Technology 53 (4e5), 193e201.
Balku, S., Berber, R., 2006. Dynamics of an activated sludge
process with nitrification and denitrification: start-up
simulation and optimisation using evolutionary algorithm.
Computer and Chemical Engineering 30 (3), 490e499.
Beraud, B., Lemoine, C., Steyer, J.P., 2009. Multiobjective genetic
algorithms for the optimisation of wastewater treatment
processes. In: Kacprzyk, Janusz (Ed.), Studies in
Computational Intelligence. Springer-Verlag, Berlin
Heidelberg, Germany, pp. 163e195.
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.
Copp, J., 2001. The COST Simulation Benchmark: Description and
Simulator Manual. Office for Official Publications of the
European Community, Luxembourg.
de Arau`jo, A.C.B., Gallani, S., Mulas, M., Olsson, G., 2011.
A systematic approach to the design of operation and control
policies in activated sludge systems. Industrial & Engineering
Chemistry Research 50, 8542e8557.
Fikar, M., Chachuat, B., Latifi, M.A., 2005. Optimal operation of
alternating activated sludge processes. Control Engineering
Practice 7 (13), 853e861.
Henze, M., Gujer, W., Mino, T., van Loosdrech, M.C.M., 2000.
Activated Sludge Models ASM1, ASM2, ASM2d and ASM3. IWA
Publishing, London, UK.
Holenda, B., Domokos, E., Redey, A., Fazakas, J., 2007. Aeration
optimisation of a wastewater treatment plant using genetic
algorithm. Optimal Control Applications and Methods 28 (3),
191e208.
Ingildsen, P., 2002. Realising full-scale control in wastewater
treatment systems using in situ nutrient sensors. PhD Thesis.
Lund University, Lund, Sweden.
Ingildsen, P., Jeppsson, U., Olsson, G., 2002. Dissolved oxygen
controller based on on-line measurements of ammonium
combining feed-forward and feedback. Water Science and
Technology 45 (4e5), 453e460.
Jeppsson, U., 1996. Modelling aspects of wastewater treatment
processes. PhD Thesis. Lund Institute of Technology, Lund,
Sweden.
Koops, H.-P., Pommerening-Roser, A., 2001. Distribution and
ecophysiology of the nitrifying bacteria emphasizing cultured
species. FEMS Microbiology Ecology 37 (1), 1e9.
Lindberg, C.F., Carlsson, B., 1996. Nonlinear and set-point control
of the dissolved oxygen concentration in an activated sludge
process. Water Science and Technology 34 (3e4), 135e142.
Machado, V.C., Gabriel, D., Lafuente, J., Baeza, J.A., 2009. Cost and
effluent quality controllers design based on the relative gain
array for a nutrient removal WWTP. Water Research 43 (20),
5129e5141.
Monod, J., 1942. Recherches sur la croissance des cultures
bacteriennes. Hermann, Paris.
., Thunberg, A., Carlsson, B., 2011. Constant is
Nordenborg, A
optimal e a simple approach for aeration control in an

activated sludge process. In: 12th Nordic Wastewater


Conference, 14e16 November, Helsinki, Finland.
Olsson, G., Jeppsson, U., 1994. Establishing cause-effect
relationships in activated sludge plants e what can be
controlled. In: 8th Forum Applied Biotechnology. Med. Fac.
Landbouww., vol. 59. University of Gent, Gent, Belgium,
pp. 2057e2070.
Olsson, G., Nielsen, M.K., Yuan, Z., Lynggaard-Jensen, A.,
Steyer, J.-P., 2005. Instrumentation, Control and Automation
in Wastewater Systems. Scientific and Technical Report No.
11. IWA Publishing, London, UK.
Samuelsson, P., Carlsson, B., 2001. Control of the aeration volume
in an activated sludge process with nitrogen removal. Water
Science and Technology 45 (4e5), 45e52.
Samuelsson, P., Halvarsson, B., Carlsson, B., 2005. Interaction
analysis and control structure selection in a wastewater
treatment plant model. IEEE Transactions on Control Systems
Technology 13 (6), 955e964.
Shen, W., Chen, X., Pons, M.N., Corriou, J.P., 2009. Model
predictive control for wastewater treatment process with
feedforward compensation. Chemical Engineering Journal 155
(1e2), 161e174.
Stare, A., Vrecko, D., Hvala, N., Strmcnik, 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 (9), 2004e2014.
Steffens, M.A., Lant, P.A., 1999. Multivariable control of nutrientremoving activated sludge systems. Water Research 33 (12),
2864e2878.
Svardal, K., Lindtner, S., Winkler, S., 2003. Optimum aerobic
volume control based on continuous in-line oxygen uptake
monitoring. Water Science and Technology 47 (11),
305e312.
Thornton, A., Sunner, N., Haeck, M., 2010. Real time control for
reduced aeration and chemical consumption: a full scale
study. Water Science and Technology 61 (9), 2169e2175.
Thunberg, A., Sundin, A.-M., Carlsson, B., 2009. Energy
optimization of the aeration process at Kappala wastewater
treatment plant. In: 10th IWA Conference on Instrumentation,
Control & Automation, 14e17 June, Cairns, Australia.
Vrecko, D., Hvala, N., Kocijan, J., 2002. Wastewater treatment
benchmark: what can be achieved with simple control? Water
Science and Technology 45 (4e5), 127e134.
Vrecko, D., Hvala, N., Carlsson, B., 2003. Feedforward-feedback
control of an activated sludge process: a simulation study.
Water Science and Technology 47 (12), 19e26.
Vrecko, D., Hvala, N., Stare, A., Burica, O., 2006. Improvement of
ammonia removal in activated sludge process with
feedforward-feedback aeration controllers. Water Science and
Technology 53 (4e5), 125e132.
Weijers, S., 2000. Modelling, identification and control of
activated sludge plants for nitrogen removal. PhD Thesis.
Technische Universiteit Eindhoven, Eindhoven, The
Netherlands.
Yamanaka, O., Obara, T., Yamamoto, K., 2006. Total cost
minimization control scheme for biological wastewater
treatment process and its evaluation based on the COST
benchmark process. Water Science and Technology 53 (4e5),
203e214.

Vous aimerez peut-être aussi