Biomass and Bioenergy 132 (2020) 105420

Contents lists available at ScienceDirect

Biomass and Bioenergy

journal homepage: [Link]

Optimization of the energy consumption in activated sludge process using

deep learning selective modeling
Rafik Oulebsir a, *, Abdelouahab Lefkir b, Abdelhamid Safri a, Abdelmalek Bermad c
a
Laboratoire D’Environnement G�eotechnique et Hydraulique, Universit�e des Sciences et de La Technologie Houari Boumediene, Algiers, Algeria
b
Laboratory of TPiTe-Ecole National des Travaux Publics, Algiers, Algeria
c
Laboratoire Mat�eriaux de G�enie Civil Et Environnement, Ecole National Polytechnique D’Alger, Algiers, Algeria

A R T I C L E I N F O A B S T R A C T

Keywords: This paper presents a method using an artificial neural network for creating an optimal model of energy con­
Energy sumption in wastewater treatment plant (WWTP) using activated sludge process. The advantage of this method is
Deep learning the use data usually measured in most of WWTP to optimize the energy consumption of the biological process.
WasteWater
This method consists of selecting the data that represent the best energy consumption using different perfor­
Activated sludge
Optimization
mance criteria then use this data to train a deep neural network. The procedure of selection is divided into two
parts, in the first selection we selected the data that respect the environmental standards, and in the second part
we selected the data with optimal energy consumption using different pollution indicators, and this data was
used to train a deep neural network, finally the best model was used to estimate the energy savings on the data
not selected. The model showed good results with a coefficient of determination that varies between 90% and
92% in training period and 74%–82% in testing period, the application of the best model on the data not selected
showed a gain in energy for the most of the data.

1. Introduction potential can be exploited [7][8]. The majority of WWTP in the world
uses conventional activated sludge process, and a viable model of
During the past twenty years, the number of papers about energy optimization of energy consumption would have the widest range of
saving in wastewater treatment plant increases, and it shows the lack of application, and it would be interesting in an economic and environ­
tools for management and optimization of energy consumption. in the mental point of view. It could also allow electricity generation from
United States, 4% of electric energy is consumed by WWTP and the biomass technologies to be considered as an alternative to satisfy
water sector [1], and in Europe, the WWTP’s consume about completely or a major part of the energy demand of WWTPs [9].
27 TWh/year [2]. Energy costs represent approximately 28% of waste­ In Ref. [10] Nguyen et al. worked on optimization of aeration time
water costs [3]. It has been estimated that energy used in the WWTPs using activated sludge model (ASM1) and benchmark simulation model
comprises around one-fifth of a municipality’s total energy use by public (BSM1), the results showed 70% of gain in energy consumption for a
utilities, and it will continue to rise by 20% in the next 15 years with the small WWTP, and 30% for large WWTP, the main disadvantage of this
increasing water consumption and more stringent regulations [3,4]. The method is the necessity of relative precise measures that describe the
aeration process in secondary treatment is the highest energy con­ variation of physicochemical characteristics and the lack of adaptation
sumption part of the wastewater treatment technology [3,5]. In most of the method to different configuration of treatment processes[11][12].
medium and large WWTPs with Conventional Activated Sludge systems, Torregrossa et al. [13] worked on energy cost modeling using machine
aeration takes up approximately 50–60% of all electricity consumption, learning techniques, they compared different models (linear-­
while sludge treatment consumes 15–25% of energy, followed by sec­ exponential-logarithmic models) with Neural network and random for­
ondary sedimentation including recirculation pumps (15%) [3,6]. est models using a database of 317 WWTP using conventional activated
Despite the use of various datasets and methodologies, studies con­ sludge process situated in the north-west of Europe, they showed that
cerning the WWTP energy efficiency agree on a specific point: the the performances indicators of the machine learning cost models
WWTPs are generally not efficient and a relevant energy-saving (MLCM) surpass those found in the literature. This indicates the power

* Corresponding author.
E-mail address: [Link]@[Link] (R. Oulebsir).

[Link]
Received 24 March 2019; Received in revised form 31 October 2019; Accepted 4 November 2019
Available online 26 November 2019
0961-9534/© 2019 Elsevier Ltd. All rights reserved.
R. Oulebsir et al. Biomass and Bioenergy 132 (2020) 105420

of machine learning techniques to model energy consumption of WWTP 2.1.2. Second selection
[8]. The data resulted from the first selection will be used to determine
Longo et al. [14] reviewed many techniques of benchmarking of the best experiences in energy consumption according to KPIs. The se­
energy consumption using Key Performance Indicator (KPI) calculated lection consists into choosing the least value of energy for the same
with data easily measurable by most of WWTP. Three different ap­ values of the corresponding KPIs. in this paper we used 3 different KPIs,
proaches were described, Normalized approach, Statistical approach each one have its representation of the efficiency of the treatment:
and Programming techniques using Data envelopment analysis. In any sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
PN 2
case, the various benchmarking methods applied so far are mainly i¼1 X ii
GP ¼ (5)
diagnostic tools that fail at prescribing any improvement strategy to N
make inefficient WWTPs efficient [14].
To overcome this issue, This paper proposes a methodology of se­ GP
PI ¼ (6)
lection best experiences of energy consumption based on different KPIs maxðGPÞ
calculated using data usually measured in WWTP at daily time-step, and
optimize energy consumption using a Deep neural network trained with
– PI: Pollution Index
the data selected. The key benefit of this method is that it can be used on
several WWTP as a diagnostic tool or to predict the daily energy con­
The pollution index represents a normalized measure of pollution of
sumption of the WWTP.
the influent calculated from the concentrations of the different
pollutant, this measure indicates the level of pollution of the incoming
2. Material and method
sewage but doesn’t indicate the amount of pollutant that enters the
WWTP.
2.1. Method of selection
ðXii Xie Þ*Q
PABi ¼ (7)
2.1.1. First selection ðDXii Þ*Qd
The first selection aims to detect data with values of effluent quality sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
that are near the design values corresponding to environmental stan­ PN 2
i¼1 PABi
dards. To minimize the complexity of the multiple variables used in the GPAB ¼ (8)
N
study that can make difficult the modeling and the selection of the best
experiences, we used Key Performance Indicators (KPIs).To measure the
efficiency of the WWTP to treat the pollution of the influent we used – Q: Flow entering the WWTP(m3 )
Treatment Yield described as : – DXii : Design value of the i-th parameter of pollution of the influent
(mg/L).
Xie
Ti ¼ 1 (1) – Qd: Design value of the flow entering the WWTP (m3 ).
Xii
– PABi: Abatement degree of the i-th pollutant(%).
– GPAB: Global degree of abatement of pollution(%).
– Xie : Concentration of the i-th pollutant of the effluent of WWTP (mg/
L) The PABi represents the amount of pollution in terms of mass of each
– Xii : Concentration of the i-th pollutant of the influent of WWTP (mg/ pollutant that was treated by the WWTP, relative to the values WWTP
L) was designed for, if the value exceeds 1 that means the amount of
– Ti: Treatment Yield of the i-th parameter (%). pollution exceeds the capacity of the WWTP. As the same for the
sffiffiffiffiffiffiffiffiffiffiffiffiffiffi pollution index, GPAB represents the global efficiency of treatment of
PN
i¼1 Ti WWTP to treat all the amount of pollutants of the influent.
GT ¼ (2)
N
ðXii Nie Þ
Ii ¼ 1 (9)
ðmaxðXii Þ Nie Þ
– GT: Global Treatment Yield (%)
YN
WQI ¼ I wi
i¼1 i
(10)
The GT represents the global efficiency of treatment of the different
pollutant parameters, the values of the GT are compared to the global Where:
Standards Treatment Yield (SGT) to determine the experiences that are
optimized. The experiences with an insufficient treatment yield and the X
N
wi ¼ 1
experiences with a treatment yield too high (that express a waste of i¼1
(11)
energy) will not be selected. PIWQI ¼ 1 WQI
Nie
STi ¼ 1 (3)
Xii – Ii : Sub-indice of the Water Quality Index.
– WQI : Water Quality Index.
– Nie : Standard concentration of the i-th parameter of the effluent of – wi : Weights relative to each sub-indices
WWTP (mg/L). – PIWQI : Pollution index calculated with the WQI.
– STi: Standardized Treatment Yield of the i-th parameter (%).
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi The WQI has been developed by Brown’s and al, This effort was
PN 2 supported by the National Sanitation Foundation, United states of
i¼1 ST i
SGT ¼ (4) America (USA) (NSF) [15]. For this reason, Brown’s index is also
N
referred to as NSF-WQI [15]. The WQI expresses the quality of water and
increases with it, unlike GPAB and PI that increases with the level of
– SGT: Standardized global Treatment Yield (%). pollution of the influent. The WQI has the advantage of controlling the
influence of each parameter measured with the weights attributed, and
the subsequent investigations tended to show that multiplicative

2
R. Oulebsir et al. Biomass and Bioenergy 132 (2020) 105420

Fig. 1. Structure of the Deep neural network used.

formulation agreed better with expert opinion than did the additive one modeling. Multiple activations function exists and each one of them is
[15]. adapted to different problems, and there’s no function superior to the
other functions in all cases.
2.2. Deep neural network
2.2.2. Loss functions
In this section, we describe briefly Neural Network and Deep neural After designing the structure of the neural network, the next step is to
network, for more details we recommend you [16], and [17] which is adapt the model to data available by finding the weights using optimi­
the main resource for this section. zation algorithms, the goal is to find the combination of weights that
minimize or maximize a loss function. Different functions can be used:
2.2.1. Description of neural network
Neural networks are a computational model that shares some prop­ – Mean Squared Error:
erties with the animal brain in which many simple units are working in
1 XN
1 XM
�2
parallel with no centralized control unit. This type of model is usually MSE ¼ y
b yij (12)
N i¼1 M j¼1 ij
used to model a complex phenomenon with a little knowledge of the
structure of the phenomenon and an amount of data or experiments
available. – Mean Absolute Error:
The structure of a classical neural network is formed by three com­
ponents or layers: 1 X 1 X
N M � �
MAE ¼ �by yij � (13)
N i¼1 M j¼1 ij
– A single input layer
– One or many hidden layers, fully connected
– A single output layer – Mean Squared Log Error:

1 XN
1 XM
�2
Each layer is composed of several neurons, these neurons are fully MSLE ¼ log b
y ij log yij (14)
N i¼1 M j¼1
connected to the neurons of the next layer, the importance of the
connection is represented by weights. Just as biological neurons don’t The MSE is widely used and is a safe bet in most cases, So is the MAE
pass on every electro-chemical impulse they receive, artificial neurons [17]. The MSLE is a good choice when we use data with a large range
are not just wires or diodes passing on a signal [17]. They are designed due to the log function used in the formulae. These are all valid choices,
to be selective. They filter the data they receive, and aggregate, convert and there certainly is no single loss function that will outperform all
and transmit only certain information to the next neuron(s) in the other loss functions for every scenario[17].
network [17]. As these filters and transformations work on data, they
convert raw input data to useful information in the context of the larger 2.2.3. Optimization algorithm
multilayer perceptron neural network [17]. Stochastic Gradient Descent (SGD) and its variants are probably the
In the first layer (input layer), the neurons represent the explanatory most used optimization algorithms for machine learning in general and
parameters, that are detected in a phase called “Features Engineering” deep learning in particular [18].
which is very important in all machine learning projects. Different improvement of the SGD has been developed, the most
In the hidden layers and the output layer, each neuron has an acti­ known algorithms are AdaGrad - RMSProp - Adam, and for now no
vation function, that is used to represent the non-linearity in the

3
R. Oulebsir et al. Biomass and Bioenergy 132 (2020) 105420

consensus on the optimization algorithm to use. The choice of the al­ easily its values, So we will use the RMSE-observations standard devi­
gorithm depends on the type of problem modeled and the user’s famil­ ation ratio (RSR):
iarity with it, but Adam and RMSProp showed good performances on PN
practical Deep Learning problem. RMSE ð ybi yi Þ2
RSR ¼ ¼ Pi¼1
N
(17)
Standarddeviationobs i¼1 ðyi yÞ2
2.2.4. Deep learning and deep neural network RSR incorporates the benefits of error index statistics and includes a
The facets that differentiate deep learning networks in general from scaling normalization factor so that the resulting statistic and reported
“canonical” feed-forward multilayer networks are as follows [17]: values can apply to various constituents [21]. RSR varies from the
optimal value of 0, which indicates zero RMSE or residual variation and
– More neurons than previous networks therefore perfect model simulation, to a large positive value. The lower
– More complex ways of connecting layers the RSR, the lower the RMSE, and the better the model simulation
– “Cambrian explosion” of computing power to train performance [21].
– Automatic feature extraction
PN
ðyi ybi Þ
PBIAS ¼ i¼1 PN (18)
When we say “more neurons,” we mean that the neuron count has i¼1 yi
risen over the years to express more complex models [17]. More con­
nections means that our networks have more parameters to optimize,
and this required the explosion in computing power that occurred over – PBIAS : Percentage bias (%).
the past 20 years [17]. All of these advances provided the foundation to
build next-generation neural networks capable of extracting features for PBIAS represents the average tendency of the simulated values to be
themselves in a more intelligent fashion [17]. This allowed deep net­ smaller or larger than the observed ones. This index can indicate if the
works to model more complex problem spaces (e.g., image recognition model is overestimating or underestimating the observed values by
advances) than previously possible [17][19]. indicating respectively a negative or a positive value. This index can’t be
used to train a model because positive and negative values of the dif­
2.2.5. Structure of DNN ference between predicted values and observed values can compensate
The structure of the neural network used described in Fig. 1 consists each other in the sum, and that can mislead the user.
of four inputs (Temperature of the incoming wastewater - Recirculated
Sludge flow - incoming flow - KPI used for the data selected), each 3. Results and discussion
parameter have an impact on biological treatment[20]. The DNN has 6
hidden layers, each hidden layer has 200 neurons, the activation func­ 3.1. Description of WWTP
tion of the neurons is the Rectified Linear unit function (ReLU), the
optimization of the parameters of the DNN was performed by ‘Adam’ The WWTP is situated in Boumerdes in Algeria and was designed to
optimizer, using MAE as loss function. The structure of the DNN has treat urban sewage using a conventional activated sludge process. First,
been selected based on ‘try and error’ method to select the number of the arrival of sewage into the storm basin, a portion of the effluent is
hidden layers and the number of neurons of each layer. pumped directly into the pre-treatment of the channel to extract
maximum suspended solids [22]. The pretreated water is directed to 3
2.2.6. Performance indicators aeration basins that are mixed with aerated biomass and kept in sus­
To compare the performances of the different DNN, we need pension, each basin comprises 3 aerators [22]. We get mixed liquor
different indices, each indices have a different interpretation and allows composed of flocculated sludge and treated water directed to the clari­
a better analysis of each model: fiers [22]. In output thereof, the biomass is separated by decantation; a
PN part of the biomass is recirculated in the basins [22]. The excess biomass
ðyi ybi Þ2 is removed from the system and constitutes the secondary sludge [22].
R2 ¼ 1 Pi¼1 Where R2 2 ½0; 1� (15)
N
i¼1 ðyi yÞ2 At the end of the process, the clarified water passes into a concrete
structure to ensure prolonged contact between the water to be dis­
infected and chlorinated water [22]. At the exit of the plant, water is
– R2 : Coefficient of Determination.
discharged into the natural environment [22].
– yi : i-th Observed value.
The Boumerdes-WWTP has put at our disposal a series of daily data
– ybi : i-th simulated value. of pollution parameters at the entry and the exit, the flow and the total
– y : Mean of the observed values. energy consumed during the treatment process for a period that extends
from January 2006 to March 2016 (4018 days of observations). In this
The coefficient of determination is the most widely used measure for paper, we used four parameters of pollution (BOD5-COD-SS-NH4), the
model performance, it measures how well the predicted data are can fit temperature of influent, influent flow, the recirculated sludge flow to
the observed data. Due to its formulation, the value of the coefficient of aeration tank, and total energy consumed (see Table 1). After data
determination is influenced by the maximum data, in other words, we cleaning operation, 318 days of observations remains and will be used to
can have a good value of the coefficient of determination and observe determine the best model. The “data cleaning” process consists of
that simulated low values are not well simulated. eliminating the missing data, dubious and extreme values. The database
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
PN contains a high number of missing data (2 measure of a parameter per
i¼1 ð y bi yi Þ2 week, was reduced to 1 measure per week at the end of the period due to
RMSE ¼ (16)
N economic reasons) that are not concurrent (the missing data for the
parameters are not on the same day of observation), which led to an
– RMSE : Root mean squared error. important reduction of the database used in this study.

The RMSE is a measure of error between values predicted by the 3.2. First selection
model and the observed ones, this measure is sometimes used to train
neural networks, by minimizing its value we approach the best perfor­ The procedure of first selection was applied to the data, the pollution
mance of the neural network. Its disadvantage is that we can’t interpret parameters used in the selection are SS-BOD5-COD-NH4, used in

4
R. Oulebsir et al. Biomass and Bioenergy 132 (2020) 105420

Fig. 3.f where we see that many points are below the x ¼ y line.
Since the PIWQI is calculated by using only the pollution entering the
WWTP, and showed results that are close to GPAB model, we can assume
that the pollution entering the WWTP contribute more to the determi­
nation of energy consumption that the effluent values of pollution
parameters.
The performance of model with GPAB selected data can be due to the
number of data used for training and testing period, and the model
trained with PIWQI can represent the best model but we can’t affirm or
contest this hypothesis.

3.4. Reduction of energy consumption

Due to its performance we used the model trained with GPAB

selected data in order to determine the optimized energy consumption
on the data not selected, Fig. 4 represent the Energy Savings with the
Fig. 2. First selection of the data. Energy measured, we can observe that the model made saving on a large
number of data, this indicates the performance of the model. Three
calculations of GT and NGT. Therefore it is impossible to have values of clusters were observed after the application of Kmeans on the data, We
GT equivalent to SGT due to the precision of the measure and the lack of see that the Energy Savings is increasing with Energy observed, and
control of the process, we used Errstandards to compensate these errors. center of the clusters follow this tendency and describe a linear function.
Fig. 2 shows that the main percentage of the data is selected (69%), It is explained by the relative low variability of pollution load which
and the main percentage of the data not selected are data that surpass leads to a small range of energy needed for the treatment.
the standards of discharge, which indicates the high level of priority The data that belongs to the first cluster k1 are almost all experiences
allowed to high treatment yield in spite of overconsumption of energy where we observe a potential of saving in energy. The third cluster
(see Table 1). represents the data that can’t be optimized, and 2 hypotheses can
explain this result, first an insufficient treatment of the wastewater due
to technical causes, Second the lack of performance of the model due to
3.3. Second selection and modeling underfitting. The second cluster k2 represents the data that are near the
optimal values, we can observe that the majority of points are repre­
The results of the application of the second selection on the data are sented in this cluster (45.45%) and the gain in energy is relatively low so
described in Table 2, it shows a relatively small number of data selected that the center of the cluster is near the zero value.
that has resulted from the low dispersion of the values of the different These Results indicate clearly that there are three classes of energy
KPIs. consumption: Underconsumption, Overconsumption and optimal con­
As mentioned above the results from the second selection were used sumption of energy, and the limit between these classes can be clearly
to train and test DNNs for each KPI. The data used were divided into two determined.
sets, 80% for training period and 20% for the testing period.
The results of training and testing of DNNs (Fig. 3) show that all the
models have good performances despite the low number of data used.
All of the performance indices show that the model trained with GPAB
selected data is the best except for the RSR in the testing period, the
Table 2
values of PBIAS indicate that the model overestimates the values of
Values of performance indicators for the training and testing period.
energy which is confirmed by Fig. 3d. We observe also that the values of
Criteria of R2 (%) RSR (%) PBIAS (%)
the performance of the first model are spaced, which denotes overfitting.
Selection
The performances of the model are close to the values of performance
of the models developed by Torregrossa et al. [13]. And the results of a Period training testing training testing training testing

comparison of different models trained with different KPIs also confirm Pollution index 92.35 74.2 33.48 52.28 0.68 2.47
that the pollution load contributes more to the prediction of energy Degree of 94.41 82.42 23.82 48.28 0.38 0.56
pollution
consumption than removal efficiency [13].
abatement
The results of the model trained with PIWQI data have performance Water quality 90.76 82.07 33.16 45.02 0.5 4
close to the GPAB model especially with R2 , but the PBIAS indicates that index
the model underestimates the values of Energy, which is confirmed by

Table 1
Statistic parameters of the Boumerdes WWTP data.
Parameters Influent Effluent

MIN MAX Mean STD MIN MAX Mean STD

Temperature (� C) 7.55 35.93 20.35 3.82 2.43 29.9 20.27 4

SS (mg/L) 24 2400 240.18 137.31 1 160 13.53 8.21
BOD5 (mg/L) 20 2500 248.64 156.38 1 55 10.69 8.28
COD (mg/L) 110 3264 472.98 251.27 1 173 42.14 24.83
NH4 (mg/L) 0.8 126.54 30.98 18.49 <0.1 60.1 5.37 8.24
Flow (m3 /d) 42 22491 11991.3 4186.17 / / / /
Energy (kWh) 11 8355 3932.84 1082.98 / / / /
Recirculated Sludge (m3 /d) 0 27170 16497 4885.05 / / / /

5
R. Oulebsir et al. Biomass and Bioenergy 132 (2020) 105420

Fig. 3. Results of the different model trained on the different criteria.

3.5. Limitation of the methodology study, but we can suggest to readers to use transfer learning in order
to transfer the knowledge generated by a model to another WWTP.
We aware that our research may have limitations: For More details we recommend the book “Hands-On Transfer
Learning with Python by Dipanjan Sarkar et al. “[23].
– First, the limited number of data restricts the applicability of the
model to other WWTP with the same characteristics. Further 4. Conclusions
research can concentrate on the use of this method on a large
database. The purpose of the current study was to determine a method using
– The method is based on one criterion in the first and second part of DNN in order to create an optimal model of energy consumption. The
the selection, which leads to judging the efficiency of the experiences results of this study support the idea that we can optimize energy con­
based only on one criterion while different parameter influences the sumption by using the data previously measured on the WWTP. We have
energy consumption of the WWTP. been able to optimize the values of energy consumption by creating a
– The transferability of the model to other WWTP is due to the char­ model using Deep Neural Network coupled with a procedure of selection
acteristics of Neural Networks. This issue is beyond the scope of this using different KPI. This model was applied to the data judged

6
R. Oulebsir et al. Biomass and Bioenergy 132 (2020) 105420

[7] D. TORREGROSSA, A Decision Support System for Energy Saving in WasteWater

Treatment Plants, PhD thesis, Universit�e du Luxembourg en Science de l’ing�enieur,
2018.
[8] R. Maachou, A. Lefkir, A. Bermad, A. Khouider, Modeling of activated sludge
process using artificial neuro-fuzzy-inference system (anfis), Desalin. Water Treat.
57 (Nov 2015) 21182–21188, [Link]
[9] P. Kosse, M. Lübken, M. Wichern, Urban lignocellulosic biomass can significantly
contribute to energy production in municipal wastewater treatment plants - gis-
based approach for a metropolitan area, Biomass Bioenergy 81 (Oct 2015)
568–573, [Link]
[10] D.H. Nguyen, C. Dagot, N. Roche, Optimisation de la conception et du
fonctionnement des stations de traitement des eaux us�ees, PhD thesis, Universit�e de
Lorraine, 2014.
[11] K.V. Gernaey, M.C. Loosdrecht, M. Henze, M. Lind, S.B. Jørgensen, Activated
sludge wastewater treatment plant modelling and simulation: state of the art,
Environ. Model. Softw 19 (Sep 2004) 763–783, [Link]
envsoft.2003.03.005.
[12] S. Gillot, J.-m. Choubert, Biodegradable organic matter in domestic wastewaters:
comparison of selected fractionation techniques, Water Sci. Technol. 62 (aug 2010)
630–639, [Link]
[13] D. Torregrossa, U. Leopold, F. Hern� andez-Sancho, J. Hansen, Machine learning for
Fig. 4. Energy savings calculated using the GPAB trained model. energy cost modelling in wastewater treatment plants, J. Environ. Manag. 223 (Oct
2018) 1061–1067, [Link]
[14] S. Longo, B.M. d’Antoni, M. Bongards, A. Chaparro, A. Cronrath, F. Fatone, J.
nonoptimal and resulted in a model that can be applied to predict the M. Lema, M. Mauricio-Iglesias, A. Soares, A. Hospido, Monitoring and diagnosis of
energy consumption needed for the treatment, and in the detection of energy consumption in wastewater treatment plants. a state of the art and
three classes of energy consumption with relatively clear limits, these proposals for improvement, Appl. Energy 179 (Oct 2016) 1251–1268, [Link]
org/10.1016/[Link].2016.07.043.
limits will warn or help the operators about the energy consumed by the
[15] T. Abbasi, S. Abbasi, Water Quality Indices, first ed., Elsevier, 2012, pp. 26–29.
different processes. The results of this research can be applied to reach [16] A. Burkov, The Hundred-Page Machine Learning Book, First Edition, Andriy
energy self-sufficiency of WWTPs by reducing the overconsumption and Burkov, 2019, pp. 70–82.
allow electricity generated from biogas to satisfy the energy needed for [17] J. Patterson, A. Gibson, Deep learning A Practitioner’s Approach, First Edition,
O’Reilly Media, Inc, 2017, pp. 23–80.
the aeration process [9,24]. [18] I. Goodfellow, Y. Bengio, A. Courville, Deep Learning, MIT Press, 2016. htt
Further work will focus on the improvement of the selection method p://[Link].
to detect much accurately optimal data, and will also focus on testing [19] Y. LeCun, Y. Bengio, G. Hinton, Deep learning, Nature 521 (May 2015) 436–444.
[Link]
different types of machine learning models like modular neural network [20] R. Maachou, A. Lefkir, A. Bermad, A. Khouider, Factorization of physicochemical
[25–27] or decision trees. Further investigations are also needed in parameters of activated sludge process using the principal component analysis,
order to estimate the distribution of the time of aeration and the dis­ Desalin. Water Treat. 57 (Nov 2015) 20292–20297, [Link]
19443994.2015.1108425.
tribution of global energy consumption on the different aerators during [21] D.N. Moriasi, J.G. Arnold, M. Liew, R.L. Bingner, R.D. Harmel, T.L. Veith, Model
the treatment. evaluation guidelines for systematic quantification of accuracy in watershed
simulations, Am. Soc. Agric. Biol. Eng. 50 (2007) 885–900, [Link]
10.13031/2013.23153.
References [22] R. Maachou, A. Lefkir, A. Bermad, T. Djaoui, A. Khouider, Statistical analysis of
pollution parameters in activated sludge process, Desalin. Water Treat. 72 (Dec
[1] R. Goldstein, W. Smith, Water and sustainability : U.s electricity consumption for 2016) 85–91, [Link]
water supply and treatment–the next half century, EPRI (2002) 6–7, tech. rep. [23] D. Sarkar, R. Bali, T. Ghosh, Hands-On Transfer Learning with Python, Implement
[2] D. Torregrossa, G. Schutz, A. Cornelissen, F. Hern�
andez-Sancho, J. Hansen, Energy Advanced Deep Learning and Neural Network Models Using TensorFlow and Keras,
saving in wwtp: daily benchmarking under uncertainty and data availability First Edition, Packt Publishing, 2018, pp. 155–169.
limitations, Environ. Res. 148 (Jul 2016) 330–337, [Link] [24] S. Kalloum, H. Bouabdessalem, A. Touzi, A. Iddou, M. Ouali, Biogas production
envres.2016.04.010. from the sludge of the municipal wastewater treatment plant of adrar city
[3] Y. Gu, Y. Li, X. Lia, P. Luo, H. Wang, X. Wang, J. Wu, F. Li, Energy self-sufficient (southwest of Algeria), Biomass Bioenergy 35 (Jul 2011) 2554–2560, [Link]
wastewater treatment plants: feasibilities and challenges, in: E. Procedia (Ed.), The org/10.1016/[Link].2011.02.012.
8th International Conference on Applied Energy – ICAE2016 vol. 105, 2017, [25] P. Melin, I. Miramontes, G. Prado-Arechiga, A hybrid model based on modular
pp. 3741–3751, [Link] neural networks and fuzzy systems for classification of blood pressure and
[4] E. Means, Water and wastewater industry energy efficiency: a research roadmap, hypertension risk diagnosis, Expert Syst. Appl. 107 (Oct 2018) 146–164, https://
Awwa Res. Found. (2004) 15–23, tech. rep. [Link]/10.1016/[Link].2018.04.023.
[5] F. Liu, A. Ouedraogo, S. Manghee, A. Danilenko, A primer on energy efficiency for [26] D. S�anchez, P. Melin, O. Castillo, A grey wolf optimizer for modular granular
municipal water and wastewater utilities, Energy Sect. Manag. Assist. Progr. neural networks for human recognition, Comput. Intell. Neurosci. 2017 (Aug
(ESMAP) (2012) 11–12, tech. rep. 2017), [Link]
[6] D. Mamais, C. Noutsopoulos, A. Dimopoulou, A. Stasinakis, T. Lekkas, Wastewater [27] D. S�anchez, P. Melin, O. Castillo, Optimization of modular granular neural
treatment process impact on energy savings and greenhouse gas emissions, Water networks using a firefly algorithm for human recognition, Eng. Appl. Artif. Intell.
Sci. Technol. 71 (Jan 2015) 303–308, [Link] 64 (Sep 2017) 172–186, [Link]

7
 ARTICLE IN PRESS

Water Research 39 (2005) 2459–2474

[Link]/locate/watres
Review
A critical comparison of systematic calibration protocols for
activated sludge models: A SWOT analysis
Gürkan Sin, Stijn W.H. Van Hulle, Dirk J.W. De Pauw, Ann van Griensven,
Peter A. Vanrolleghem
BIOMATH, Department of Applied Mathematics, Biometrics and Process Control, Ghent University, Coupure Links 653,
B-9000 Gent, Belgium
Received 14 September 2004; received in revised form 3 May 2005; accepted 4 May 2005

Abstract

Keywords: Activated sludge models (ASMs); Calibration; Systematic protocols; SWOT analysis

Contents

1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2460
2. Systematic protocols for activated sludge model calibration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2461
2.1. Summary of the four systematic calibration protocols . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2461
2.1.1. The BIOMATH calibration protocol (Vanrolleghem et al., 2003) . . . . . . . . . . . . . . . . . . . 2461
2.1.2. The STOWA calibration protocol (Hulsbeek et al., 2002) . . . . . . . . . . . . . . . . . . . . . . . . 2463

Corresponding author. Tel.: +32 9 264 5937; fax: +32 9 264 6220.
E-mail address: [Link]@[Link] (G. Sin).

0043-1354/$ - see front matter r 2005 Elsevier Ltd. All rights reserved.
doi:10.1016/[Link].2005.05.006
 Journal of Water Process Engineering 31 (2019) 100819

Contents lists available at ScienceDirect

Journal of Water Process Engineering

journal homepage: [Link]/locate/jwpe

Plant-wide modeling and optimization of a large-scale WWTP using T

BioWin’s ASDM model
⁎
Abdelsalam Elawwad , Minerva Matta, Mohamed Abo-Zaid, Hisham Abdel-Halim
Environmental Engineering Dept., Faculty of Engineering, Cairo University El-Gamaa St., 12613, Giza, Egypt

A R T I C LE I N FO A B S T R A C T

Keywords: In this paper, a plant-wide model has been created for an Egyptian WWTP located in Cairo, namely the Gabal El-
Mathematical modeling Asfar, using BioWin 5.3 software. The model has been created according to the Good Modeling Practice protocol
Wastewater treatment where design reports, effluent analysis, and historical data have been collected - along with various site visits and
Plant-wide simulation a detailed sampling campaign. The collected data was used to acquire the necessary analysis for wastewater
BioWin
characterization and model calibration. The wastewater has been characterized and the model has been cali-
Optimization
brated according to STOWA protocol for calibration. Then, the model has been dynamically validated using the
analysis acquired over an entire year. The model’s accuracy was acceptable and therefore was created to perform
optimization studies. Multiple operational scenarios were simulated to achieve the highest effluent quality with
nitrogen removal, as the plant currently is not able to perform nitrification-denitrification. Different DO con-
centrations, RAS flows, SRTs and consequently WAS flows were varied to optimize the nitrogen removal. It was
found that nitrification and denitrification percentages of 94 and 62.4%, respectively, could be achieved easily
without modifying the WWTP’s configuration or unit capacities. This was achieved by increasing the SRT from
2.7 to 7 days, while converting some compartments of the aeration tank from aerobic to anoxic. Sludge and
water lines were both included in to assess the effect of any change in the operating conditions on all components
of the plant including sludge units.

1. Introduction BioWin© software (Envirosim, Canada) is a powerful simulator that has

plant-wide, built-in IWA models such as ASM1, ASM2d, and ASM3.
Wastewater contains many water pollutants and must be treated However, BioWin has its own general activated sludge-digestion model
prior to being discharged into the environment. Biological treatment is (ASDM) which allows users to model different aerobic and anaerobic
the most common method for treating wastewater and among the dif- biological processes without the coupling of more than one model [5].
ferent types of biological treatments, the activated sludge process is the The ASDM model has several processes which are divided into six
method most often applied as it removes organic matter and nutrients main models: an activated sludge model, an anaerobic digestion model,
from the wastewater [1]. Anaerobic digestion is the most widely-used settling models, a chemical precipitation model, a pH model, and an
biological process for sludge stabilization and energy recovery in was- aeration and gas transfer model. ASDM has a special anaerobic sludge
tewater treatment plants [2]. model that is characterised by different processes. These processes can
The various biological treatment processes used to treat wastewater be classified into three major categories: heterotrophic growth through
can be described by using mathematical equations and models [3]. fermentation, the growth and decay of propionic acid degrading
Different mathematical models for biological wastewater treatment acetogenic bacteria, and the growth and decay of methanogenic bac-
have evolved in recent decades. Among them, activated sludge models teria. BioWin is used frequently in the literature for optimizing muni-
(ASMs) are those used most often in optimization studies for municipal cipal WWTPs [6,7] and industrial WWTPs [8,9].
and industrial WWTPs [4]. ASM is a representation of microbial growth Few researches applied full-scale plant-wide modeling including
and substrate utilization within an activated sludge system through a anaerobic digesters because of its complexity [10]. In plant-wide
dynamic mathematical expression. ASMs are incorporated at different modeling, the WWTP is treated as a system that is completely in-
recent simulation platforms such as Simba, GPS-X, BioWin, etc. tegrated, where primary settlers and final clarification units, biological

⁎
Corresponding author.
E-mail addresses: elawwad@[Link] (A. Elawwad), minervaav@[Link] (M. Matta), eng2abozaid@[Link] (M. Abo-Zaid),
hishama.halim2011@[Link] (H. Abdel-Halim).

[Link]
Received 8 December 2018; Received in revised form 2 March 2019; Accepted 29 March 2019
Available online 08 April 2019
2214-7144/ © 2019 Elsevier Ltd. All rights reserved.
A. Elawwad, et al. Journal of Water Process Engineering 31 (2019) 100819

reactors, sludge digester units, sludge thickening units, sludge dewa- between 70 and 80% of the biodegradable organic pollutants. After
tering systems, and any other related processes are linked together that, the flow is directed to Zone #3 which represents the remaining
[11]. Regardless of its complexity, plant-wide modeling of WWTPs can two-thirds of the AT’s total volume and receives the other half of the
be used as a base for any future applications, optimization studies, or aeration capacity to biodegrade the rest of organic pollutants and sti-
for developing a process for achieving the highest effluent quality levels mulate the nitrification process. Fig. 1 shows a schematic of the AT and
and energy savings. Use of the BioWin general model is expected to its current aeration routine. Moreover, it has six secondary clarifiers
reduce the time and complexity of plant-wide modeling, especially (FST) that each has a diameter of 52 m and water depth of 3.9 m. Gabal
huge and multi-process modeling. So, in this study, BioWin’s ASDM El-Asfar Contract 19 is equipped with a sludge thickening system
model was used to optimize the case study of Gabal El-Asfar WWTP composed of gravity thickeners (STT) and dissolved air floatation (DAF)
(operated by Degrémont) which is a mega WWTP and the largest in for thickening of primary and secondary sludge, respectively. Then the
Middle East. It treats wastewater with an activated sludge process and thickened sludge is digested in eight primary anaerobic digesters (PD),
has anaerobic sludge digesters for sludge treatment. The effluent is each with a volume of 9760 m3, plus two secondary digesters (SD) each
discharged into a nearby agricultural canal, yet the effluent eventually having a volume of 8753 m3. Finally, the digested sludge is dewatered
reaches El-Manzala Lake through the Bahr El-Baqar agricultural drain - in a mechanical dewatering house (MDH) using a belt press system.
which is already suffering from environmental problems. Operational
decisions for such huge plants should not be taken lightly since they
2.2. Historical analysis and sampling campaign
directly impact the receiving bodies of water. To the best of our
knowledge, plant-wide modelling studies for Mega WWTPs WWTP are
In Gabal El-Asfar Contract 19 WWTP, the flow, oxygen concentra-
rare. In this paper, different operational routines were investigated, and
tion, and pH are measured online, while most of the other parameters
multiple optimization alternatives were proposed with the purpose of
are measured in the lab. These are composite samples that are collected
achieving better effluent quality for Gabal El-Asfar WWTP.
daily from different sampling points throughout the treatment process.
As per the calibration protocol, a sampling campaign was performed to
2. Material and methods verify the historical data; to complete the plant’s routine data; to
measure the mass balance requirements such as flow data, sludge
For the purposes of Gabal El-Asfar modeling, Good Modeling production, and SRT; to calculate the key conversions of N and COD; to
Practice (GMP) protocol is followed to show clear and organized steps. measure the influent characterization, including raw and settled in-
Historical data and design reports have been collected and 11 site visits fluent; to measure the effluent characterization; and to set up a good
have been made, in addition to a sampling campaign for performing base for model calibration.
important analyses that are not included in the plant’s routine analyses The sampling program was executed in five days: 24-hour compo-
for proper wastewater characterization, and a successful model cali- site samples were used for raw water influent, PST effluent and FST
bration according to GMP protocol requirements. treated effluent, while the rest of the measurements were performed
using four grab samples, due to insufficient tools and resources. Sludge
2.1. WWTP process description analysis was also performed for the thickeners influent, effluent, DAF
influent, DAF effluent, primary, and secondary digesters, and finally at
Gabal El-Asfar has different stages, where Gabal El-Asfar Contract the mechanical dewatering houses. During the sampling program, the
19, the case study of this paper, has a design capacity of 500,000 m3/ weather was favourable. There was no rainfall and the plant operated
day. Gabal El-Asfar Contract 19 has six circular primary settling tanks through dry weather conditions at an average air temperature of 24 °C.
(PST), each with a diameter of 53 m and a water depth of 3.2 m, and Wastewater measurements were performed following the Standard
eight biological aeration tanks (AT) with a total volume 109,840 m3. As Methods for the Examination of Water and Wastewater [12].
shown in Fig. 1, AT has a special configuration where its three-zoned The influent and effluent wastewater characteristics obtained from
circular aeration tanks include a central circular zone divided into two the sampling campaign are summarized in Table 1. Generally, Gabal El-
subzones (Zone #1 and #2), and an external plug-flow ring zone (Zone Asfar has a high effluent quality in terms of BOD, COD, and TSS re-
#3). Zone #1 and #2 represent one-third of the AT’s volume, receive moval, and the influent wastewater is very stable in terms of flow and
half of the aeration capacity, and are constantly aerated to eliminate pollutants. Also, the PST and FST are operated with high solids removal

Fig. 1. Schematic of the Aeration Tanks in Gabal El-Asfar Contract 19.