Biochemical Engineering Journal 133 (2018) 47–58

Contents lists available at ScienceDirect

Biochemical Engineering Journal

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

Regular article

Development of artificial neural networks to predict membrane

fouling in an anoxic-aerobic membrane bioreactor treating domestic
wastewater
Félix Schmitt a,b , Rajesh Banu c , Ick-Tae Yeom d , Khac-Uan Do a,∗
a
School of Environmental Science and Technology, Hanoi University of Science and Technology, Hanoi, Vietnam
b
Energy and Environmental Department, National Institute of Applied Sciences of Lyon, 69621 Villeurbanne Cedex, France
c
Department of Civil Engineering, Regional Centre of Anna University, Tirunelveli, India
d
Graduate School of Water Resources, Sungkyunkwan University, South Korea

a r t i c l e i n f o a b s t r a c t

Article history: An artificial neural network (ANN) was first developed to predict the transmembrane pressure in an
Received 17 October 2017 anoxic-aerobic membrane bioreactor (AO-MBR) treating domestic wastewater. A few studies about pre-
Received in revised form 10 January 2018 diction of membrane fouling in MBRs using ANNs have been published so far, even though our recent
Accepted 2 February 2018
work indicates that ANNs show a great potential for this application. In this study, 10 parameters linked to
Available online 5 February 2018
wastewater treatment and measured in the different parts of the AO-MBR system were used as the input
variables of the ANN. The goal was to select the most relevant input parameters to predict the evolution of
Keywords:
the transmembrane pressure based on the performances of the ANN. An ANN model was selected for its
Artificial neural network
Membrane bioreactor
satisfying performances (R2 = 0.850). In conclusion, ANNs could be a valid method to predict membrane
Membrane fouling fouling in AO-MBR systems treating domestic wastewater.
Transmembrane pressure © 2018 Elsevier B.V. All rights reserved.
Wastewater

1. Introduction brane fouling phenomenon has been widely developed. Numerous

Membrane bioreactor (MBR) technology is an increasing and
attractive option for the treatment and reuse of industrial and
municipal wastewater [1–5]. However, membrane fouling still
remains the most challenging issue in MBR operation and attracts
considerable attention in MBR studies [5–12]. Several techniques
are currently employed to mitigate membrane fouling in MBRs,
such as applying a pre-treatment to the influent, increasing the
aeration or chemically modifying the mixed liquor [7,13,14]. There-
fore, prediction of membrane fouling will play an important role in
adjusting the operational conditions to prolong the membrane fil-
tration time effectively. Currently, prediction of membrane fouling
in MBRs using a model which would be able to simulate the mem-
Abbreviations: Alk, alkalinity; ANN, artificial neural network; AO-MBR, anoxic-
aerobic membrane bioreactor; COD, chemical oxygen demand; DO, dissolved
oxygen; HRT, hydraulic retention time; MBR, membrane bioreactor; MLSS, mixed
liquor suspended solids; MLVSS, mixed liquor volatile suspended solids; MSE, mean
squared error; RMSE, root mean squared error; SRT, sludge retention time; TMP,
transmembrane pressure; TN, total nitrogen; TP, total phosphorus.
∗ Corresponding author.
E-mail address: uan.dokhac@hust.edu.vn (K.-U. Do).
simulations of the transmembrane pressure (TMP) or permeate flux
evolution using mathematical models [15–18] and computational
fluid dynamics simulations [19,20] gave satisfying and promising
results. However, the development of these models was limited
by the complexity of membrane fouling phenomenon and many
simplifications on the operational conditions and the wastewater
characteristics have to be made in order to obtain usable dynamic
equations.
Facing the difficulty to develop such models, artificial neu-
ral networks (ANNs) represented a good compromise to evaluate
membrane fouling in MBRs. ANNs are a statistical tool inspired by
biological neural systems. Basically, calculating units called neu-
rons, gathered in several layers, compose the network which takes
some variables as inputs and compute their outputs. The structure
of the network is adapted through a training phase by comparing
its outputs to experimental values which correspond to the target
outputs. When applied to MBRs, ANNs can allow to describe the
complex membrane fouling phenomenon by reducing it to a black-
box and considering only input parameters linked to the MBR and
the wastewater characteristics which could affect it.
The results of the following literature is summarized in Table 1
and further information about the ANNs developed is also given.

https://doi.org/10.1016/j.bej.2018.02.001
1369-703X/© 2018 Elsevier B.V. All rights reserved.
 48
Table 1
A summarization of the inputs, structure, output, and limitations for the ANNs used in MBRs.

Reference Outputs Inputs Structure Performance Comments/Limitations

Piron et al. [21] Rt TMP Recurrent Error = 11% • Dynamic simulation of membrane fouling but constant
Crossflow velocity operational parameters, which does not represent real
SS conditions in MBRs
Hamachi et al. [22] Ep TMP Recurrent Error = 10%
J Crossflow velocity

F. Schmitt et al. / Biochemical Engineering Journal 133 (2018) 47–58

SS
Çinar et al. [23] COD Ammonia COD Ammonia Cascade forward ANN Perfect match but no • Small database (only 30 data over 100 days of operation)
Nitrate Nitrate value of R2 or RMSE • Membrane fouling characteristics taken as inputs
Phosphate Phosphate
In the effluent In the influent
SRT
HRT
J
TMP
Pendashteh et al. [24] COD OLR Basic R2 > 0.99 • Optimization of the ANN structure with a genetic
TOC TDS ANN algorithm (GA)
oil in sludge reaction time operational time • Operational time taken as input
MLSS • Membrane fouling characteristics not considered either
In the effluent as inputs or outputs
Giwa et al. [25] COD DO Basic R2 > 0.97 • Membrane fouling characteristics not considered either
NH4 + -N PO4 3- P MLVSS ANN as inputs or outputs
In the effluent pH • Small database (only 16 data over 60 days of operation)
Electric conductivity
Geissler et al. [26] J TMP, dTMP/dt, TMP during backwash, Filtration Elman network Error = 2.7% • Database size >1000, which can lead to overfitting
cycle length, Backwash cycle length, SRT, TSS, 55 neurons • Influent and mixed liquor characteristics not considered
Temperature, Oxygen decay rate
Mirbagheri et al. [27] TMP TSS Basic ANN TMP: • Optimization with a GA
Perm CODin Error = 1% • 31 data only
SRT Perm: • Operational time taken as input
MLSS Error = 1%
Time
Radial basis ANN TMP:
Error = 2%
Perm:
Error = 3%
 F. Schmitt et al. / Biochemical Engineering Journal 133 (2018) 47–58
In recent years, many different types of ANNs have been used to
model membrane systems [21,22,28–32]. In particular, ANNs were
developed to predict the evolution of the hydraulic resistance and
the evolution of deposit thickness and permeate flux respectively
with the help of recurrent networks by taking TMP and other influ-
ent characteristics as inputs [21,22]. However, in these ANN models
only a single set of the initial input values was used to predict by
recurrence the whole set of the outputs. A cascade forward ANN
has been developed to predict the performance of a small-scale
submerged MBR treating cheese whey. The ANN input variables
were the influent total chemical oxygen demand (COD), ammo-
nia, nitrate and total phosphate concentrations, sludge retention
time (SRT), hydraulic retention time (HRT), the permeate flux and
the TMP [23]. A perfect match between measured and estimated
outputs was obtained for both training and testing phases. How-
ever, TMP and flux evolutions were taken as inputs, thus the ANN
was not considering them as outputs. The effluent characteristics
of a SBR-MBR treating synthetic hypersaline wastewater were pre-
dicted by an ANN which was composed of a single hidden layer.
Nevertheless, the prediction of membrane fouling characteristics
was not considered in the study [24]. Similarly, feed forward neu-
ral networks composed of a single hidden layer were developed to
predict the COD, NH4 + -N and PO4 3− -P concentrations in the efflu-
ent of a submerged MBR treating municipal wastewater. However,
the membrane fouling characteristics such as TMP or permeate
flux were also not considered in the model [25]. Geissler et al. [26]
developed an ANN to predict the permeate flux evolution in a pilot
MBR. The ANN consists of a single hidden layer. The inputs were
composed of TMP, rate of TMP change, filtration cycle length, back-
wash cycle length, SRT, total suspended solids, temperature and
oxygen decay rate in the aerobic zone. The permeate flux was the
output of the ANN. However, in this study, the parameters related
to wastewater characteristics such as COD and dissolved oxygen
(DO) concentrations or related to the mixed liquor such as mixed
liquor suspended solids (MLSS) concentration were not considered,
although they have a significant influence on membrane fouling.
Moreover, the predicted values of the permeate flux were likely to
overfit the experimental data, because the number of hidden neu-
rons was high. More recently, Mirbagheri et al. [27] investigated
membrane fouling in a pilot submerged MBR plant. The main goal
of the study was to determine the effect of simultaneous aeration on
submerged MBR performances by developing an ANN which would
be able to predict the TMP and the permeability of the membrane
over time. The problem was approached with two different mod-
els including a multilayer perceptron neural network and a radial
basis function neural network. For both simulations, the input vari-
ables were the operational time, total suspended solids, COD in
the influent, SRT and MLSS. The outputs consisted of permeabil-
ity and TMP. The ANN models were composed of one input layer,
one output layer and a single hidden layer. Then the prediction of
the permeability and TMP was considered. Nevertheless, the whole
experiment was based on a single series of measures which means
that the ANN models obtained could possibly not be able to predict
membrane fouling for various working conditions of the reactor.
The development of ANNs to predict membrane fouling in a
basic filtration system stayed until now limited to simple models
taking few and constant parameters as inputs. To the best of our
knowledge, a few studies about prediction of membrane fouling in
MBRs using ANNs have been published so far. Our recent work indi-
cates that ANNs show a great potential to predict membrane fouling
in MBRs [33]. Therefore, the aim of this study was to develop an ANN
model to predict membrane fouling from long-term anoxic-aerobic
MBR (AO-MBR) operational data. In particular, the ANN was trained
to predict the TMP considering the effect of different inputs such
as pH, alkalinity (Alk), MLSS, COD, total nitrogen (TN), ammoniacal
49
nitrogen (NH4 -N), nitrate (NO3 -N), total phosphorus (TP) on this
target.
2. Materials and methods
2.1. MBR operation
A lab-scale AO-MBR system has been used in this study (Fig. S1).
The working volumes of the anoxic and aerobic basin were 25 and
50 L, respectively. The HRT of the anoxic and aerobic basins were
3 and 6 h, respectively. In order to facilitate nitrogen removal, the
reactor was able to internally recycle the mixed liquor (IR = 300%Q)
using a peristaltic pump (Masterflex, Model GT-300D, Green Tech)
connected to the aerobic and anoxic basins. The anoxic basin was
equipped with a low speed mixer (Model MS 30120, 0–1500 rpm,
Lab. Stirrer) to maintain the MLSS in suspension. Five flat mem-
brane modules (Table S1) were submerged in the aeration basin.
The operating flux of the membrane was maintained at 17 L/m2 h
(LMH) by a peristaltic pump (Masterflex, Model GT-150D, Green
Tech). The pump was operated using cycles of 10 min of suc-
tion followed by 2 min of pause. In order to minimize membrane
fouling due to the cake layer formation, an aeration system was
placed below the membrane module, with the cross velocity of
uplifting air flow maintained within the range 20–22 cm/s. The DO
in the aerobic basin was maintained within the range 3–4 mg/L.
In order to control the DO concentration in the anoxic basin at
less than 0.5 mg/L for efficient denitrification, a degassing tank
(Volume = 12.5 L; HRT = 30 min) was added in the internal recycle
line, before the anoxic basin. The system was operated for about
180 days to treat domestic wastewater. The wastewater character-
istics used in this study are shown in Table 2.
2.2. ANN modelling development
An algorithm sketch of the model development process in this
study was presented in Fig. 1. The procedure includes several steps.
(i) Data collection and preprocessing: The main objective of the
data preprocessing is to determine the suitable locations for the
data required for the modelling activities. In this step, the inputs
and target outputs were normalized between 0 and 1 using the
function mapminmax. (ii) Model design and network training: The
model architecture, training method, and training rates were deter-
mined using a trial-and-error approach. In this study, the Neural
Network Toolbox V9.1 of MATLAB® mathematical software was
used to develop ANNs with the aim of predicting the TMP evolu-
tion in the AO-MBR by considering different input variables related
to the wastewater treatment. The overall structure of a basic ANN
along with the diagram of a neuron are represented on Fig. 2. In
the first phase of the ANN development, simple ANNs were created
using the fitnet function, which returns a basic backpropagation
network presenting a single hidden layer. For each test, the num-
ber of hidden neurons was fixed to 5 in order to study the influence
of the input variables chosen to feed the ANN and optimize the
selection of the wastewater treatment parameters which have to
be taken into account for TMP prediction. Defining a small number
of neuron allows indeed to avoid overfitting, which is the conse-
quence of a large number of weights which tends to give the ANN
the capacity to be flexible and easily fit the training dataset [34]. As
seen in Fig. 2, each connection between the nodes (or neurons) is
assigned with a weight and a bias, specific to each neuron, is added
to the sum of the weighted inputs. Moreover, an activation function
is defined for each layer of the ANN. Its role is to compute the output
of a neuron in function of its input and to introduce non-linearity
to the network, since these functions are commonly sigmoids. The
activation functions are specific to a layer and can be different for
50 F. Schmitt et al. / Biochemical Engineering Journal 133 (2018) 47–58

Table 2
Characteristics of the influent wastewater used in the study.

Range pH DO (mg/L) COD (mg/L) MLSS (mg/L) TN (mg/L) TP (mg/L) Alkalinity (mgCaCO3 /L) NO3 -N (mg/L) NH4 -N (mg/L)

Max 7.68 1.24 290.0 112.0 66.4 5.80 280.0 1.40 57.10
Min 7.36 0.79 154.0 18.0 42.8 3.80 230.0 0.10 30.10
Average 7.51 1.05 226.7 61.2 55.2 4.93 252.6 0.52 42.55
Standard deviation 0.07 0.11 27.0 20.8 5.4 0.37 10.5 0.34 6.43

Fig. 1. Algorithm sketch of the model development process.

Fig. 2. Graphical representation of the structure of a basic ANN (A) and of a neuron (B).

the hidden layers and the output layer. Eq. (1) gives the expression In Eq. (2), X is the normalized input, X is the original input and
of the output yi obtained from the i-th neuron of a defined layer as Xmin and Xmax are the minimum and maximum values respectively
a function of the input xj coming from the j-th neuron of the previ- of the input variables over the whole range of input values of the
ous layer (over n neurons) and entering this neuron, the weight wij dataset. The learning (or training) phase consists in adapting the
specific to the interconnection between the i-th neuron and the j-th weights and the biases of the ANN which are randomly defined at
neuron of the previous layer, the bias bi and the activation function the creation of the network, in order to minimize its performance
f. factor often defined as the mean squared error (MSE). The learning
procedure was launched with the function train which initialized
 

j=1 the random initial weights and biases for the ANN and applied the
yi = f bi + xj × wij (1) LM algorithm to minimize the MSE. The MSE, whose expression is
n given by Eq. (3), allows to compare, for the same inputs, the pre-
dicted outputs given by the ANN with the target outputs measured
In order to fit the demands of the activation functions, inputs and during the experiment.
targets of the ANN are normalized between 0 and 1. The normal-
ization function is given by Eq. (2).
i=1  2
 (i)
X − Xmin MSE = 1/N Y (i) − Yt (3)
X = (2)
Xmax − Xmin N
 F. Schmitt et al. / Biochemical Engineering Journal 133 (2018) 47–58 51

Table 3 Table 4
Summary of the parameters used as input variables for the study. Variations in the performances of the ANN trained 10 times.

Parameter Part of the AO-MBR Test N◦ R2 RMSE (cmHg) Order of importance

Influent Degassing tank Anoxic tank MBR Effluent 1 0.39 0.201 2−1-4-3
2 0.43 0.192 4−2-3-1
COD (mg/L) x x x x x 3 0.55 0.171 1-2-3-4
MLSS (mg/L) x x x x x 4 0.28 0.231 2-4−1-3
MLVSS (mg/L) x x x 5 0.43 0.194 4−2-1-3
pH x x x x 6 0.31 0.212 1-4-3-2
DO (mg/L) x x x 7 0.54 0.176 4−1-2-3
Alkalinity (mgCaCO3 /L) x x x x 8 0.36 0.209 3−1-4-2
TN (mg/L) x x x x 9 0.42 0.197 2−1-3-4
TP (mg/L) x x x x 10 0.42 0.195 3-4−1-2
NO3 -N (mg/L) x x x x
NH4 -N (mg/L) x x x x

(i)
In Eq. (3), Y (i) and Yt are the i-th predicted and target outputs
respectively over N data.
The training phase is operated with a precise learning algo-
rithm, the most common one being the Levenberg-Marquardt (LM)
algorithm, used for various ANN applications. Generally, a part of
the input database is used to stop the learning phase when the
minimum MSE over this part of the database has reached a local
minimum, even if the MSE over the training dataset keeps decreas-
ing. This part forms the validation dataset and is mainly used to
avoid overfitting of the network [35]. Another part of the dataset
is used to test the network after the training phase and evaluate
its accuracy. The performances of the ANN can be evaluated using
the root-MSE (RMSE) along with the coefficient of determination
R2 (Eq. (4)) and the relative error (Eq. (5)).
i=1  (i)
2
1− N
Yt − Y (i)
R2 =
i=1  2
(i)
Yt − Ŷ
N and 4 respectively. Table 4 presents the values of R2 and RMSE for
(i)
Yt − Y (i)
Error (%) = | | × 100
(i)
Yt
In Eq. (4), Ŷ is the average of the predicted outputs over the N data.
(iii) Model testing: The actual values of the inputs in the testing
datasets are compared to predicted values by the ANN models, to
evaluate the models performance. In this study, the whole exper-
imental database was adapted for each test, taking into account
the parameters used as the input variables of the ANN and their
frequency of measure, so that each input matched its correspond-
ing output. Then the input database was divided in a training, a
validation and a test dataset with ratios of 0.70, 0.15 and 0.15
respectively for the learning process. The whole database was dis-
tributed randomly to these 3 datasets. Table 3 indicates the input
variables which were available to use with the ANN predicting the
TMP evolution in the AO-MBR.
2.3. Analysis methods
Samples for the analyses of soluble constituents were cen-
trifuged at 5000 rpm for 5 min, and the supernatant then filtered
through a membrane with a pore size of 0.45 mm (GD/X PVDF,
Whatman). MLSS, mixed liquor volatile suspended solids (MLVSS),
COD, pH and nitrate were determined in accordance with Standard
Methods [36]. TN and TP were determined using a Hach spec-
trophotometer (Model: DR-2800). The ammonia concentrations in
the influent and effluent were measured using an ion-selective
electrode (Thermo Orion, Model 95-12). The cross flow velocity of
uplifting air was measured with an electromagnetic flow veloc-
ity meter (ACM-250-DL, Alec Electronics, Japan). Other operational
parameters such as permeate flow rate and TMP were monitored by
sensors and gauges installed in the system. All sensors and gauges
were cleaned and calibrated weekly.
3. Results and discussion
3.1. Selection of the input parameters
The different tests could not be performed by studying only one
ANN model each time. The performances of the trained ANN indeed
highly depend on the initial values of weights and biases which are
randomly defined at the first step of the training phase. The division
of the database operated by the fitnet function has also an impact on
the final values of weights and biases. To illustrate this characteris-
tic, 10 successive correlation tests for TMP were carried out using 4
variables as inputs: the permeate flux (J), COD concentration in the
MBR (CODmbr ), MLSS concentration in the MBR (MLSSmbr ) and the
(4)
OLR calculated from CODin and J, associated with numbers 1, 2, 3
these tests. The order of importance of the variables according to
the Garson’s algorithm, developed by Garson in 1991 [37], are also
(5)
indicated. This algorithm allows to determine the relative impor-
tance of each input variable of the ANN by partitioning the neural
network connection weights [38].
Despite a similar ANN architecture, each test led to different
performances. R2 ranged from 0.28 to 0.55 with a mean of 0.413
and a standard deviation of 0.0864. For these 10 successive tests,
the order of importance of the variables was never the same, despite
that several tests led to similar performances such as tests N◦ 2 and
5 and tests N◦ 9 and 10. For this last case, the importance of each
variable for one test was even at the opposite of the one for the
other test. Thus the order of importance does not seem to be linked
to the correlation efficiency of the ANN. The Garson’s algorithm
was therefore not considered to study the influence of the input
variables on the prediction of TMP.
For the next section, the ANN model selected for each test on
different sets of input variables was chosen to present a coefficient
of determination close to the mean of the coefficients of determi-
nation obtained over 20 successive ANNs trained using the same
inputs. Each parameter presented in Table 3 was studied sepa-
rately and the procedure was the same for each test: a first ANN
was trained using the whole variables whose concentrations were
measured in the AO-MBR, then different tests were carried out by
considering several combinations of these same variables in order
to determine which one was the most relevant for TMP prediction.
Finally, the last step of this part of the study consisted in combin-
ing the selected variables together in order to obtain an ANN model
which presents a high potential to predict TMP in the AO-MBR. In
the next section, the different parts of the AO-MBR were referenced
as –in for the influent, -deg for the degassing tank, -an for the anoxic
tank, -mbr for the MBR and –eff for the effluent.
52 F. Schmitt et al. / Biochemical Engineering Journal 133 (2018) 47–58
3.2. Correlation of pH and alkalinity input variables on TMP
prediction
During the experiment, pHin , pHan , pHmbr and pHeff have been
monitored. A test was carried out to assess the relative influence of
pH on TMP prediction. Fig. 3 shows the correlation performances
obtained by considering the 4 pH input variables for the ANN
(R2 = 0.427 and RMSE = 0.194cmHg). After using the ANN with com-
binations of the pH variables, pHan and pHmbr are the most influent
pH variables for TMP prediction as the ANN reached the highest
performances by considering only these 2 variables as inputs, with
R2 = 0.431 and RMSE = 0.203cmHg (Table S2).
The alkalinity in each part of the AO-MBR (in the influent, in the
anoxic tank, in the MBR and in the effluent) has been measured
along the experiment. A test using these concentrations (measured
as mg CaCO3 /L) as input variables of the ANN has been carried
out, the results being shown on Fig. 4. The performances of the
ANN reached R2 = 0.617 and RMSE = 0.158cmHg. The best correla-
tion results were reached considering Alkan , Alkmbr and Alkeff as
input variables, with R2 = 0.645 and RMSE = 0.152cmHg. However,
the performances reached with Alkmbr and Alkeff only were almost
similar (R2 = 0.626 and RMSE = 0.159cmHg) (Table S2). This combi-
nation can be selected as it allows to get rid of one variable, i.e. of one
series of measurements during the operation. The influence of the
DO variables was quickly analysed, as the performances were very
low for all the tests. They reached R2 = 0.162 and RMSE = 0.239cmHg
when the 3 DO variables where taken as input variables (Table S2).
DO used as input has no impact on membrane fouling. These results
were predictable because the amounts of DO in the MBR and in the
anoxic tank were controlled over the whole operation, and thus
were independent from the membrane fouling phenomenon.
3.3. Correlation of MLSS input variables on TMP prediction
Then MLSS concentrations have been used as input vari-
ables for TMP prediction. The input database was composed
of MLSSin , MLSSdeg , MLSSan , MLSSmbr and MLSSeff . The predic-
tion results are shown on Fig. 5. The correlation performances
when these variables were considered were very poor (R2 = 0.163
and RMSE = 0.238cmHg). After having gathered the performances
obtained for several combinations of MLSS input variables, MLSS
concentrations in the AO-MBR seemed not to be suitable for TMP
prediction, as the values of R2 stayed very low, under 0.20. Nev-
ertheless, MLSSan and MLSSmbr taken as the two input variables
presented the best correlation results. Additionally, when MLVSS
concentrations in the anoxic tank and in the MBR were added
to MLSSan and MLSSmbr as input variables, correlation perfor-
mances reached R2 = 0.210 and RMSE = 0.228cmHg and were thus
slightly better than those reached without considering MLVSS. But
the improvement is not significant enough to definitely consider
MLVSS as input variables of the ANN.
3.4. Correlation of COD input variables on TMP prediction
The importance of COD concentrations as input variables for
TMP prediction was assessed. The input database was composed
of CODin , CODdeg , CODan , CODmbr and CODeff . The results of the test
are shown on Fig. 6. As seen on Fig. 6, when the five COD variables
were considered, the correlation performances were R2 = 0.709 and
RMSE = 0.143cmHg. Knowing CODin and CODeff is not sufficient to
predict TMP evolution (R2 = 0.165). In other words, membrane foul-
ing which is directly linked to the value of TMP has no impact
on the COD removal rate. If CODmbr is added to the two previous
COD input variables, the prediction of TMP is improved but stays
very limited (R2 = 0.409). The best correlation performances were
obtained by using an ANN taking CODdeg , CODan and CODmbr as
input variables (R2 = 0.673). But according to the tests, only CODan
and CODmbr can be considered as input variables because the result-
ing performances are overall similar to those obtained with the best
correlation model, with R2 a bit lower (0.659 against 0.673) but a
RMSE slightly lower too (0.151 cmHg against 0.153 cmHg) (Table
S2). It can be inferred from the previous tests that COD concentra-
tions in the anoxic tank and in the MBR have much more impact on
TMP prediction than pH, Alk and MLSS ones.
3.5. Correlation of TN, NO3-N, NH4-N input variables on TMP
prediction
Removal of TN from the influent represents a major issue in
wastewater treatment. During the experiment, the TN amounts
were monitored in the AO-MBR. A first correlation test was car-
ried out using the 4 TN variables as inputs. The correlation results
are shown on Fig. 7. This test led to the best correlation results
until now, with R2 = 0.820 and RMSE = 0.109cmHg. This means that
TN measurements in the AO-MBR present a high potential to pre-
dict TMP evolution using ANNs. The performances reached by using
TNin and TNeff as input variables (R2 = 0.792 and RMSE = 0.118cmHg)
were practically similar to those obtained by using the 4 TN input
variables. These results showed the potential of these variables for
TMP prediction and highlighted the link between membrane foul-
ing and the TN removal rate in the AO-MBR. The TN removal rate
increased along with the TMP during the experiment from 47% to
66% approximately (Figure S2). However, a transitory phase has
to be noticed during the first 40 days, most likely due to the period
needed to achieve the normal operating conditions for the AO-MBR.
Thus, as a matter of fact, membrane fouling tends to improve TN
removal in the AO-MBR. In a second phase, the input variables TNin
and TNan on one hand and TNin and TNmbr in the other hand have
been considered to study the influence of the TN removal rate in
the anoxic tank and in the MBR respectively on TMP prediction. The
performances reached by the ANN were not satisfying in both cases,
with R2 close to 0.60. Thus the TN removal rates in the anoxic tank
and in the MBR are not reliable variables to be used for the ANN.
As well as TN, NO3 -N and NH4 -N concentrations have been mea-
sured in the influent, in the anoxic tank, in the MBR and in the
effluent. Four NO3 -N variables were first considered and taken as
inputs of the ANN to predict TMP. The results of this test are shown
on Fig. 8. The correlation performances reached R2 = 0.626 and
RMSE = 0.165cmHg. The amount of nitrate production in each part of
the AO-MBR seemed to have a mitigated influence on TMP predic-
tion. In other terms, membrane fouling did slightly affect nitrate
production throughout the whole treatment. Indeed, R2 did not
exceed 0.673 among all the tests considering several combinations
of variables. These performances were reached using Nitratembr
and Nitrateeff as input variables. The concentrations of nitrate in
the MBR and in the effluent decreased from 18.0 mg/L at the begin-
ning of the operation to around 6.0 mg/L until 180 days (Figure S3).
Thus nitrification in the MBR seemed to be limited by the increase in
membrane fouling. However, no link can be established between
the nitrification rate from the MBR to the effluent and the TMP
because the two corresponding concentrations were practically the
same and evolved similarly.
In a second time, the 4 variables related to NH4 -N concentrations
in the AO-MBR have been used as input variables of the ANN. The
results of this test are represented on Fig. 9. The performances of
this model were poor and lower than those obtained with the NO3 -
N variables. Indeed, the ANN gave R2 = 0.450 and RMSE = 0.201cmHg.
Moreover, the performances obtained with several combinations
of NH4 -N variables did not exceed R2 = 0.404. Thus, NH4 -N con-
centrations throughout the whole treatment are not likely to be
 F. Schmitt et al. / Biochemical Engineering Journal 133 (2018) 47–58 53

Fig. 3. Correlation plot (A) and evolution (B) of the target and the output TMP for pH input variables.

Fig. 4. Correlation plot (A) and evolution (B) of the target and the output TMP for alkalinity input variables.

Fig. 5. Correlation plot (A) and evolution (B) of the target and the output TMP for MLSS input variables.
54 F. Schmitt et al. / Biochemical Engineering Journal 133 (2018) 47–58

Fig. 6. Correlation plot (A) and evolution (B) of the target and the output TMP for COD input variables.

Fig. 7. Correlation plot (A) and evolution (B) of the target and the output TMP for TN input variables.

Fig. 8. Correlation plot (A) and evolution (B) of the target and the output TMP for NO3 -N input variables.
 F. Schmitt et al. / Biochemical Engineering Journal 133 (2018) 47–58
Fig. 9. Correlation plot (A) and evolution (B) of the target and the output TMP for NH4 -N input variables.
dependent on membrane fouling and in this way cannot be used as
input variables of the ANN to predict TMP evolution.
3.6. Correlation of TP input variables on TMP prediction
Similar tests have been led with TP variables. Fig. 10 shows
the correlation results obtained by using the 4 variables linked to
TP measurements in the AO-MBR as input variables: TPin , TPan ,
TPmbr and TPeff . Similarly to TN variables, TP variables taken as
inputs allowed to reach high performances with R2 = 0.805 and
RMSE = 0.114cmHg. However, unlike the TN removal rate, the pre-
diction of TMP using TPin and TPeff were not conclusive. Indeed,
the performances given by the ANN did not exceed R2 = 0.496 and
RMSE = 0.185cmHg. According to these results, the TP removal rates
in the AO-MBR are not linked to membrane fouling, or at least
to the TMP evolution. Indeed, despite a transitory phase during
the first 40 days of the experiment approximately, where the TP
removal rate progressively increased from 25% to around 40%,
the values seemed to vary randomly in a range from 38% to 46%.
Thus one cannot conclude on any relation between TMP evolution
linked to membrane fouling and TP removal rates in the AO-MBR.
However, the only input variables which resulted in high perfor-
mances were TPin and TPan . The performances obtained with the
ANN taking TPin and TPan as input variables reached R2 = 0.801 and
RMSE = 0.115cmHg, which indicates that the TP removal rate in the
anoxic tank presents potential for the prediction of TMP in the AO-
MBR. However, knowing the TP removal rate in the MBR is not
sufficient to predict TMP, as the ANN taking TPin and TPmbr as input
variables presented low performances which were R2 = 0.608 and
RMSE = 0.161cmHg (Table S2).
3.7. Effect of operational time as input on TMP prediction
The operational time is a central issue for the prediction of TMP
using an ANN. Indeed, the operational time can be taken as an
input variable along with other parameters linked to the AO-MBR
in order to predict the TMP evolution. This evolution implies indeed
a dynamic approach as membrane fouling increases with time and
presents several fouling phases which depend on many parame-
ters. Mirbagheri et al. [27] built a basic ANN composed of a single
hidden layer which aimed to predict the evolution of TMP in their
pilot sMBR by considering TSS, COD concentration in the influent,
SRT, MLSS concentration in the MBR and the operational time as
input variables. The ANN performances were very satisfying, with
55
Table 5
Comparison between ANN models either taking operational time as input variable
or not.
Set of input variables R2 without time R2 with time
CODan – CODmbr 0.659 0.950
MLSSan – MLSSmbr 0.210 0.959
pHan – pHmbr 0.431 0.911
TNin – TNeff 0.792 0.959
TPin – TPan 0.801 0.960
DOin – DOmbr 0.159 0.938
R2 values over 0.97 and a relative error below 3%. In order to eval-
uate the relevance of considering the time as an input variable of
the ANN predicting TMP in the AO-MBR, different models obtained
previously for different input variables and which presented var-
ious potentials (low to high) were trained by adding operational
time to the set of input variables. The performances obtained for
each model are reported in Table 5. In every case, which means for
either low performances ANNs such as DO one or high potential
ANNs such as TP one, considering the operational time as an addi-
tive input variable allowed to obtained ANNs with high correlation
performances with R2 above 0.90. These tests emphasized that the
operational time is a key parameter for TMP prediction using ANNs,
and that its influence on the fitting performances should be care-
fully studied. An ANN taking only the operational time as input
variable has been created, and the results can be seen on Fig. 11.
Considering this ANN model, the fitting performances between
the experimental TMP and the predicted TMP were almost per-
fect, with R2 = 0.991 and RMSE = 0.024cmHg. Knowing that, the ANN
appeared to be overfitting the predicted TMP with the operational
time, because the experiment was carried out on a single period
of time and was not repeated. In other terms, the ANN adapted to
the input database during the learning phase to fit with this specific
experiment. Thus an ANN model taking the operational time as one
of its input variables and whose database is constituted after a sin-
gle experiment corresponding to specific conditions could not be
used to predict TMP for other wastewater treatment conditions in
the AO-MBR. The database used by [27] to train their ANN (which
included the operational time) was obtained from a single exper-
imental period, that is why their ANN model was most certainly
overfitting the input data and gave such good performances (with
R2 values over 0.97). Therefore, a classification was proposed in
this study for the selected variables according to their potential in
membrane fouling prediction when they will be combined with
56 F. Schmitt et al. / Biochemical Engineering Journal 133 (2018) 47–58

Fig. 10. Correlation plot (A) and evolution (B) of the target and the output TMP for TP input variables.

Fig. 11. Correlation plot (A) and evolution (B) of the target and the output TMP obtained with the ANN taking the operational time as input variable.

Table 6 3.8. Combination of input variables on predicting TMP by ANN

Classification of the input variables selected for TMP prediction.

Selected variables R2 Class The final step for the selection of input variables of the ANN
CODan – CODmbr 0.659 Medium predicting the TMP consisted in combining the different couples of
MLSSan – MLSSmbr 0.196 Weak variables gathered in Table 6 together and select the optimal com-
pHan – pHmbr 0.431 Weak bination, i.e. the combination which led to the best performances
TNin – TNeff 0.792 Strong after training the ANN. Only the variables classified as medium and
TPin – TPan 0.817 Strong
strong were considered, the weak variables being inappropriate for
Nitratembr – Nitrateeff 0.673 Medium
Ammoniain – Ammoniaan 0.404 Weak TMP prediction. Table 7 presents a synthesis of the results obtained
Alkmbr – Alkeff 0.626 Medium from this study. According to the results obtained, the most relevant
DOin – DOmbr 0.159 Weak combination of variables to use for TMP prediction should be TNin ,
TNeff , Alkmbr and Alkeff , whose performances reached R2 = 0.875 and
RMSE = 0.091cmHg. However, it has been decided to let the Alk vari-
ables apart because the input database used when Alk variables
were considered was constituted of 46 data against 69 data for the
other variables (Table 6). Thus if for a couple of variables, R2 < 0.60,
other variables database. This is due to less frequent measures of Alk
it could be considered that it was too low to have any potential, even
in the AO-MBR compared to the other parameters. Following this
coupled to other variables. Then if for a single couple of variables,
choice, two combinations of variables presented equivalent perfor-
R2 > 0.75, then they present a high potential for the final perfor-
mances: TN – TP – COD and TN – TP – Nitrate. The TN – TP – COD
mances when coupled to other variables. For the medium potential,
model reached a higher R2 than the TN – TP – Nitrate one (0.855
it was proposed to take the range between weak and strong (tran-
against 0.850) but its RMSE was higher too (0.101 cmHg against
sition range). It was admited that the choice of the values (0.6 and
0.099 cmHg). Eventually, the TN – TP – Nitrate model was selected
0.75) was a bit empirical.
 F. Schmitt et al. / Biochemical Engineering Journal 133 (2018) 47–58
Table 7
Performances obtained for different combinations of the medium and strong vari-
ables taken as input of the ANN for TMP prediction.
Input variables R2
COD – TN – TP – Nitrate – Alk 0.834
TN – TP – Nitrate – Alk 0.852
TN – TP – COD – Alk 0.854
TN – TP – COD – Nitrate 0.831
TN – TP – COD 0.855
TN – COD 0.815
TP – COD 0.790
TN – TP – Nitrate 0.850
TN – Nitrate 0.822
TP – Nitrate 0.817
TN – TP – Alk 0.864
TN – Alk 0.875
TP – Alk 0.753
COD – Nitrate – Alk 0.802
TN – TP 0.843
because Nitrate variables gave better results than COD variables
when they were considered separately.
After having assessed the accuracy of each parameter measured
during the experiment, several conclusions can be made. MLSS,
COD, pH and DO which are common parameters of wastewater
treatment processes could not be linked to TMP as the perfor-
mances obtained by the ANNs taking them as input variables were
not satisfying (from R2 = 0.169 for DO to less than 0.70 for COD).
Moreover, one can obtain only an order of magnitude of the per-
formances of an ANN for a certain input database, because they
highly depend on the initial set of weights and biases created at the
beginning of the training phase, along with the database division
for training, validation and testing.
4. Conclusions
The prediction of TMP in the pilot scale AO-MBR using ANNs
has been shown to have potential. After having assessed the accu-
racy of each parameter measured during the experiment, several
conclusions can be made. MLSS, COD, pH and DO which are com-
mon parameters of wastewater treatment processes could not be
linked to TMP as the performances obtained by the ANNs taking
them as input variables were not satisfying (from R2 = 0.169 for DO
to less than 0.70 for COD). The combination of TNin – TNeff , TPin –
TPan and Nitratembr – Nitrateeff was performant for TMP prediction,
with R2 = 0.850. One can obtain only an order of magnitude of the
performances of an ANN for a certain input database, because they
highly depend on the initial set of weights and biases created at
the beginning of the training phase, along with the database divi-
sion for training, validation and testing. Further investigations on
the prediction of TMP in MBRs using ANNs are therefore needed.
Particularly, the input database should be constituted of data col-
lected during several independent operations of the MBR, in order
to avoid the ANN to overfit one specific experiment. A particular
attention should be payed to choosing the correct parameters for
training the ANN such as the initial set of weights and biases and
the division of the database and finding the optimal structure, i.e.
the number of hidden neurons or the number of hidden layers.
Acknowledgements
The authors thank for supports from EXPLO’RA Sup scholarship
provided by Région Auvergne-Rhône-Alpes (No. 2017/03/00061,
France). The additional supports from Hanoi University of Science
and Technology was highly appreciated (No. 09/QD-DHBK-
VKHCNMT, Vietnam).
57
Appendix A. Supplementary data
Supplementary data associated with this article can be found, in
RMSE the online version, at https://doi.org/10.1016/j.bej.2018.02.001.
0.102
0.115
0.097
References
0.113
0.101
[1] S. Judd, The status of membrane bioreactor technology, Trends Biotechnol. 26
0.113
(2008) 109–116.
0.116
[2] Z. Wang, Z. Wu, S. Mai, C. Yang, X. Wang, Y. An, Z. Zhou, Research and
0.099 applications of membrane bioreactors in China: progress and prospect, Sep.
0.106 Purif. Technol. 62 (2008) 249–263.
0.108 [3] S.J. Judd, The status of industrial and municipal effluent treatment with
0.098 membrane bioreactor technology, Chem. Eng. J. 305 (2016) 37–45.
0.091 [4] H. Lin, W. Peng, M. Zhang, J. Chen, H. Hong, Y. Zhang, A review on anaerobic
0.124 membrane bioreactors: applications, membrane fouling and future
0.112 perspectives, Desalination 314 (2013) 169–188.
0.100 [5] Q. Dong, W. Parker, M. Dagnew, Long term performance of membranes in an
anaerobic membrane bioreactor treating municipal wastewater,
Chemosphere 144 (2016) 249–256.
[6] F. Meng, S. Zhang, Y. Oh, Z. Zhou, H.-S. Shin, S.-R. Chae, Fouling in membrane
bioreactors: an updated review, Water Res. 114 (2017) 151–180.
[7] R. Habib, M.B. Asif, S. Iftekhar, Z. Khan, K. Gurung, V. Srivastava, M. Sillanpää,
Influence of relaxation modes on membrane fouling in submerged membrane
bioreactor for domestic wastewater treatment, Chemosphere 181 (2017)
19–25.
[8] M.-A. Yun, K.-M. Yeon, J.-S. Park, C.-H. Lee, J. Chun, D.J. Lim, Characterization
of biofilm structure and its effect on membrane permeability in MBR for dye
wastewater treatment, Water Res. 40 (2006) 45–52.
[9] E. Poorasgari, K. König, P. Fojan, K. Keiding, M.L. Christensen, Fouling of
enhanced biological phosphorus removal–membrane bioreactors by
humic-like substances, Chemosphere 117 (2014) 144–150.
[10] A. Sofia, W.J. Ng, S.L. Ong, Engineering design approaches for minimum
fouling in submerged MBR, Desalination 160 (2004) 67–74.
[11] K.-J. Hwang, C.-Y. Liao, K.-L. Tung, Effect of membrane pore size on the
particle fouling in membrane filtration, Desalination 234 (2008) 16–23.
[12] K.-U. Do, R.J. Banu, D.-H. Son, I.-T. Yeom, Influence of ferrous sulfate on
thermochemical sludge disintegration and on performances of wastewater
treatment in a new process: anoxic–oxic membrane bioreactor coupled with
sludge disintegration step, Biochem. Eng. J. 66 (2012) 20–26.
[13] S. Judd, The MBR Book: Principles and Applications of Membrane Bioreactors
for Water and Wastewater Treatment, 1st ed., Elsevier Science, Oxford, 2006.
[14] R. Campo, M. Capodici, G. Di Bella, M. Torregrossa, The role of EPS in the
foaming and fouling for a MBR operated in intermittent aeration conditions,
Biochem. Eng. J. 118 (2017) 41–52.
[15] I.M. Griffiths, A. Kumar, P.S. Stewart, A combined network model for
membrane fouling, J. Colloid Interface Sci. 432 (2014) 10–18.
[16] G.A. Pimentel, M. Dalmau, A. Vargas, J. Comas, I. Rodriguez-Roda, A. Rapaport,
A. Vande Wouwer, Validation of a simple fouling model for a submerged
membrane bioreactor, IFAC-PapersOnLine 48 (2015) 737–742.
[17] M.F.R. Zuthi, W. Guo, H.H. Ngo, D.L. Nghiem, F.I. Hai, S. Xia, J. Li, J. Li, Y. Liu,
New and practical mathematical model of membrane fouling in an aerobic
submerged membrane bioreactor, Bioresour. Technol. 238 (2017) 86–94.
[18] A. Charfi, M. Aslam, G. Lesage, M. Heran, J. Kim, Macroscopic approach to
develop fouling model under GAC fluidization in anaerobic fluidized bed
membrane bioreactor, J. Ind. Eng. Chem. 49 (2017) 219–229.
[19] A. Boyle-Gotla, P.D. Jensen, S.D. Yap, M. Pidou, Y. Wang, D.J. Batstone,
Dynamic multidimensional modelling of submerged membrane bioreactor
fouling, J. Membr. Sci. 467 (2014) 153–161.
[20] W. Zhang, X. Ruan, Y. Ma, X. Jiang, W. Zheng, Y. Liu, G. He, Modeling and
simulation of mitigating membrane fouling under a baffle-filled turbulent
flow with permeate boundary, Sep. Purif. Technol. 179 (2017) 13–24.
[21] E. Piron, E. Latrille, F. René, Application of artificial neural networks for
crossflow microfiltration modelling: black-box and semi-physical
approaches, Comput. Chem. Eng. 21 (1997) 1021–1030.
[22] M. Hamachi, M. Cabassud, A. Davin, M. Mietton Peuchot, Dynamic modelling
of crossflow microfiltration of bentonite suspension using recurrent neural
networks, Chem. Eng. Process. Process Intensif. 38 (1999) 203–210.
[23] Ö Çinar, H. Hasar, C. Kinaci, Modeling of submerged membrane bioreactor
treating cheese whey wastewater by artificial neural network, J. Biotechnol.
123 (2006) 204–209.
[24] A.R. Pendashteh, A. Fakhru’l-Razi, N. Chaibakhsh, L.C. Abdullah, S.S. Madaeni,
Z.Z. Abidin, Modeling of membrane bioreactor treating hypersaline oily
wastewater by artificial neural network, J. Hazard. Mater. 192 (2011)
568–575.
[25] A. Giwa, S. Daer, I. Ahmed, P.R. Marpu, S.W. Hasan, Experimental investigation
and artificial neural networks ANNs modeling of electrically-enhanced
membrane bioreactor for wastewater treatment, J. Water Process Eng. 11
(2016) 88–97.
[26] S. Geissler, T. Wintgens, T. Melin, K. Vossenkaul, C. Kullmann, Modelling
approaches for filtration processes with novel submerged capillary modules
in membrane bioreactors for wastewater treatment, Desalination 178 (2005)
125–134.
58 F. Schmitt et al. / Biochemical Engineering Journal 133 (2018) 47–58
[27] S.A. Mirbagheri, M. Bagheri, Z. Bagheri, A.M. Kamarkhani, Evaluation and
prediction of membrane fouling in a submerged membrane bioreactor with
simultaneous upward and downward aeration using artificial neural
network-genetic algorithm, Process Saf. Environ. Prot. 96 (2015) 111–124.
[28] M. Dornier, M. Decloux, G. Trystram, A. Lebert, Dynamic modeling of crossflow
microfiltration using neural networks, J. Membr. Sci. 98 (1995) 263–273.
[29] S. Chellam, Artificial neural network model for transient crossflow
microfiltration of polydispersed suspensions, J. Membr. Sci. 258 (2005) 35–42.
[30] Q.-F. Liu, S.-H. Kim, Evaluation of membrane fouling models based on
bench-scale experiments: a comparison between constant flowrate blocking
laws and artificial neural network (ANNs) model, J. Membr. Sci. 310 (2008)
393–401.
[31] Q.-F. Liu, S.-H. Kim, S. Lee, Prediction of microfiltration membrane fouling
using artificial neural network models, Sep. Purif. Technol. 70 (2009) 96–102.
[32] B.K. Nandi, A. Moparthi, R. Uppaluri, M.K. Purkait, Treatment of oily
wastewater using low cost ceramic membrane: comparative assessment of
pore blocking and artificial neural network models, Chem. Eng. Res. Des. 88
(2010) 881–892.
[33] F. Schmitt, K.-U. Do, Prediction of membrane fouling using artificial neural
networks for wastewater treated by membrane bioreactor technologies:
bottlenecks and possibilities, Environ. Sci. Pollut. Res. 24 (2017) 22885–22913.
[34] O. Giustolisi, D. Laucelli, Improving generalization of artificial neural
networks in rainfall–runoff modelling, Hydrol. Sci. J. 50 (2005) 439–457.
[35] A.P. Piotrowski, J.J. Napiorkowski, A comparison of methods to avoid
overfitting in neural networks training in the case of catchment runoff
modelling, J. Hydrol. 476 (2013) 97–111.
[36] APHA, Standard Methods for the Examination of Water and Wastewater, 21st
ed., American Water Works Association, Water Pollution and Control
Federation, Washington DC, USA, 2005.
[37] G.D. Garson, Interpreting neural-network connection weights, AI Expert 6
(1991) 46–51.
[38] J.D. Olden, D.A. Jackson, Illuminating the black box: a randomization approach
for understanding variable contributions in artificial neural networks, Ecol.
Modell. 154 (2002) 135–150.