Vous êtes sur la page 1sur 14

Arabian Journal for Science and Engineering

https://doi.org/10.1007/s13369-018-3484-8

RESEARCH ARTICLE - CHEMICAL ENGINEERING

A Model Based on Bootstrapped Neural Networks for Modeling the


Removal of Organic Compounds by Nanofiltration and Reverse
Osmosis Membranes
Yamina Ammi1,2 · Latifa Khaouane1 · Salah Hanini1

Received: 19 October 2017 / Accepted: 25 July 2018


© King Fahd University of Petroleum & Minerals 2018

Abstract
The present paper illustrates the use of single neural networks (SNN) and bootstrap aggregated neural networks (BANN)
for modeling the removal of organic compounds by nanofiltration and reverse osmosis. A set of 278 data points was used to
build the SNN and BANN. Bootstrap aggregated neural networks are used to enhance the accuracy and robustness of neural
network models built from a limited amount of training data. The training dataset is re-sampled using bootstrap re-sampling
with replacement to form several sets of training data. For each set of training data, a neural network model is developed.
The individual neural networks are then combined together to form a bootstrap aggregated neural network. Experimental
removals were compared against calculated removals and excellent R correlation coefficients were found (0.9890, 0.9836,
and 0.9841) for the training, test, and total dataset, respectively. The performance of the models (INN, BANN, and SNN) is
shown that models built from BANN are more accurate and robust than those built from individual neural networks (INN)
single neural networks (SNN).

Keywords Bootstrap · Neural networks · Modeling · Removal · Organic compounds · Membranes

1 Introduction actions [4]. An important aspect to deal with the problem


has been the identification of compound physicochemical
Augmentation of water sources using reclaimed water is a properties and membrane characteristics to explain transport
developing trend in the world due to the shortage of clean and removal of micropollutants by different mechanisms,
water sources worldwide. On the other hand, the application explicitly size/steric exclusion, hydrophobic adsorption and
of this water supply has a major concern, which is the occur- partitioning, and electrostatic repulsion [5,6].
rence of trace emerging contaminants. To date, many studies The literature contains a number of articles that have pro-
have focused on the application of nanofiltration (NF) and posed a mechanistic understanding of the interaction between
reverse osmosis (RO) membranes to remove trace emerg- membranes and organic compounds; others have tried to
ing contaminants, and the reported results vary greatly in a apply fitting parameter models to model removal. However,
different water matrix [1]. there have been few models to “predict” the removal of com-
Fewer studies have examined micropollutant removal as a pounds. In recent years, there have been various attempts
function of the interactions between organic compounds and to advance the use of multiple linear regressions “MLR”
membrane surfaces [2,3]. In general, three major interactions as a viable approach to developing data-driven models to
primarily affecting solute–membrane removal can be pointed describe the retention of organic compounds by nanofiltration
out as follows: steric hindrance (sieving effect), electrostatic and reverse osmosis membranes [7–12]. Several authors have
repulsion (charge effect), and hydrophobic/adsorptive inter- investigated the use of single neural network models (SNN)
for modeling the retention of organic compounds by NF/RO
B Yamina Ammi [5,7,9,13]. Flyborg et al. [14] was used a partial least squares
ammi.yamina@yahoo.fr
model (PLS) for predicting rejection of trace organic com-
1 Laboratory of Biomaterials and Transport Phenomena pounds by nanofiltration using treated wastewater as feed.
(LBMPT), University of Médéa, Médéa, Algeria Artificial neural networks have been widely used as a
2 University Center Ahmed Zabana Relizane, Relizane, Algeria statistical modeling tool to approximate complex functions,

123
Arabian Journal for Science and Engineering

particularly nonlinear ones between system inputs and out- gated neural networks (BANN) for modeling the removal
puts. A typical neural network comprises of an input layer of organic compounds by nanofiltration and reverse osmo-
with multiple inputs, an output layer with at least one output, sis membranes, taking into account three porosity indicators
and at least one hidden layer with multiply hidden neurons, (molecular weight cutoff “MWCO,” salt rejection “MgSO4 ,”
which are all associated by weights and biases [15]. The and salt rejection “NaCl”). The bootstrap aggregated neu-
application of ANNs in pollutant removal has been gaining ral network “BANN” was compared with individual neural
increasing attention. Nevertheless, ANN has also been crit- networks and single neural network “SNN.” This paper
icized for its “black box” nature such that the linearity or is organized as follows: Sect. 2 introduces materials and
quadratic dependence of the transfer equations may not be methods. Section 3 presents modeling procedure with sig-
well understood. In addition, the computational burden and nal neural networks and modeling procedure with bootstrap
the overfitting issue have also been identified [16]. aggregated neural networks. Results and discussion are pre-
The accuracy and robustness of neural network models are sented in Sect. 4. Section 5 concludes the work.
largely affected by the available training data and the train-
ing process. When the total of training data is limited, the
neural network would tend to over-fit noise in the training 2 Materials and Methods
data and exhibit significant generalization errors. An attrac-
tive approach to improving neural network model robustness The database (DB) included 278 removals of 23 organic com-
is to develop a set of neural network models and then com- pounds used in this study available in the literature [8,9,29–
bine them. Combining multiple neural network models have 33]. The selection of the input and output variables was based
been investigated by several researchers [17]. on interactions between organic compound properties, mem-
To date, there is a shortage of studies that present a brane characteristics, and filtration operating conditions for
model to predict the removal performance of trace organic the removal of organic compounds by NF/RO membranes.
compounds by NF/RO membranes using various poros- These solute–membrane interactions are determined by
ity indicators, two types of membranes (nanofiltration and organic compound solute properties, membrane properties,
reverse osmosis), and nature of organic compounds (neutral and operating conditions [4–6]. The inputs considered in this
and ionic organics). Both [5–14] have presented models that work are: compound properties describing hydrophobicity
give good predictions for the removal organic compounds (log Kow , logD), polarity (dipole moment), and size (molec-
by NF/RO membranes. However, most of these models have ular weight “Mw,” effective diameter of organic compound
been limited. in water “dc ,” molecular length, molecular depth, molecular
A limitation of single neural network models (SNN), width, molecular equivalent width “eqwidth”); membrane
multiple linear regressions (MLR), and partial least squares characteristics (molecular weight cutoff “MWCO,” magne-
model (PLS) for modeling the retention of organic com- sium sulfate salt rejection “SR (MgSO4 ),” sodium chloride
pounds by NF/RO membranes is that they can lack gener- salt rejection “SR (NaCl),” surface membrane charge “zeta
alization when applied to unseen data, the trained neural potential”; membrane hydrophobicity “contact angle”); and
network gives good performance on the training data but operating conditions (pH, pressure, temperature and recov-
gives an unsatisfactory performance on unseen data which are ery).
not used in the training process. Over the past years, many The molecular weight (MW ) and logD values were ob-
techniques have been developed to improve generalization tained through the online program (http://www.chemspider.
capability, such as bootstrap aggregated neural networks. com). Kowwin, epa (http://www.epa.gov/oppt/exposure/
A number of scientific researchers have attempted to pubs/epissuite.htm) was used for computing logKow . Dipole
use the bootstrap aggregated neural networks (BANN) in moment was calculated by Chem3D Ultra. The ChemAxon
different procedures: for optimization of electrical power software (http://www.chemicalize.org) was used to generate
in energy-deficient scenarios [18], for modeling the poly- length, width, and depth.
merization process and fermentation process [19–23], for The values of equivalent molecular width “eqwidth” were
optimization of a fed-batch bioreactor [24], for optimization calculated by the following equation [34].
of atmospheric crude distillation system operations [25], for
hierarchical prediction of monthly runoff in upper Damodar 
eqwidth = molecular width × molecular depth. (1)
valley catchment [26], for melt index prediction [27], for
modeling of a post-combustion CO2 capture process [28],
and for modeling the retention of organic compounds by The effective diameter “dc ” of organic compound in water
nanofiltration and reverse osmosis membranes [6]. is defined as follows [31]:
To the best of our knowledge, no studies have been
reported in the literature that has used the bootstrap aggre- dc = 0.065Mw 0.438 . (2)

123
Arabian Journal for Science and Engineering

Fig. 1 Procedure of the design and optimization of the architecture neural network

3 Modeling Procedure determine its predictive performance. Table 2 shows the error
values obtained for the removal under the influence of the
3.1 Modeling with Single Neural Networks division of the database. It has been observed that the third
division represents the best result.
The neural network is a mathematical model composed of
computing units (neural units) and linkage (dendrite), which 3.2 Modeling with Bootstrap Aggregated Neural
has a similar construction with the human brain. In numerous Networks (BANN)
neural networks, back-propagation (BP) neural network has
been used throughout this paper with forecasting horizon An attractive approach to improving neural network model
and supervised learning for several significant advantages: robustness is to develop a set of neural network models
nonlinear mapping capability; self-learning and adaptive and then combine them. Combining multiple neural network
capabilities; generalization ability; fault tolerance [35]. models have been investigated by several researchers [6,17].
Characteristically, it has three layers including the input To build bootstrap aggregated neural network models, the
layer, hiding layer, and output layer. The layers are made training data were re-sampled using bootstrap re-sampling
up with compute units and linked by transfer functions. The with replacement [37] to form 30 sets of training.
NN model functions as follows: Each neuron receives sig- A diagram of the bootstrap aggregated neural networks
nals from the neurons in the previous layer, and every one is shown in Fig. 2, where several individual neural network
of those signals is multiplied by a separate weight value. models are developed to model the same relationship and are
The weighted signals are summed and passed to next layer combined together using the following equation:
through a transfer function. Therefore, the knowledge of the n
NN model is stored in the transfer function, weight matrix, i=1 yi
y= , (3)
and threshold of NN model [36]. n
The procedure based on the design and optimization of the
architecture of single neural network “SNN” and individual where yi is the output of the individual neural networks
neural networks (INN) is described in Fig 1. “INN,” y represents the output of the BANN, and n is the
The statistical analysis of inputs and output (minimum number of INN models.
“min,” mean, maximum “max,” and standard deviations For each model (SNN and BANN), a neural network was
“STD”) data is given in Table 1. developed. Each NN contains three layers of neurons or
We studied the effect of the division of the database on nodes: one input layer with 15 neurons for NNs in the input
the optimization of the neural model; for this, the database layer, one hidden layer with a number of active neurons opti-
was divided into two subbases: a training base that allows mized during training, and one output layer with one unit
optimization of the NN architecture and a test base used to that generated the estimated value of removal. The number
of hidden neurons was varied from 3 to 25 neurons. The tan-

123
Arabian Journal for Science and Engineering

Table 1 Statistical analysis of


Min. Mean Max. SD
inputs and output
Inputs
dc (g/mol) 0.5280 0.7819 0.9409 0.1060
LogD − 2.8600 1.0846 3.3900 1.5039
Dipole moment (Debye) 0.0000 4.5636 10.9100 3.7806
Length (nm) 0.6310 1.3247 1.7030 0.3013
Eqwidth (nm) 0.2397 0.3791 0.4753 0.0600
MWCO (Dalton) 100.0000 147.5719 225.0000 52.3883
SR (MgSO4 ) 0.1000 0.8756 0.9970 0.2772
SR (NaCl) 0.1210 0.7443 0.9960 0.3115
Zeta potential (mv) − 41.3000 − 26.6204 3.1000 11.2961
Contact angle (◦ ) 20.1000 41.4421 65.0000 13.4837
pH 4.0000 6.6650 8.0100 1.0040
Pressure (KPa) 240.0000 1138.3827 4100.0000 615.5188
Recovery (%) 0.0400 25.0292 92.2600 28.9102
Temperature (◦ C) 20.0000 22.5000 25.0000 2.2481
Output
Removal (%) 1.0000 87.1323 100.0000 23.5705

Table 2 Division of the


Database Percentage Errors
database
EAM R

Division 1
Training phase: 111 data points 60 0.0003 0.9910
Test phase: 167 data points 40 0.0030 0.9609
Division 2
Training phase: 83 data points 70 0.0014 0.9703
Test phase:195 data points 30 0.0023 0.9714
Division 3
Training phase: 55 data points 80 0.0006 0.9869
Test phase: 223 data points 20 0.0018 0.9839

layer. The pure-linear (purelin) transfer function was used in


the output layer. Each neural network was trained using the
BFGS quasi-Newton (trainbfg).
The bootstrap aggregated neural networks output is taken
as the average of the individual neural network outputs.
The application of NNs modeling of the removal of organic
compounds by NF/RO membranes was performed using
STATISTICA.

4 Results and Discussion

4.1 Performance Models


Fig. 2 Bootstrap aggregated neural networks

Table 3 shows the structures of the optimized individual


gent hyperbolic (tanh), the log sigmoid (logsig), the sin, and neural networks “INNs” models and single neural network
the exponential transfer functions were used in the hidden “SNN” model. It can be seen that these networks are not

123
Arabian Journal for Science and Engineering

Table 3 Structures of the optimized INNs and SNN models linear regression are, straightforwardly, obtained using the
INNs Hidden layer MATLAB function “postreg”. Figure 3 shows comparison
between experimental and calculated removals with agree-
Neurons number Activation function
ment vectors approaching the ideal [i.e., α = 1 (slope), β = 0
INN1 12 Tanh (intercept), R = 1 (correlation coefficient)] in the adjustment
INN2 22 Exponential of the profiles of the neural networks, for SNN ([α, β, R] =
INN3 16 Tanh [1.0052, − 0.3855, 0.9818] for training phase, [α, β, R] =
INN4 21 Exponential [1.0769, − 7.1965, 0.9804] for test phase, and [α, β, R] =
INN5 10 Exponential [1.0262, − 2.3813, 0.9810] for total) and for BANN ([α, β,
INN6 6 Tanh
R] = [1.0063, − 0.5524, 0.9890] for training phase, [α, β, R]
= [1.0752, − 7.2152, 0.9836] for test phase, and [α, β, R] =
INN7 25 Logistic
[1.0728, − 6.5418, 0.9841] for total).
INN8 18 Logistic
INN9 3 Tanh
INN10 6 Tanh
INN11 13 Exponential 4.2 Comparison Between Individual Neural
INN12 22 Logistic Networks
INN13 14 Tanh
INN14 13 Tanh Figure 4 shows the correlation coefficient “R,” mean abso-
INN15 12 Tanh
lute error (MAE), the root mean squared error (RMSE), the
standard error of prediction (SEP), and the model predictive
INN16 9 Tanh
error (MPE), of the individual neural networks on the total
INN17 14 Tanh
datasets. It can be seen that these networks are not consistent
INN18 9 Tanh
and give various performances. Equations of those parame-
INN19 18 Tanh
ters are given as follows:
INN20 25 Logistic
INN21 23 Exponential
INN22 10 Tanh
1  
n

INN23 18 Exponential MAE = yi,exp − yi,cal  , (4)
INN24 17 Exponential n
i=1

INN25 4 Tanh n  2
INN26 7 Tanh i=1 Yi,exp − Yi,cal
RMSE = , (5)
INN27 15 Tanh n
INN28 7 Tanh RMSE
SEP (%) = × 100, (6)
INN29 25 Exponential Ye
n  
INN30 10 Tanh 100   (yi,exp − yi,cal 
SNN 25 Tanh
MPE (%) =  , (7)
n y i,exp
i=1

consistent and give various structures. Nineteen neural net- where n is the total number of data, Yi,exp is the experimental
works were used the transfer function tangent hyperbolic value, Yi,cal represents the calculated value from the neural
(tanh), eight neural networks were used the transfer function network model, and Ye is the mean value of experimental
exponential, and four neural networks were used the trans- data.
fer function log sigmoid (logistic) in the hidden layer, while For comparison purpose, the bootstrap aggregated neural
neural networks did not use the function sin in the hidden network and ten “good” individual neural networks (neural
layer. Eighteen individual neural networks INN (1, 5, 6, 9, network Nos. 1, 2, 3, 5, 10, 13, 14, 20, 25, and 27) are pro-
10, 11, 13, 14, 15, 16, 17, 18, 22, 25, 26, 27, 28, and 30) have posed. Figure 4 shows that these ten individual networks give
adequate cache numbers (fewer than or equal to 14 neurons) the various and the best performances among the 30 individ-
to obtain a satisfactory approximation. ual networks. It is shown that the individual neural networks
The single neural network “SNN” and bootstrap aggre- give the best performances among the 30 individual networks
gated neural networks “BANN” models were developed with on the total. However, it gives the worst performance on the
the aim of predicting of the removal of organic compounds unseen datasets. This indicates the non-robust nature of indi-
by NF/RO membranes. The plot and the parameters of the vidual neural network models.

123
Arabian Journal for Science and Engineering

Fig. 3 Comparison between experimental and calculated removals: a SNN “training phase,” b SNN “test phase,” c SNN “total,” d BANN “training
phase,” e BANN “test phase,” f BANN “total phase”

123
Arabian Journal for Science and Engineering

(a)
0.995
0.99

R (total data base)


0.985
0.98
0.975
0.97
0.965
0.96
0.955
0.95
0.945

Individual neural networks (INNs)

(b)
MAE (total data base) (%)

4.5
4
3.5
3
2.5
2
1.5
1
0.5
0

Individual neural networks (INNs)

(c)
RMSE (total data base) (%)

Individual neural networks (INNs)

(d)
SEP (total data base) (%)

8
7
6
5
4
3
2
1
0

Individual neural networks (INNs)

(e)
30
MPE (total data base) (%)

25

20

15

10

Individual neural networks (INNs)

Fig. 4 Correlation coefficient and errors obtained for each individual neural networks: a R, b MAE, c RMSE, d SEP, e MPE

123
Arabian Journal for Science and Engineering

Table 4 Errors of INNs,


R MAE (%) RMSE (%) SEP (%) MPE (%)
BANN, and SNN models
INN1
Training phase 0.9902 1.6150 3.2905 3.7462 2.7515
Test phase 0.9830 3.4239 6.1772 7.4767 32.4247
Total 0.9871 1.9729 4.0292 4.6418 8.6223
INN2
Training phase 0.9912 1.3842 2.4305 2.6964 2.9008
Test phase 0.9827 3.7708 5.7644 6.9771 23.8671
Total 0.9876 1.8563 3.3634 3.7940 7.0488
INN3
Training phase 0.9894 1.4249 2.9478 2.9478 2.5180
Test phase 0.9724 3.3812 7.2258 8.7458 16.6716
Total 0.9824 1.8119 4.0003 4.5093 5.3182
INN4
Training phase 0.9887 1.5368 2.8361 3.1695 2.3851
Test phase 0.9559 4.7197 8.6053 10.4155 36.7488
Total 0.9769 2.1665 4.5938 5.2128 9.1837
INN5
Training phase 0.9947 0.9586 2.2410 2.5333 1.8997
Test phase 0.9757 3.8534 6.4385 7.7930 9.4242
Total 0.9889 1.5313 3.4971 4.0056 3.3883
INN6
Training phase 0.9907 1.2188 2.6505 2.9860 4.1785
Test phase 0.9246 6.4261 11.5663 13.9994 69.1321
Total 0.9664 2.2490 5.6659 6.4716 17.0290
INN7
Training phase 0.9828 2.6336 4.2497 4.8945 8.7053
Test phase 0.9625 5.2835 8.0031 9.6866 39.2476
Total 0.9770 3.1578 5.2114 6.0602 14.7478
INN8
Training phase 0.9918 1.4691 2.7757 3.1414 2.6565
Test phase 0.9726 4.5166 6.9602 8.4244 20.5937
Total 0.9858 2.0721 3.9705 4.5520 6.2052
INN9
Training phase 0.9741 2.5770 4.4503 4.9793 6.7392
Test phase 0.9632 4.9556 8.2894 10.0332 57.7921
Total 0.9696 3.0476 5.4297 6.1673 16.8396
INN10
Training phase 0.9933 1.2040 2.2302 2.4867 1.8011
Test phase 0.9532 4.9382 9.1166 11.0344 32.0988
Total 0.9787 1.9428 4.5203 5.1198 7.7952
INN11
Training phase 0.9814 2.6672 4.4112 5.0866 12.4721
Test phase 0.9666 4.8931 7.7814 9.4183 51.0547
Total 0.9768 3.1076 5.2525 6.1138 20.1054

123
Arabian Journal for Science and Engineering

Table 4 continued
R MAE (%) RMSE (%) SEP (%) MPE (%)

INN12
Training phase 0.9853 1.9867 3.4646 3.8928 4.7845
Test phase 0.9708 3.9593 7.0679 8.5548 45.0234
Total 0.9806 2.3769 4.4172 5.0346 12.7455
INN13
Training phase 0.9923 1.2201 2.8313 3.2393 1.9856
Test phase 0.9849 3.0242 5.0731 6.1403 17.8128
Total 0.9902 1.5770 3.3944 3.9261 5.1168
INN14
Training phase 0.9968 0.9745 1.8754 2.1596 2.6465
Test phase 0.9771 3.8190 6.8093 8.2418 23.8640
Total 0.9905 1.5373 3.4633 4.0269 6.8442
INN15
Training phase 0.9810 2.4220 4.1846 4.7527 7.0948
Test phase 0.9544 5.1043 8.7760 10.6221 38.2285
Total 0.9729 2.9526 5.4114 6.2220 13.2543
INN16
Training phase 0.9949 1.0134 2.0548 2.2943 1.6643
Test phase 0.9616 4.3900 8.1002 9.8042 17.3213
Total 0.9839 4.4466 4.4478 4.4645 4.4633
INN17
Training phase 0.9911 1.8478 3.3562 3.8906 4.9282
Test phase 0.9752 3.8568 6.7235 8.1379 10.6698
Total 0.9869 2.2453 4.2402 4.9567 6.0641
INN18
Training phase 0.9906 1.9280 3.5846 4.2287 8.0919
Test phase 0.9726 3.8482 7.1170 8.6141 31.6707
Total 0.9859 2.3079 4.5087 5.3456 12.7567
INN19
Training phase 0.9914 1.1214 2.6359 2.9602 2.2119
Test phase 0.9694 4.1692 7.3046 8.8413 43.2602
Total 0.9838 1.7244 4.0162 4.5756 10.3329
INN20
Training phase 0.9956 1.0438 2.2444 2.5721 2.7518
Test phase 0.9859 3.5974 5.5803 6.7542 31.0320
Total 0.9922 1.5490 3.1940 3.6993 8.3468
INN21
Training phase 0.9842 1.5961 3.3874 3.7702 2.6636
Test phase 0.9649 4.8219 7.6595 9.2707 43.1403
Total 0.9776 2.2343 4.5619 5.1595 10.6716
INN22
Training phase 0.9934 1.5696 2.8184 3.2375 3.2635
Test phase 0.9835 3.9377 6.4042 7.7514 33.4930
Total 0.9894 2.0381 3.8060 4.4166 9.2442

123
Arabian Journal for Science and Engineering

Table 4 continued
R MAE (%) RMSE (%) SEP (%) MPE (%)

INN23
Training phase 0.9846 1.8932 3.7093 4.2540 3.4191
Test phase 0.9262 5.2326 11.5705 14.0045 30.3533
Total 0.9647 2.5539 6.1256 7.0989 8.7478
INN24
Training phase 0.9887 1.4679 3.3008 3.7225 6.1056
Test phase 0.9702 5.0123 7.5959 9.1938 27.5142
Total 0.9823 2.1691 4.4894 5.1323 10.3411
INN25
Training phase 0.9920 1.5261 2.5770 2.8832 3.3843
Test phase 0.9657 3.8383 7.6362 9.2425 22.6801
Total 0.9833 1.9835 4.1065 4.6643 7.2018
INN26
Training phase 0.9915 1.4374 2.8109 3.1544 3.1914
Test phase 0.9717 4.5813 7.8448 9.4950 25.1046
Total 0.9837 2.0594 4.3027 4.8991 7.5268
INN27
Training phase 0.9970 0.88744 1.5601 1.7488 3.3482
Test phase 0.9714 3.9926 7.3030 8.8393 23.8465
Total 0.9879 1.5017 3.5361 4.0226 7.4036
INN28
Training phase 0.9792 2.3207 4.4276 5.0155 7.5122
Test phase 0.9512 5.0610 9.2274 11.1685 23.0072
Total 0.9705 2.8629 5.7070 6.5479 10.5777
INN29
Training phase 0.9809 2.6320 4.4252 5.0610 8.8290
Test phase 0.9523 6.0490 9.5103 11.5109 54.5873
Total 0.9714 3.3680 5.9097 6.8533 18.2672
INN30
Training phase 0.9820 2.1329 3.9043 4.3721 9.0957
Test phase 0.9443 5.7174 11.7376 14.2067 101.3501
Total 0.9630 2.8421 6.2837 7.1423 27.3475
BANN (Staking 30 nets)
Training phase 0.9890 0.4475 0.5794 0.6560 0.5088
Test phase 0.9836 3.1789 5.7007 6.9000 11.6995
Total 0.9841 0.9878 2.5882 2.9684 2.7228
SNN
Training phase 0.9818 2.2573 4.1346 4.6854 7.7226
Test phase 0.9804 3.3441 6.1491 7.4426 20.2211
Total 0.9810 2.4723 4.6037 5.2835 10.1954

4.3 Comparison Between INNs, BANN, and SNN sistent. A network having small errors in the training dataset
may have quite large errors in the test dataset. The minimum
Table 4 shows the errors of the 30 individual neural networks root mean squared error (RMSE) of the individual neural
“INNs” models, the bootstrap aggregated neural networks network (INN20 ) on the training data, test data, and total
“BANN,” and the signal neural networks “SNN” on training datasets is 2.2444, 5.5803, and 3.1940%, respectively. The
data, test data, and total datasets. It can be observed that the RMSE from the aggregated neural networks on the train-
performances of these networks on the total sets are not con- ing data, test data, and total datasets is 0.5794, 5.7007, and

123
Arabian Journal for Science and Engineering

2.5882%, respectively. Thus, the model (BANN) accuracy (a) 7.7226


is significantly improved by combining multiple imperfect 8

models. 7
In order to establish the developed BANNs models as 6
4.6854
a plausible alternative to the SNN models, a comparison 5
4.1346
between the two approaches was made in terms of the mean 4
absolute error (MAE), the root mean squared error (RMSE), 3
the Standard error of prediction (SEP), and the model predic- 2
2.2573
tive error (MPE). Figure 5 shows the comparison between the 1
BANN and the signal neural networks models. Comparison 0.656 0.5088
0 0.5794
of training, test, and total results between the bootstrap aggre- 0.4475
gated neural networks models and the signal neural network MPE (%)
SEP (%)
models has shown the superiority of the bootstrap aggregated RMSE
MAE
neural network model. This indicates the non-robust nature BANN SNN
of the single neural network model. The BANN model has
a higher precision and can describe the removal of organic (b)
compounds by nanofiltration and reverse osmosis more accu- 25
20.2211
rately compared with the single neural networks model. It is
shown that such bootstrap aggregated neural networks model 20

can give more accurate and robust predictions than the single 15
neural networks model when applied to unseen data. 11.6995
10 7.4426
4.4 Comparison with Other Models 6.1491
6.9
5
3.3441 5.7007
For a significance of the problem encountered during the sep- 0 3.1789
aration of organic compounds by nanofiltration and reverse
osmosis membranes, models have been developed to study MPE (%)
SEP (%)
RMSE
the performance of these processes. Among these models, we MAE
meet single neural network models (SNN), multiple linear BANN SNN
regressions (MLR), and bootstrap aggregated neural net-
works (BANN). But there are a small number of works cited (c)
in the literature relating to work on modeling the retention of 12 10.1954
organic compounds by NF/RO membranes during the years
10
2008–2017. Table 5 lists the only notable examples of studies
predicting the removal of organic compounds during NF/RO 8
by the SNN, MLR, and BANN models published over the 5.2835
6
years. This table shows that our model is the only model 4.6037
that brings together the three porosity indicators (molecular 4
2.4723 2.9684 2.7228
weight cutoff “MWCO,” salt rejection “MgSO4 ,” and salt 2 2.5882
rejection “NaCl”); it is applicable for two types of mem-
0
branes nanofiltration and reverse osmosis and for neutral and 0.9878
ionic organics compounds. It shows that the errors of our MPE (%)
SEP (%)
model are lower than those obtained by other models. MAE
RMSE
In our current study, we used the bootstrap aggregated BANN SNN
neural networks (BANN) for modeling the removal of
organic compounds by nanofiltration and reverse osmosis Fig. 5 MAE, RMSE, PSE, and MPE of bootstrap aggregated neural
networks “BANN” and signal neural networks “SNN”: a training data,
membranes, taking into account three porosity indicators b test data, and c total data
(molecular weight cutoff “MWCO,” salt rejection “MgSO4 ,”
and salt rejection “NaCl”), but in our earlier study we were
using only one porosity indicators (salt rejection “NaCl”). than a BANN model with only porosity indicators. Our model
Our BANN model with tree porosity indicators gives more is more accurate and robust than those models cited in the
accurate and robust prediction of the removal of organic com- literature for modeling the removal of organic compounds by
pounds by nanofiltration and reverse osmosis membranes nanofiltration and reverse osmosis membranes.

123
123
Table 5 Overview of various works on models for predicting the retention of organic compounds by NF/RO
Name of model Modeling method Porosity indicators Membrane type Data base size Number and R MAE (%) RMSE (%) SEP (%)
nature of organic
compounds

Libotean et al. [7] SNN – RO 144 50 Neutral 0.9899 6.6000 – –


Yangali-Quintanilla et al. [8] MLR – NF–RO 64 14 Neutral and ionic 0.7924 – 5.5800 –
MLR – NF–RO 64 14 Neutral and ionic 0.6595 – 1.4500 –
MLR – NF–RO 64 14 Neutral and ionic 0.9757 – 10.1400 –
MLR – NF–RO 64 14 Neutral and ionic 0.9316 – 12.5600 –
Yangali-Quintanilla et al. [9] SNN SR(MgSO4 ) NF–RO 161 50 Neutral 0.9848 – 5.0000 –
Yangali-Quintanilla et al. [10] MLR MWCO NF 106 66 Neutral and ionic 0.8660 – – –
MLR MLR NF 106 66 Neutral and ionic 0.9195 – – –
Arash and Christopher [13] SNN – NF 255 67 Neutral 0.9273 – – –
Sadmani et al. [11] MLR – NF 51 17 Neutral and ionic 0.8916 – – 9.1740
Sadmani et al. [12] MLR – NF 51 17 Neutral and ionic 0.8814 – – 13.6060
Ammi et al. [5] SNN MWCO NF–RO 965 82 Neutral and ionic 0.9128 7.9138 11.2430 15.0730
SNN SR (MgSO4 ) NF–RO 701 81 Neutral and ionic 0.9419 6.6125 8.2047 12.1087
SNN MWCO and SR NF–RO 701 81 Neutral and ionic 0.9527 6.0995 9.0742 10.9485
(MgSO4 )
BANN SR (NaCl) NF–RO 436 42 Neutral and ionic 0.9704 1.5164 2.9757 3.5759
Khaouane et al. [6] BAMLR SR (NaCl) NF–RO 436 42 Neutral and ionic 0.8156 4.7350 7.7058 9.2601
Flyborg et al. [14] PLS – NF 156 32 0.9633 – – –
Our Work BANN MWCO, NF/RO 278 23 Neutral and ionic 0.9836 0.9878 2.5882 2.9684
SR (MgSO4 ),
and SR (NaCl)
Arabian Journal for Science and Engineering
Arabian Journal for Science and Engineering

5 Conclusion 5. Ammi, Y.; Khaouane, L.; Hanini, S.: Prediction of the rejection of
organic compounds (neutral and ionic) by nanofiltration and reverse
osmosis membranes using neural networks. Korean J. Chem. Eng.
The present study was aimed at modeling the removal of 32(11), 2300–2310 (2015). https://doi.org/10.1007/s11814-015-
organic compounds by nanofiltration and reverse osmosis by 0086-y
bootstrap aggregated neural networks (BANN). Comparison 6. Khaouane, L.; Ammi, Y.; Hanini, S.: Modeling the retention of
between BANN and single neural network (SNN) revealed organic compounds by nanofiltration and reverse osmosis mem-
branes using bootstrap aggregated neural networks. Arab. J. Sci.
that BANN model gave the best performance: The root mean Eng. 42(4), 1443–1453 (2017). https://doi.org/10.1007/s13369-
squared errors for the total dataset were 2.5882 for BANN 016-2320-2
and 4.6037 for SNN. The BANN model has a higher preci- 7. Libotean, D.; Giralt, J.; Rallo, R.; Cohen, Y.; Giralt, F.; Ridg-
sion and can describe the removal of organic compounds by way, H.F.; Rodriguez, G.; Phipps, D.: Organic compounds passage
through RO membranes. J. Membr. Sci. 313(1–2), 23–43 (2008).
nanofiltration and reverse osmosis more accurately compared
https://doi.org/10.1016/j.memsci.2007.11.052
with the single neural network model. It shows that such a 8. Yangali-Quintanilla, V.; Kennedy, M.; Amy, G.; Kim, T.U.: Mod-
bootstrap aggregated neural networks model can give more eling of RO/NF membrane rejections of PhACs and organic
accurate and robust predictions than a single neural networks compounds: a statistical analysis. Drink. Water Eng. Sci. 1(1), 7–15
(2008)
model when applied to unseen data. 9. Yangali-Quintanilla, V.; Verliefde, A.; Kim, T.U.; Sadmani, A.;
Also, the BANN model with the effective diameter of Kennedy, M.; Amy, G.: Artificial neural network models based
organic compound in water (dc) and with three porosity on QSAR for predicting rejection of neutral organic compounds
indicators (MWCO, SR“MgSO4,” SR“NaCl”) gives more by polyamide nanofiltration and reverse osmosis membranes. J.
Membr. Sci. 342(1–2), 251–262 (2009). https://doi.org/10.1016/j.
accurate and robust prediction the removal of organic com- memsci.2009.06.048
pounds by nanofiltration and reverse osmosis membranes 10. Yangali-Quintanilla, V.; Sadmani, A.; McConville, M.; Kennedy,
than a BANN model with molecular weight (Mw) and with M.; Amy, G.: A QSAR model for predicting rejection of emerging
only porosity indicators (SR“NaCl”). contaminants (pharmaceuticals, endocrine disruptors) by nanofil-
tration membranes. Water Res. 44(2), 373–384 (2010). https://doi.
Our model BANN is more precise and correct than those org/10.1016/j.watres.2009.06.054
models cited in the literature for modeling the removal of 11. Sadmani, A.H.M.A.; Andrews, R.C.; Bagley, D.M.: Impact of
organic compounds by nanofiltration and reverse osmosis natural water colloids and cations on the rejection of pharmaceuti-
membranes. cally active and endocrine disrupting compounds by nanofiltration.
J. Membr. Sci. 450, 272–281 (2014). https://doi.org/10.1016/j.
memsci.2013.09.017
Acknowledgements Authors gratefully acknowledge the team of Lab-
12. Sadmani, A.H.M.A.; Andrews, R.C.; Bagley, D.M.: Nanofiltration
oratory of Biomaterials and Transport Phenomena, the University of
of pharmaceutically active and endocrine disrupting compounds
Medea, and University Center of Relizane for their help throughout
as a function of compound interactions with DOM fractions and
this project. The authors also thank the anonymous reviewers for their
cations in natural water. Sep. Purif. Technol. 122, 462–471 (2014).
constructive comments which helped to improve the quality and pre-
https://doi.org/10.1016/j.seppur.2013.12.003
sentation of this paper.
13. Arash, S.; Christopher, B.: Application of quantitative structure-
property relationships (QSPRs) to predict the rejection of organic
solutes by nanofiltration. Sep. Purif. Technol. 118, 627–638 (2013)
14. Flyborg, L.; Björlenius, B.; Ullner, M.; Persson, K.M.: A PLS
model for predicting rejection of trace organic compounds by
References nanofiltration using treated wastewater as feed. Sep. Purif. Tech-
nol. 174, 212–221 (2017). https://doi.org/10.1016/j.seppur.2016.
1. Lin, Y.-L.: Effects of organic, biological and colloidal fouling on 10.029
the removal of pharmaceuticals and personal care products by 15. Lin, W.; Jing, L.; Zhu, Z.; Cai, Q.; Zhang, B.: Removal of heavy
nanofiltration and reverse osmosis membranes. J. Membr. Sci. 542, metals from mining wastewater by Micellar-Enhanced Ultrafiltra-
342–351 (2017). https://doi.org/10.1016/j.memsci.2017.08.023 tion (MEUF): experimental investigation and Monte Carlo-based
2. Verliefde, A.R.; Heijman, S.G.; Cornelissen, E.R.; Amy, G.; Van artificial neural network modeling. Water Air Soil Pollut. 228(6),
der Bruggen, B.; van Dijk, J.C.: Influence of electrostatic inter- 206 (2017). https://doi.org/10.1007/s11270-017-3386-5
actions on the rejection with NF and assessment of the removal 16. Elmolla, E.S.; Chaudhuri, M.; Eltoukhy, M.M.: The use of artificial
efficiency during NF/GAC treatment of pharmaceutically active neural network (ANN) for modeling of COD removal from antibi-
compounds in surface water. Water Res. 41(15), 3227–3240 (2007). otic aqueous solution by the Fenton process. J. Hazard. Mater.
https://doi.org/10.1016/j.watres.2007.05.022 179(1), 127–134 (2010). https://doi.org/10.1016/j.jhazmat.2010.
3. Verliefde, A.R.D.; Cornelissen, E.R.; Heijman, S.G.J.; Verberk, 02.068
J.Q.J.C.; Amy, G.L.; Van der Bruggen, B.; van Dijk, J.C.: The role 17. Zhang, J.: Inferential estimation of polymer quality using bootstrap
of electrostatic interactions on the rejection of organic solutes in aggregated neural networks. Neural Netw. 12, 927–938 (1999)
aqueous solutions with nanofiltration. J. Membr. Sci. 322(1), 52–66 18. Tahir, M.F.; Tehzeeb ul, H.; Saqib, M.A.: Optimal scheduling of
(2008). https://doi.org/10.1016/j.memsci.2008.05.022 electrical power in energy-deficient scenarios using artificial neural
4. Gur-Reznik, S.; Koren-Menashe, I.; Heller-Grossman, L.; Rufel, network and Bootstrap aggregating. Int. J. Electr. Power Energy
O.; Dosoretz, C.G.: Influence of seasonal and operating conditions Syst. 83, 49–57 (2016). https://doi.org/10.1016/j.ijepes.2016.03.
on the rejection of pharmaceutical active compounds by RO and 046
NF membranes. Desalination 277(1–3), 250–256 (2011). https:// 19. Zhang, J.: Developing robust non-linear models through bootstrap
doi.org/10.1016/j.desal.2011.04.029 aggregated neural networks. Neurocomputing 25, 93–113 (1999)

123
Arabian Journal for Science and Engineering

20. Zhang, J.: Batch-to-batch optimal control of a batch polymerisation 30. Dolar, D.; Vuković, A.; Ašperger, D.; Košutić, K.: Effect of water
process based on stacked neural network models. Chem. Eng. Sci. matrices on removal of veterinary pharmaceuticals by nanofiltra-
63, 1273–1281 (2008) tion and reverse osmosis membranes. J. Env. Sci. 23(8), 1299–1307
21. Zhang, J.; Feng, Y.; Al-Mahrouqi, M.H.: Reliable optimal control (2011). https://doi.org/10.1016/S1001-0742(10)60545-1
of a fed-batch fermentation process using ant colony optimization 31. Dolar, D.; Ignjatic Zokic, T.; Kosutic, K.; Asperger, D.; Mutavdzic
and bootstrap aggregated neural network models. In: Valadi, J., Pavlovic, D.: RO/NF membrane treatment of veterinary pharma-
Siarry, P. (eds.) Applications of Metaheuristics in Process Engineer- ceutical wastewater: comparison of results obtained on a laboratory
ing, pp. 183–200. Springer International Publishing, Cham (2014) and a pilot scale. Environ. Sci. Pollut. Res. Int. 19(4), 1033–1042
22. Zhang, J.; Xu, Q.J.Y.: Inferential estimation of polymer melt index (2012). https://doi.org/10.1007/s11356-012-0782-7
using sequentially trained bootstrap aggregated. Neural Netw. 32. Dolar, D.; Kosutic, K.; Asperger, D.: Influence of adsorption of
29(4), 442–448 (2006) pharmaceuticals onto RO/NF membranes on their removal from
23. Mohammed, K.-J.R.; Zhang, J.: Reliable optimisation control of a water. Water Air Soil Pollut Int. J. Environ. Pollut. 224(1), 1–13
reactive polymer composite moulding process using ant colony (2013)
optimisation and bootstrap aggregated neural networks. Neural 33. Dolar, D.; Košutić, K.; Periša, M.; Babić, S.: Photolysis of
Comput. Appl. 23(7), 1891–1898 (2013). https://doi.org/10.1007/ enrofloxacin and removal of its photodegradation products from
s00521-012-1273-y water by reverse osmosis and nanofiltration membranes. Sep. Purif.
24. Al-Mahrouqi, M.H.; Zhang, J.: Reliable optimal control of a fed- Technol. 115(13), 1–8 (2013). https://doi.org/10.1016/j.seppur.
batch bio-reactor using ant colony optimization and bootstrap 2013.04.042
aggregated neural networks. IFAC Proc. Vol. 41(2), 8407–8412 34. Santos, J.L.C.; de Beukelaar, P.; Vankelecom, I.F.J.; Velizarov, S.;
(2008). https://doi.org/10.3182/20080706-5-KR-1001.01421 Crespo, J.G.: Effect of solute geometry and orientation on the
25. Osuolale, F.N.; Zhang, J.: Multi-objective optimisation of atmo- rejection of uncharged compounds by nanofiltration. Sep. Purif.
spheric crude distillation system operations based on bootstrap Technol. 50(1), 122–131 (2006). https://doi.org/10.1016/j.seppur.
aggregated neural network models. In: Gernaey, K.V., Huusom, 2005.11.015
J.K., Gani, R. (eds.) Computer Aided Chemical Engineering, vol. 35. Mandal, S.; Sivaprasad, P.V.; Venugopal, S.; Murthy, K.P.N.; Raj,
37, pp. 671–676. Elsevier, New York (2015) B.: Artificial neural network modeling of composition-process-
26. Sharma, S.K.; Tiwari, K.N.: Bootstrap based artificial neural net- property correlations in austenitic stainless steels. Mater. Sci. Eng.
work (BANN) analysis for hierarchical prediction of monthly A 485(1), 571–580 (2008). https://doi.org/10.1016/j.msea.2007.
runoff in Upper Damodar Valley Catchment. J. Hydrol. 374(3), 08.019
209–222 (2009). https://doi.org/10.1016/j.jhydrol.2009.06.003 36. Liu, G.; Jia, L.; Kong, B.; Guan, K.; Zhang, H.: Artificial neu-
27. Zhang, Z.; Wang, T.; Liu, X.: Melt index prediction by aggregated ral network application to study quantitative relationship between
RBF neural networks trained with chaotic theory. Neurocomputing silicide and fracture toughness of Nb-Si alloys. Mater. Des. 129,
131, 368–376 (2014). https://doi.org/10.1016/j.neucom.2013.10. 210–218 (2017)
006 37. Efron, B.; Tibshirani, R.: An Introduction to Bootstrap. Chapman
28. Bai, Z.; Li, F.; Zhang, J.; Oko, E.; Wang, M.; Xiong, Z.; Huang, and Hall, London (1993)
D.: Modelling of a post-combustion CO2 capture process using
bootstrap aggregated extreme learning machines. In: Kravanja, Z.,
Bogataj, M. (eds.) Computer Aided Chemical Engineering, vol. 38,
pp. 2007–2012. Elsevier, New York (2016)
29. Tang, C.Y.; Kwon, Y.-N.; Leckie, J.O.: Effect of membrane chem-
istry and coating layer on physiochemical properties of thin film
composite polyamide RO and NF membranes: II. Membrane phys-
iochemical properties and their dependence on polyamide and
coating layers. Desalination 242(1–3), 168–182 (2009). https://doi.
org/10.1016/j.desal.2008.04.004

123

Vous aimerez peut-être aussi