Documenti di Didattica
Documenti di Professioni
Documenti di Cultura
a r t i c l e i n f o a b s t r a c t
Article history: This paper describes the effect of simultaneous upward and downward aeration on
Received 8 November 2014 the membrane fouling and process performances of a submerged membrane bioreac-
Received in revised form 23 March tor. Trans-membrane pressure (TMP) and membrane permeability (Perm) were simulated
2015 using multi-layer perceptron and radial basis function artificial neural networks (MLPANN
Accepted 25 March 2015 and RBFANN). Genetic algorithm (GA) was utilized in order to optimize the weights and
Available online 12 May 2015 thresholds of the models. The results indicated that the simultaneous aeration does not sig-
nificantly improve the removal efficiency of contaminants. The removal efficiencies of BOD,
Keywords: COD, total nitrogen, NH+
4 − N and TSS were 97.5%, 97%, 94.6%, 96% and 98%, respectively. It
Simultaneous aeration was observed that the TMP increases and the Perm decreases as operational time increases.
Hollow fiber membrane The TMP increasing rate (dTMP/dt) and the Perm decreasing rate (dPerm/dt) for the upward
Membrane fouling aeration were 2.13 and 2.66 times higher than that of simultaneous aeration, respectively.
Trans-membrane pressure The training procedures of TMP and Perm models were successful for both RBFANN and
Membrane permeability MLPANN. The train and test models by MLPANN and RBFANN showed an almost perfect
Artificial neural networks match between the experimental and the simulated values of TMP and Perm. It was illus-
trated that the GA-optimized ANN predicts TMP and permeability more accurately than a
network with a trial-and-error approach calibration.
© 2015 The Institution of Chemical Engineers. Published by Elsevier B.V. All rights reserved.
∗
Corresponding author. Tel.: +98 9181331137; fax: +98 2188770006.
E-mail address: bagherimajead@yahoo.com (M. Bagheri) .
http://dx.doi.org/10.1016/j.psep.2015.03.015
0957-5820/© 2015 The Institution of Chemical Engineers. Published by Elsevier B.V. All rights reserved.
112 Process Safety and Environmental Protection 9 6 ( 2 0 1 5 ) 111–124
Fig. 2 – Topological architectures of the neural networks used in this study: MLP (a), and RBF (b).
116 Process Safety and Environmental Protection 9 6 ( 2 0 1 5 ) 111–124
linear, threshold linear and bipolar linear) was iteratively mutation: A sudden change in a step of the solution’s feature
determined by developing several networks. Each network is called mutation (Enayatifar et al., 2014; John, 1992).
was trained until the network average RMSE was minimum Algorithm is started with a set of random solutions called
and R2 was equal to 1. Other parameters for network were population (Kurek and Ostfeld, 2013). Population initialization
chosen as the default values of the software (Pendashteh et al., is the process of generating initial parameter sets for the first
2011). population of the GA run. Two methods, namely random and
The performances of the ANN models were measured by heuristic are used to initiate the population. The commonly
R2 and RMSE between the predicted values of the network and used random method generates parameters randomly with-
the experimental values, which were calculated by Eqs. (3) and out any prior knowledge of the likely ‘optimum’ parameter set.
(4), respectively (Pendashteh et al., 2011). The heuristic method, on the other hand, requires some prior
knowledge of the likely ‘optimum’ parameter set, and there-
n (i)
2 fore, provides the ‘optimum’ solution faster (Liu et al., 2007;
1− i=1
y∗i − yp
Mendes et al., 2014). The number of variables, population size
R2 = n 2 (3)
y∗i − ȳ and generation gap in this study were 5, 30 and 1, respectively.
i=1
The initial set of strings will generally not provide an opti-
mal solution (Liu et al., 2007). Selection is the survival of
n 2
1
(i) the fittest within the GA. The selection process determines
RMSE = yp − y∗ i
(4)
n which chromosomes are preferred for generating the next
i=1
population, according to their fitness values in the current
population. The key notion in selection is to give a higher prior-
(i)
where ȳ is the average of y over the n data, and y∗i and yp are ity or preference to better individuals (Rani et al., 2013). There
the ith target and predicted responses, respectively. are different selection methods as stochastic uniform, remain-
In order to obtain convergence within a reasonable number der, uniform, shift linear, roulette wheel and tournament (Kılıç
of cycles, the input and output data should be normalized and et al., 2014). The tournament method is preferred in this study
scaled to the range of 0–1 by the following equation (Sahoo and and the selection probability is equal to 0.5. The tournament
Ray, 2006; Xi et al., 2011): method can be described by the following equation:
xi − xmin Fi
xni =
xmax − xmin
(5) i = Nk (7)
F
j=1 j
where xi is the initial value, xmax and xmin are the maximum
where i is the weight of ith individual within population.
and minimum of the initial values, and xni is the scaled value.
Moreover, the sum of the elective probabilities of all the
After the training and testing of the ANN, the output data
individuals within population is 1 as is determined by the
were scaled to the real-world values through the following
following equation:
equation:
Nk
xi = xni (xmax − xmin ) + xmin (6) i = 1 (8)
i=1
2.5. Genetic algorithm optimization procedure
The crossover operator is used to create new chromosomes
GAs are procedures based on the mimetic of mechanics of for the next generation by combining randomly two selected
natural selection and genetics. Theoretically developed by chromosomes from the current generation (Rani et al., 2013).
Holland (John, 1992), genetic algorithms emulate the biolo- Once the surviving individuals are determined, the population
gist evolutionary theory to solve optimization problems. This is completed with new individuals obtained through cross-
method can find optimal solutions in a large solution space over mechanisms performed on two parents randomly paired
by evaluating only a relatively small number of potential in the whole current population (Montastruc et al., 2004).
solutions. It works directly with the fitness of each solution There are several crossover methods available for reproduc-
instead of derivatives or other auxiliary characteristics used ing the next generation such as single pint, double point,
with traditional deterministic approaches. Also, GA searches multi point, uniform, random and arithmetic crossover. The
for the global optimum solution over the whole solution space, crossover procedure adopted here is the uniform permuta-
instead of focusing in a part or boundaries of solution space tion operation Single-point and multipoint crossover define
that may lead to a local, rather than global, optimal solutions crossover points between the first and last bit of two chromo-
(Kurek and Ostfeld, 2013). The main advantages of Gas are the somes to exchange the bits between them. Uniform crossover
possibility of solving every optimization problems where the generalizes this scheme to make every bit position a potential
solutions can be represented in chromosomes, the possibility crossover point. In uniform crossover, one offspring is con-
to solve problems with multiple solutions and with large num- structed by choosing every bit with a probability Pc from either
ber of variables, the absence of requirements regarding the parent (Rani et al., 2013). The crossover probability equal to 0.4
availability of mathematical knowledge of the optimization was chosen in this study.
problem (Mendes et al., 2014). Basically, a GA is categorized Mutation is used in support of crossover to ensure that
into four main steps: (1) Creating population: The numbers of the full range of allele is accessible for each gene. Mutation
the initial populations are generated in this step, (2) selection: is applied at a certain probability, Pm , to each gene of the
The solution for creating the offspring is chosen in this step, offspring, xi (t) to produce the mutated offspring Xi (t). The
(3) crossover: This section is dedicated to creating new solu- mutation probability, also referred to as the mutation rate, is
tions by considering the solutions from the selection step, (4) usually a small value, Pm ∈ [0,1], to ensure that good solutions
Process Safety and Environmental Protection 9 6 ( 2 0 1 5 ) 111–124 117
are not distorted too much. Given that each gene is mutated
at probability Pm , the probability that an individual will be
mutated, taking into account that the individual contains nx
genes, is given by the following equation (Ordóñez Galán et al.,
2011):
Fig. 6 – Train (a) and test (b) models by RBFANN, and train (c) and test (d) models by MLPANN for TMP.
Process Safety and Environmental Protection 9 6 ( 2 0 1 5 ) 111–124 119
3.3. Modeling of membrane fouling Then the autocorrelation functions of the simulated values
are compared to the observed values, and the performance of
The operational time, TSS, influent COD, SRT and MLSS were the one-step-ahead predictions are analyzed carefully. More-
the network inputs in order to model the TMP and mem- over, another criterion applied for judging the validity of the
brane permeability (Perm) using the RBFANN and MLPANN. model is the assessment of the goodness of fit according to
The RBFANN applied newrbe function to the input data and different available criteria. The RMSE values for train and test
the spread of radial basis function was considered equal to its (verification) models by MLPANN were 0.208 kPa and 0.312 kPa,
default value, namely 1. A large spread results in a smooth respectively. Moreover, the R2 values for train and test models
function approximation, but, by contrast, a large spread can were 0.97 and 0.98, respectively. The MLPANN simulated TMP
cause numerical problems (Pendashteh et al., 2011). The so precisely that the mean average error for train and test mod-
newrbe function selected 70% of normalized data to train els were 1% and 3%, respectively. The RMSE values for train
and 30% to test the RBFANN models (Çinar et al., 2006). The and test (verification) models by RBFANN were 0.352 kPa and
RBFANN was designed in a loop that applied newrbe to the 0.329 kPa, respectively. Furthermore, the R2 values for train
data for more than 100 times in order to minimize error. and test models were 0.97 and 0.98, respectively. The RBFANN
The optimal network was chosen on the basis of the mini- simulated TMP so precisely that the mean average error for
mum average error. The MLPANN applied newff function to train and test models did not exceed 3%.
the input data and, created a feed-forward back propagation The results indicated that the MLPANN has stronger
network. The newff function selected 70% of normalized data approximation and generalization ability than the RBFANN
to train, 15% to test and 15% to validate the MLPANN models with regard to our experimental data in the modeling of TMP.
(Çinar et al., 2006). The MLPANN was trained by different learn- It was observed that the precision of MLPANN and RBFANN
ing algorithms for maximum 1000 epochs. Nevertheless, the models is determined not only by the number of input vari-
Levenberg–Marquardt algorithm resulted in the most accu- ables of the network but also by correlation of these variables.
rate models for training and testing data after less than 40 The high correlation of the input data affected the precision
iterations. Based on our research, the optimal models were of the MLPANN models more than RBFANN models. On the
obtained with the hidden layer consisting of 8–12 neurons. other hand, the number of the input data affected the preci-
The best transfer function for the hidden layer was found sion of the RBFANN models more than MLPANN models. Fig. 7
to be hyperbolic tangent sigmoid (tansig) function while the shows the TMP models by MLPANN and RBFANN according to
best transfer function for the output layer was a linear one all experimental data for 60 days.
(purelin). The TMP was modeled separately by considering different
single variables as inputs of MLPANN and RBFANN in order
3.3.1. Trans-membrane pressure models to investigate the effect of each variable on the changes of
The training procedures for the TMP models were success- TMP. Similarly, separate models were performed in order to
ful for both RBFANN and MLPANN. The train and test models show the effects of joint input variables on the changes of TMP.
by MLPANN and RBFANN showed an almost perfect match These inputs were used to train the networks in groups of two,
between the experimental and the simulated values of TMP three and four variables. The results indicated that operational
(Fig. 6). It is obvious that a careful verification is required time among single input variables, and operational time and
for the proper use of models in practical applications. In the MLSS among groups of two variables significantly affected the
current research, the model is verified by evaluating its per- TMP models. Furthermore, operational time, MLSS and TSS
formance in propitiously simulating the statistical features of among groups of three variables, and operational time, MLSS,
the observed data. In particular, the observed high quantiles TSS and SRT between groups of four variables had the most
are compared with the simulated ones (Weiss et al., 2012). significant effects on the TMP models (Table 5).
120 Process Safety and Environmental Protection 9 6 ( 2 0 1 5 ) 111–124
Table 5 – Effects of different single and joint parameters on the trans-membrane pressure (TMP).
Input parameter no. Radial basis function (RBF) models Multi-layer perceptron (MLP) models Importance order
2 2
R RMSE (kPa) R RMSE (kPa)
It has been proved that the model fit statistics are not a each variable on the RBFANN and MLPANN models compared
good guide to how well a model will simulate and predict a with the other parameters was determined by its importance
time-series phenomenon. High values of R2 do not necessar- order. Table 5 shows the importance order of each single and
ily lead to a favorable model, although it can be a sign of a joint input variable for the TMP models. The variable with
propitious model. Additionally, a way to measure the predic- higher rank of importance showed not only an almost perfect
tive capability of a model is to test it on a set of data not used match between experimental and both MLPANN and RBFANN
in simulation process. This has been described in literature as models but also less RMSE and more R2 values. The results
test set and the data used for simulation is training set. As a of modeling demonstrated that the variation of the TMP is
result, to verify our models, a set of the TMP and membrane influenced by operational time, MLSS, SRT, TSS and influent
permeability data was used to examine the predictive ability of COD. The current research shows that the operational time
models. Furthermore, sensitivity analyses (Saltelli et al., 2008) and MLSS significantly affect the TMP predicting models as
were performed to determine the sensitivity of TMP and mem- they have been examined experimentally in earlier studies
brane permeability to changes of input variables. The effect of (Park et al., 2010).
Fig. 8 – Train (a) and test (b) models by RBFANN, and train (c) and test (d) models by MLPANN for Perm.
Process Safety and Environmental Protection 9 6 ( 2 0 1 5 ) 111–124 121
Table 6 – Effects of different single and joint parameters on the membrane permeability (Perm).
Input parameter no. Radial basis function (RBF) models Multi-layer perceptron (MLP) models Importance order
2 2
R RMSE (LMH/kPa) R RMSE (LMH/kPa)
Table 7 – Effect of applying GA on the precision of TMP and membrane permeability models.
Type of ANN model R2 RMSE (kPa) or (LMH/kPa) Percentage of error (%) Rank of model (accuracy)
Trans-membrane
pressure (TMP)
models
RBFANN 0.95 0.97 0.98 0.497 0.352 0.329 5 4
RBFANN-GA 0.98 0.98 0.98 0.267 0.251 0.283 2 2
MLPANN 0.98 0.99 0.98 0.293 0.208 0.312 3 3
MLPANN-GA 0.98 0.99 0.99 0.256 0.214 0.202 1 1
Membrane
permeability
(Perm) models
RBFANN 0.97 0.95 0.99 0.041 0.050 0.028 4 4
RBFANN-GA 0.98 0.98 0.99 0.026 0.019 0.029 3 2
MLPANN 0.98 0.99 0.99 0.037 0.016 0.031 3 3
MLPANN-GA 0.99 0.99 0.99 0.016 0.011 0.014 1 1
(Soleimani et al., 2013), and it can be seen that the esti- 4. Conclusions
mated TMP and permeability values, that were achieved by GA,
corresponding to optimum operating parameters are almost A vertically oriented hollow fiber SMBR with combined aera-
greater than those observed in experiments and also it can be tion/filtration was used to examine the effect of simultaneous
observed that some estimated TMP and permeability values, upward and downward aeration on the membrane fouling
that were achieved by GA, corresponding to optimum operat- characteristics and process performance of the system. More-
ing parameters are almost near to the minimum quantities over, TMP and Perm were simulated using GA based MLPANN
those observed in experiments. and RBFANN models. We concluded that there is no signif-
In the current research, the tournament selection method, icant difference in the removal efficiency of BOD, COD, TN,
a mutation method consisting of replacing a randomly NH+ 4 − N and TSS between upward and simultaneous upward
selected gene by its binary complement and uniform crossover and downward aeration systems in the municipal wastewa-
method were used as genetic operators. For the GA operators ter treatment. It was observed that the TMP increases and
investigated in this study, the selection probability was ini- Perm decreases with any increase in the operational time. The
tially set to 0.5; the crossover probability was set to 0.4; the TMP increasing rates (dTMP/dt) for the upward and simultane-
mutation probability was set to 0.1; the population size was ous aeration were 0.16333 and 0.07666, respectively while the
set to 30; the generation gap was set to 1; the maximum num- Perm decreasing rates (dPerm/dt) were 0.01333 and 0.00511,
ber of generation was set to 10 and the stopping criterion was respectively. We concluded that the (dTMP/dt) increasing rate
set to either exceed the maximum number of generations or and the (dPerm/dt) decreasing rate for the upward aeration
stop after 6 generations without improvement of fitness val- were 2.13 and 2.66 times higher than that of simultaneous
ues. Like any stochastic procedure to solve a given problem aeration, respectively. The training procedures for TMP and
(Montastruc et al., 2004), the GA has to be run 90 times in this Perm models were successful for both RBFANN and MLPANN.
study with different initial populations. It was observed that The train and test models by MLPANN and RBFANN showed
the fixed percentage of individuals that survive in the new gen- an almost perfect match between the experimental and the
eration and mutation rates do not have a significant influence simulated values of TMP and Perm. We concluded that the
on the method convergence which is confirmed by 60 runs of MLPANN has stronger approximation and generalization abil-
GA. ity than RBFANN with regard to our experimental data in the
Based on the result of this study, the optimum models are modeling of the TMP and Perm.
obtained with the hidden layer consisting of 10 neurons com-
pared with the results of previous studies, which varied from Acknowledgements
10 to 20 neurons (Badrnezhad and Mirza, 2014; Bagheri et al.,
2014). The residuals of the models attained by the RBFANN The authors are grateful to Ekbatan wastewater treatment
and the MLPANN when GA applied for train and test data plant for their technical and logistical assistance during this
sets showed an approximately normal distribution. A normal work which was supported by authors. And also, we wish to
distribution of variation results in a Gaussian curve (specific thank Ali Reza Jafari and Mohammad Reza Ghanbari for tech-
bell-shaped curve), with the highest point in the middle and nical help with modeling by artificial neural network.
smoothly curving symmetrical slopes on both sides of center
(Chamkalani et al., 2013). Gaussian curve reveals our results Appendix A. Supplementary data
are symmetrical and their axis round around zero (Ratkovich
et al., 2012). The effects of applying GA on the RBFANN and Supplementary data associated with this article can be
the MLPANN in the simulation of trans-membrane pressure found, in the online version, at http://dx.doi.org/10.1016/j.
and membrane permeability have been presented in Table 7. psep.2015.03.015.
The results showed that the precision and accuracy of all
ANN models increased when GA is applied to the models. It References
was illustrated that the GA-optimized ANN predicts TMP and
permeability more accurately than a network with a trial-and- Andrew, D., 2005. Standard Methods for the Examination of
error approach calibration. Water and Wastewater. APHA-AWWA-WEF, Washington, DC.
Process Safety and Environmental Protection 9 6 ( 2 0 1 5 ) 111–124 123
Artiga, P., Ficara, E., Malpei, F., Garrido, J., Mendez, R., 2005. Enayatifar, R., Abdullah, A.H., Isnin, I.F., 2014. Chaos-based image
Treatment of two industrial wastewaters in a submerged encryption using a hybrid genetic algorithm and a DNA
membrane bioreactor. Desalination 179, 161–169. sequence. Opt. Lasers Eng. 56, 83–93.
Azmy, A.M., Erlich, I., Sowa, P., 2004. Artificial neural Ferrari, S., Bellocchio, F., Piuri, V., Borghese, N.A., 2010. A
network-based dynamic equivalents for distribution systems hierarchical RBF online learning algorithm for real-time 3-D
containing active sources. IEE Proc. Gener. Transm. Distrib. scanner. IEEE Trans. Neural Networks 21, 275–285.
151, 681–688. Gander, M., Jefferson, B., Judd, S., 2000. Membrane bioreactors for
Badrnezhad, R., Mirza, B., 2014. Modeling and optimization of use in small wastewater treatment plants: membrane
cross-flow ultrafiltration using hybrid neural network-genetic materials and effluent quality. Water Sci. Technol. 41, 205–211.
algorithm approach. J. Ind. Eng. Chem. 20, Geissler, S., Wintgens, T., Melin, T., Vossenkaul, K., Kullmann, C.,
528–543. 2005. Modelling approaches for filtration processes with novel
Bagheri, M., Mirbagheri, S., Ehteshami, M., Bagheri, Z., 2014. submerged capillary modules in membrane bioreactors for
Modeling of a sequencing batch reactor treating municipal wastewater treatment. Desalination 178, 125–134.
wastewater using multi-layer perceptron and radial basis Guadix, A., Zapata, J.E., Almecija, M.C., Guadix, E.M., 2010.
function artificial neural networks. Process Saf. Environ. Prot., Predicting the flux decline in milk cross-flow ceramic
http://dx.doi.org/10.1016/j.psep.2014.04.006. ultrafiltration by artificial neural networks. Desalination 250,
Barello, M., Manca, D., Patel, R., Mujtaba, I.M., 2014. Neural 1118–1120.
network based correlation for estimating water permeability Guglielmi, G., Saroj, D.P., Chiarani, D., Andreottola, G., 2007.
constant in RO desalination process under fouling. Sub-critical fouling in a membrane bioreactor for municipal
Desalination 345, 101–111. wastewater treatment: experimental investigation and
Börger, E., Päppinghaus, P., Schmid, J., 2000. Report on a Practical mathematical modelling. Water Res. 41, 3903–3914.
Application of ASMs in Software Design Abstract State Han, H.-G., Qiao, J.-F., 2012. Prediction of activated sludge bulking
Machines-Theory and Applications. Springer, based on a self-organizing RBF neural network. J. Process
Berlin/Heidelberg, pp. 361–366. Control 22, 1103–1112.
Bouhabila, E.H., Ben Aïm, R., Buisson, H., 1998. Microfiltration of Henze, M., Grady, C., Gujer, W., Marais, G., Matsuo, T., 1987. A
activated sludge using submerged membrane with air general model for single-sludge wastewater treatment
bubbling (application to wastewater treatment). Desalination systems. Water Res. 21, 505–515.
118, 315–322. John, H., 1992. Holland, Adaptation in Natural and Artificial
Caro, J., 1998. Basic principles of membrane technology. Z. für Systems. MIT Press, Cambridge, MA.
Phys. Chem. 203, 263. Judd, S., 2010. The MBR Book: Principles and Applications of
Chamkalani, A., Amani, M., Kiani, M.A., Chamkalani, R., 2013. Membrane Bioreactors for Water and Wastewater Treatment.
Assessment of asphaltene deposition due to titration Elsevier, Amsterdam.
technique. Fluid Phase Equilib. 339, 72–80. Kashaninejad, M., Dehghani, A., Kashiri, M., 2009. Modeling of
Chang, I.-S., Le Clech, P., Jefferson, B., Judd, S., 2002. Membrane wheat soaking using two artificial neural networks (MLP and
fouling in membrane bioreactors for wastewater treatment. J. RBF). J. Food Eng. 91, 602–607.
Environ. Eng. 128, 1018–1029. Kılıç, U., Ayan, K., Arifoğlu, U., 2014. Optimizing reactive power
Chen, H.-W., Yu, R.-F., Ning, S.-K., Huang, H.-C., 2010. Forecasting flow of HVDC systems using genetic algorithm. Int. J. Electr.
effluent quality of an industry wastewater treatment plant by Power Energy Syst. 55, 1–12.
evolutionary grey dynamic model. Resour. Conserv. Recycl. 54, Kurek, W., Ostfeld, A., 2013. Multi-objective optimization of water
235–241. quality, pumps operation, and storage sizing of water
Chen, H., Kim, A.S., 2006. Prediction of permeate flux decline in distribution systems. J. Environ. Manage. 115, 189–197.
crossflow membrane filtration of colloidal suspension: a Le Clech, P., Jefferson, B., Chang, I.S., Judd, S.J., 2003. Critical flux
radial basis function neural network approach. Desalination determination by the flux-step method in a submerged
192, 415–428. membrane bioreactor. J. Membr. Sci. 227, 81–93.
Cho, B., Fane, A., 2002. Fouling transients in nominally Lipták, B.G., 2010. Instrument Engineers’ Handbook Volume Two:
sub-critical flux operation of a membrane bioreactor. J. Process Control and Optimization. CRC Press, Boca Raton.
Membr. Sci. 209, 391–403. Liu, Q.-F., Kim, S.-H., 2008. Evaluation of membrane fouling
Chon, K., Cho, J., Shon, H.K., 2013. A pilot-scale hybrid municipal models based on bench-scale experiments: a comparison
wastewater reclamation system using combined coagulation between constant flowrate blocking laws and artificial neural
and disk filtration, ultrafiltration, and reverse osmosis: network (ANNs) model. J. Membr. Sci. 310, 393–401.
removal of nutrients and micropollutants, and Liu, Q.-F., Kim, S.-H., Lee, S., 2009. Prediction of microfiltration
characterization of membrane foulants. Bioresour. Technol. membrane fouling using artificial neural network models.
141, 109–116. Sep. Purif. Technol. 70, 96–102.
Çinar, Ö., Hasar, H., Kinaci, C., 2006. Modeling of submerged Liu, S., Butler, D., Brazier, R., Heathwaite, L., Khu, S.-T., 2007.
membrane bioreactor treating cheese whey wastewater by Using genetic algorithms to calibrate a water quality model.
artificial neural network. J. Biotechnol. 123, 204–209. Sci. Total Environ. 374, 260–272.
Ct, P., Thompson, D., 2000. Wastewater treatment using Martin, K.J., Nerenberg, R., 2012. The membrane biofilm reactor
membranes: the North American experience. Water Sci. (MBfR) for water and wastewater treatment: principles,
Technol. 41, 209–215. applications, and recent developments. Bioresour. Technol.
Cui, Z., Chang, S., Fane, A., 2003. The use of gas bubbling to 122, 83–94.
enhance membrane processes. J. Membr. Sci. 221, Mendes, J., Araújo, R., Matias, T., Seco, R., Belchior, C., 2014.
1–35. Automatic extraction of the fuzzy control system by a
Delgrange, N., Cabassud, C., Cabassud, M., Durand-Bourlier, L., hierarchical genetic algorithm. Eng. Appl. Artif. Intell. 29,
Laine, J., 1998a. Modelling of ultrafiltration fouling by neural 70–78.
network. Desalination 118, 213–227. Meng, F., Yang, F., Shi, B., Zhang, H., 2008. A comprehensive study
Delgrange, N., Cabassud, C., Cabassud, M., Durand-Bourlier, L., on membrane fouling in submerged membrane bioreactors
Laine, J., 1998b. Neural networks for prediction of operated under different aeration intensities. Sep. Purif.
ultrafiltration transmembrane pressure–application to Technol. 59, 91–100.
drinking water production. J. Membr. Sci. 150, 111–123. Meng, F., Zhang, H., Yang, F., Liu, L., 2007. Characterization of
Dong, W.-y., Wang, H.-j., Li, W.-g., Ying, W.-c., Gan, G.-h., Yang, Y., cake layer in submerged membrane bioreactor. Environ. Sci.
2009. Effect of DO on simultaneous removal of carbon and Technol. 41, 4065–4070.
nitrogen by a membrane aeration/filtration combined Metcalf, I., 2003. Wastewater Engineering; Treatment and Reuse.
bioreactor. J. Membr. 344, 219–224. McGraw-Hill, New York.
124 Process Safety and Environmental Protection 9 6 ( 2 0 1 5 ) 111–124
Montastruc, L., Azzaro-Pantel, C., Pibouleau, L., Domenech, S., Sarma, M., Sahai, V., Bisaria, V., 2009. Genetic algorithm-based
2004. Use of genetic algorithms and gradient based medium optimization for enhanced production of fluorescent
optimization techniques for calcium phosphate precipitation. pseudomonad R81 and siderophore. Biochem. Eng. J. 47,
Chem. Eng. Process: Process Intensif. 43, 1289–1298. 100–108.
Moral, H., Aksoy, A., Gokcay, C.F., 2008. Modeling of the activated Sha, W., Edwards, K., 2007. The use of artificial neural networks in
sludge process by using artificial neural networks with materials science based research. Mater. Des. 28,
automated architecture screening. Comput. Chem. Eng. 32, 1747–1752.
2471–2478. Shetty, G.R., Chellam, S., 2003. Predicting membrane fouling
Naessens, W., Maere, T., Nopens, I., 2012. Critical review of during municipal drinking water nanofiltration using artificial
membrane bioreactor models—Part 1: Biokinetic and filtration neural networks. J. Membr. Sci. 217, 69–86.
models. Bioresour. Technol. 122, 95–106. Shokrian, M., Sadrzadeh, M., Mohammadi, T., 2010. C3H8
Nandi, B., Moparthi, A., Uppaluri, R., Purkait, M., 2010. Treatment separation from CH4 and H2 using a synthesized PDMS
of oily wastewater using low cost ceramic membrane: membrane: experimental and neural network modeling. J.
comparative assessment of pore blocking and artificial neural Membr. Sci. 346, 59–70.
network models. Chem. Eng. Res. Des. 88, 881–892. Soleimani, R., Shoushtari, N.A., Mirza, B., Salahi, A., 2013.
Nourbakhsh, H., Emam-Djomeh, Z., Omid, M., Mirsaeedghazi, H., Experimental investigation modeling and optimization of
Moini, S., 2014. Prediction of red plum juice permeate flux membrane separation using artificial neural network and
during membrane processing with ANN optimized using RSM. multi-objective optimization using genetic algorithm. Chem.
Comput. Electron. Agric. 102, 1–9. Eng. Res. Des. 91, 883–903.
Ordóñez Galán, C., Rodríguez-Pérez, J.R., Martínez Torres, J., Suchacz, B., Wesołowski, M., 2006. The recognition of similarities
García Nieto, P.J., 2011. Analysis of the influence of forest in trace elements content in medicinal plants using MLP and
environments on the accuracy of GPS measurements by using RBF neural networks. Talanta 69, 37–42.
genetic algorithms. Math. Comput. Modell. 54, 1829–1834. Trussell, R.S., Merlo, R.P., Hermanowicz, S.W., Jenkins, D., 2006.
Park, C., Lee, Y.H., Lee, S., Hong, S., 2008. Effect of cake layer The effect of organic loading on process performance and
structure on colloidal fouling in reverse osmosis membranes. membrane fouling in a submerged membrane bioreactor
Desalination 220, 335–344. treating municipal wastewater. Water Res. 40,
Park, H.-D., Lee, Y.H., Kim, H.-B., Moon, J., Ahn, C.-H., Kim, K.-T., 2675–2683.
Kang, M.-S., 2010. Reduction of membrane fouling by Ueda, T., Hata, K., Kikuoka, Y., Seino, O., 1997. Effects of aeration
simultaneous upward and downward air sparging in a on suction pressure in a submerged membrane bioreactor.
pilot-scale submerged membrane bioreactor treating Water Res. 31, 489–494.
municipal wastewater. Desalination 251, 75–82. Vaca Mier, M., López Callejas, R., Gehr, R., Jiménez Cisneros, B.E.,
Park, J.-W., Venayagamoorthy, G.K., Harley, R.G., 2005. MLP/RBF Alvarez, P.J., 2001. Heavy metal removal with mexican
neural-networks-based online global model identification of clinoptilolite:: multi-component ionic exchange. Water Res.
synchronous generator. Ind. Electron. IEEE Trans. 52, 35, 373–378.
1685–1695. Venayagamoorthy, G.K., 2007. Online design of an echo state
Patsios, S., Karabelas, A., 2011. An investigation of the long-term network based wide area monitor for a multimachine power
filtration performance of a membrane bioreactor (MBR): the system. Neural Netw. 20, 404–413.
role of specific organic fractions. J. Membr. Sci. 372, 102–115. Verrecht, B., Maere, T., Nopens, I., Brepols, C., Judd, S., 2010. The
Pendashteh, A.R., Fakhru’l-Razi, A., Chaibakhsh, N., Abdullah, cost of a large-scale hollow fibre MBR. Water Res. 44,
L.C., Madaeni, S.S., Abidin, Z.Z., 2011. Modeling of membrane 5274–5283.
bioreactor treating hypersaline oily wastewater by artificial Wallhäußer, E., Hussein, W.B., Hussein, M.A., Hinrichs, J., Becker,
neural network. J. Hazard. Mater. 192, 568–575. T.M., 2011. On the usage of acoustic properties combined with
Psoch, C., Schiewer, S., 2006. Anti-fouling application of air an artificial neural network—a new approach of determining
sparging and backflushing for MBR. J. Membr. Sci. 283, presence of dairy fouling. J. Food Eng. 103,
273–280. 449–456.
Rani, D., Jain, S.K., Srivastava, D.K., Perumal, M., 2013. 3—Genetic Wang, X., Chen, Y., Zhang, J., Li, X., Ren, Y., 2013. Novel insights
algorithms and their applications to water resources systems. into the evaluation of submerged membrane bioreactors
In: Yang, Xin-She, Gandomi, Amir Hossein, Talatahari, Siami, under different aeration intensities by carbon emission.
Alavi, Amir Hossein (Eds.), Metaheuristics in Water, Desalination 325, 25–29.
Geotechnical and Transport Engineering. Elsevier, Oxford, pp. Weiss, J., Bernardara, P., Andreewsky, M., Benoit, M., 2012.
43–78. Seasonal autoregressive modeling of a skew storm surge
Ratkovich, N., Chan, C., Bentzen, T.R., Rasmussen, M.R., 2012. series. Ocean Modell. 47, 41–54.
Experimental and CFD simulation studies of wall shear stress Xi, X., Cui, Y., Wang, Z., Qian, J., Wang, J., Yang, L., Zhao, S., 2011.
for different impeller configurations and MBR activated Study of dead-end microfiltration features in sequencing
sludge. Water Sci. Technol. 65, 2061–2070. batch reactor (SBR) by optimized neural networks.
Rocha, I., Parente Jr., E., Melo, A., 2014. A hybrid Desalination 272, 27–35.
shared/distributed memory parallel genetic algorithm for Zarragoitia-González, A., Schetrite, S., Alliet, M., Jáuregui-Haza,
optimization of laminate composites. Compos. Struct. 107, U., Albasi, C., 2008. Modelling of submerged membrane
288–297. bioreactor: conceptual study about link between activated
Sahoo, G.B., Ray, C., 2006. Predicting flux decline in crossflow slugde biokinetics, aeration and fouling process. J. Membr. Sci.
membranes using artificial neural networks and genetic 325, 612–624.
algorithms. J. Membr. Sci. 283, 147–157. Zhu, X., Wang, Z., Wu, Z., 2011. Characterization of membrane
Saltelli, A., Ratto, M., Andres, T., Campolongo, F., Cariboni, J., foulants in a full-scale membrane bioreactor for supermarket
Gatelli, D., Saisana, M., Tarantola, S., 2008. Global Sensitivity wastewater treatment. Process Biochem. 46, 1001–1009.
Analysis: The Primer. Wiley.com, Chichester, West Sussex.