Modeling and optimization of ceramic membrane

microfiltration using neural networks and genetic

algorithms

S. Strugholtz*, S. Panglisch**, J. Gebhardt***, R. Gimbel*

*University of Duisburg-Essen, Institute for Energy and Environmental Engineering,
Bismarckstrasse 90, 47057 Duisburg, Germany, [email protected]
**IWW Rhenish-Westfalian Instiute for Water Research, Moritzstrasse 24, 45476 Muelheim an
der Ruhr, Germany, [email protected]
***aquatune - Dr. Gebhardt & Co. GmbH, Untig Mühl, 65326 Aarbergen,
[email protected]

Abstract Membrane plants for drinking water treatment should not only deliver a good water
quality but should also operate cost effective. Therefore a two stage procedure was applied
for optimization of a ceramic membrane microfiltration processes with coagulation
pretreatment. First neural networks were applied for prediction of the course of
transmembrane pressure (TMP) over several cycles with high precision. With a sensitivity
analysis relationships between influencing parameters could be shown. In a second step
these models were applied for operational costs optimization by genetic algorithms. Based on
the idea of Darwin’s survival of the fittest, settings for filtration time, flux and aluminum dosage
were optimized leading to minimized operational costs with a costs reduction of about 30 %.
The selected study proved the effectiveness of genetic algorithms and the applicability for
online optimization being planned for further studies.

Keywords: neural networks, genetic algorithms, drinking water, ceramic membranes,

optimization

Introduction
Drinking water treatment plants have to provide a good water quality and at the same time
low operational costs. Due to the complexity of the raw water matrix processes are often
difficult to handle and reliable predictions for the course of process are difficult to obtain.
Therefore great importance is attached to the application of neural networks as black box
models in a wide field of drinking water treatment facilities (Baxter et al. 2001).
In recent years the use of membrane technology has become an economic alternative to
conventional processes. However, before realization of large-scale plants usually extensive
and time consuming studies in bench-scale or pilot plants studies are necessary to adapt the
plant to the on-site conditions. This is not only necessary for the membrane step itself but
also for possible pre-treatment facilities like e.g. coagulation/flocculation. Up to now these
studies are performed by a mixture of "trial and error" and operators’ experiences from other
sites. Additionally, membrane plants in dead-end mode are usually operated based on fixed
time schedules instead of relating operational parameters to transmembrane pressure (TMP)
development or incoming changing water conditions.

Water Practice & Technology Vol 1 No 4 © IWA Publishing 2006 doi: 10.2166/WPT.2006083
Modeling and optimization of ceramic membrane microfiltration using neural networks and genetic
algorithms

Due to the lack of knowledge about influencing parameters neural networks as black box
models are capable for modeling fouling (Cabassud 1998, Shetty 2003) as well as for
performance prediction (Teodosiu 2000, Oh 2004, Veerapaneni 2004) from operational and
water quality data derived by online data logging and laboratory measurements during pilot
or bench-scale testing.
In surface water treatment coagulation is a common pre-treatment step for micro- and
ultrafiltration (MF/UF) processes. Neural network approaches also exist for modeling
coagulant dosage (Gagnon 1997, Maier 2004) but so far not in combination with a membrane
treatment step. Additionally, in dead-end membrane processes chemical enhanced
backwashes (CEB) in defined intervals are necessary to improve process stability.
Veerapaneni et al. (2004) included backwash frequencies also for CEB in their networks but
without detailed examination of single cycles.
In a further step existing neural models can be used for process control for example to
avoid irreversible fouling and improve productivity (Cabassud 2002). For existing models
many optimization approaches can be applied, e.g. conjugate gradient methods, Newton’s
method or quadratic programming. All these methods have the drawback that only one
solution per iteration is generated which may be located in a local optimum. This limitation
can be avoided by using genetic algorithms where a variety of settings in form of
chromosomes is presented to the model over several generations. The optimization is gained
by the principle of Darwin’s survival of the fittest. Genetic algorithms have already been
used for membrane processes optimization on the basis mathematical models (Yuen 2000,
Murthy 2006). In MF/UF drinking water treatment processes up to now no reliable
mathematical model exists to describe permeate flow or transmembrane pressure. Therefore
the combination of neural network modeling and genetic optimization is a promising way to
increase membrane process productivity
Within this study neural networks were trained for predicting the TMP at cycle start as
well as the TMP at cycle end, considering process parameters, water quality parameters as
well as coagulation and CEB parameters. A sensitivity analysis of the neural network shows
the influence of different input parameters on the targets which may also contain hints on
mechanisms in membrane filtration. With a reliable prediction quality genetic algorithms can
be applied to the models for the purpose of finding the optimal combination of control
variables (e.g. coagulant dose, filtration flux and time, backwash frequency, chemical
demand). By this means the productivity can be optimized and chemically enhanced
backwashes (CEB) can be initiated by necessity instead of a rigid time schedule leading to
reduced operational costs.

Methods
Neural networks
For this study commercially available software (NeuroModel, Atlan-tec, Germany) was used
for neural network modeling. It works with the back propagation algorithm containing
special methods for convergence which are based on the method of conjugate gradients. The
network structure consists of an input layer, one hidden layer and an output layer. The
number of neurons and training settings are fixed by an expert system of the software.
Further information on neural networks can be found in Bishop (1995).
Modeling and optimization of ceramic membrane microfiltration using neural networks and genetic
algorithms

Before network training typical data are generated from the raw data by hierarchical data
clustering, assuring that all settings, also non stationary, are equally presented to the network
as training data.

Figure 1.1 SecurityNet to evaluate prediction accuracy

To use neural networks for optimization and control information on the prediction
security is needed. The more typical are trained in certain areas of the parameter range the
better will the prediction quality be. Therefore a SecurityNet is integrated (Froese, 1997).
This patented method as shown in figure 1.1 works with five parallel trained networks with
different initializations from which the obtained errors are evaluated. The evaluation of these
errors provides prediction accuracy in every point of the parameter range. For optimization
and control only parameter settings within defined prediction accuracy can be used. This
system can also be applied for an aimed design of experiments in regions where the
prediction quality is low.

Genetic algorithms
Genetic algorithms are inspired by Darwin’s evolutional theory that in nature always the
fittest individuals survive. By transferring possible process settings to binary coded
chromosomes, lots of possible operating points can be tested at the same time, striving for
the fittest chromosome, i.e. striving for the global maximum of the parameter range. In this
way several manipulable variables can be optimized at the same time at a comparable low
computational complexity.
From an initial population the fittest solution of the optimization problem, e.g. in case of
this study the minimization of operational costs, is searched over a defined number of
generations. From generation to generation in each case the fitness of two chromosomes is
compared and the one with the higher fitness is selected for the next generation. Additionally
the selected chromosomes undergo genetic operations, namely crossover and mutation,
gaining new chromosomes for the children’s’ generation. A scheme of the process is shown
in figure 1.2.

Figure 1.2 Selection and Mutation of genetic algorithms

Using neural networks as model for genetic optimization, the input parameters are divided
into constants or disturbance variables and into manipulable variables which are the ones to
Modeling and optimization of ceramic membrane microfiltration using neural networks and genetic
algorithms

be optimized for the given optimization problem. The manipulable variables provide a data
range in which they may be varied. This should be selected within the area of training
parameters, assuring that no extrapolation is possible. Additionally, the output variable of the
network must be inside a defined range of the SecurityNet value obtained from network
training, so that each solution is near a solution that was already present in the training data
set. If there are other logic or operational barriers to be taken care of these must also be
implemented as barriers for the search area. For optimization a target function including the
manipulable variables is defined. Each time a better fitness is generated the chromosomes are
decoded back to the original values.

Ceramic membrane microfiltration

Due to mechanical robustness and chemical resistance the application of ceramic
membranes would offer operational advantages over polymeric membranes which are
currently dominant on the water treatment market. In Japan a new membrane system with a
high packing density was introduced by NGK Insulators Ltd., Japan, allowing economic
large scale applications. This membrane system was tested by IWW Water Center, Mülheim,
Germany (IWW) on behalf of NGK in a pilot plant study to investigate the technological and
economical performance for hybrid surface water treatment process with coagulation
pretreatment over one year. In the plant up to 6 test-modules with a membrane surface of 0.4
m² each could be operated in parallel. The membranes have a nominal pore size of 0.1 µm,
the diameter of the feed channels is 2.5 mm. The coagulant was first dosed into a mixing
tank; after a half year coagulation was changed to inline-coagulation. The pilot plant was
fully automated and connected to a data logger. Besides the operational parameters flux and
pressure the water quality parameters turbidity, coagulation pH and temperature were
measured online. Additional information on plant design and operational/economical results
can be found in Lerch et al. (2005) and Loi-Bruegger et al. (2006)

Modeling and optimization of membrane microfiltration

For the operator of membrane plants two things are of interest:
• The prediction of membrane processing is desirable to evaluate how the process will
operate several cycles later assuming constant conditions. Prediction quality is compared
to experimental results (small dashed quadrate in figure 1.3).
• Online optimization of manipulable variables is desirable to ensure economical settings
for an optimal productivity. In this case precisely trained networks are used for genetic
optimization of a given cost function by modification of manipulable parameters. The
optimized specific costs can be compared to the real operational costs for energy,
chemical demand and effluent (big dashed quadrate in figure 1.3).
Modeling and optimization of ceramic membrane microfiltration using neural networks and genetic
algorithms

60
Present cycle
55
TMPe_1 TMPe
50

TMP [kPa]
mTMP_1 mTMP
45

40
TMPs_1 TMPs TMPs+1
35

30 tbw tf
TMP calculated TMP measured Backwash
CEBalkaline CEBacidic

Figure 1.3 Modeling and optimization Figure 1.4 Calculation of characteristic points

To describe the membrane process, detailed knowledge on each point of filtration is not
necessary as during the filtration step TMP rises linearly and in case of this ceramic
membrane process the backwash step follows the same procedure each time but with various
CEB intervals. Therefore characteristic finger prints were calculated for each filtration cycle
consisting of (i) backwash (neutral or CEB) and (ii) filtration. The slope of TMP (mTMP) is
obtained by linear regression from minutely logged TMP data. From slope and filtration time
TMPs at cycle start and TMPe at cycle end can be determined. As the pH of CEB was not
continuously measured, a pH value was set constant for neutral (pH7), alkaline (pH10.5) and
acidic (pH2), giving the network a parameter for differentiation between the different
programs. Important parameters do describe the backwash are the backwash efficiency
coefficient (BEC), the difference between two following TMPs (DTMPs) to describe fouling
or cleaning benefit from one cycle to the other as well as the irreversible fouling TMPirr as
difference between TMPs and the constant start TMP (TMPo) after an intensive cleaning in
place (CIP) or installation of new modules. Also the coagulant dosing had to be set constant
as it was not measured, but the error from this is quite low as a high precision dosing pump
was used. All parameters with description, calculation and limit values are given in table 1.
Training was conducted with about 3250 datasets collected over one year of operation. Some
process parameters are shown in figure 1.5 and for selected parameters in more detail in
figure 1.6. Before training about 600 datasets from different times of operation with different
settings for filtration time, flux, backwashing program and aluminum dose were separated for
testing and parameter prediction with untrained data. Model deviation is calculated by

Prediction − Measure
Deviation =
Data range of training parameters
Modeling and optimization of ceramic membrane microfiltration using neural networks and genetic
algorithms

Table 1.1 Model parameters with limit values and their usage as network input parameters
Description Abbrevation Unit min max Measure TMPs_all TMPs_red TMPe_all TMPe_red TMPe_opt
Water quality parameters Network input parameters
Temperature Feed_T °C 2 25 online _1 _1 _1 _1 0
Turbidity Feed_Tur FNU 0.5 100 online _1 x _1 _1 0
Load with aluminum
Load_Al mg/m² 235 1575 calculated _1 _1 _1 _1 _1
c(Al)*Flux*tf
Load with Turbidity
Load_Tur FNU*l/m² 80 30000 calculated _1 _1 _1 x x
Feed_Tur*Flux*tf
Process parameters
Flocculation pH Feed_pH - 6.8 7.5 online _1 x _1 x _1
Feed pressure Feed_P kPa 20 175 online _1 _1 _1 _1 _1
TMP cycle start TMPs kPa 8 110 calculated _1 _1 _1,0 _1,0 _1,0
TMP cycle end TMPe kPa 8 205 calculated _1 _1 _1 _1 _1
TMP cycle slope mTMP kPa/min 0.005 2 calculated _1 _1 _1 _1 _1
TMPs-TMPs_1 DTMPs kPa -15 15 calculated _1 _1 _1 _1 _1
Irreversible fouling
TMPirr kPa -5 90 calculated _1 _1 _1 _1 _1
TMPs-TMPo
Backwash time tbw min 2 19 calculated _1 x _1 x x
Backwash pH bw_pH - 2 10.5 set _1 _1 _1 _1 _1
Last backwash with
t_same_pH min 47 7000 calculated _1 _1 _1 _1 _1
the same pH
Backwash efficiency
(TMPs-TMPe_1)/ BEC - 0.4 9 calculated _1 _1 _1 _1 _1
(TMPs_1-TMPe_1)
Manipulable variables
Flux Flux l/(m²h) 80 300 online _1 _1 _1 _1 0
Filtration time tf min 45 120 calculated _1 x _1 x 0
Al-concentration Al mg/l 1.5 3.5 set _1 _1 _1 _1 0
_1: input parameter from the cycle before the present one
0: input parameter from the present cycle
x: parameter not used for modeling

300 120

250 100
TMP [kPa], Flux[l/m²h]

Filtration Time [min],

Turbidity [FNU]
200 80

150 60

100 40

50 20

0 0
04

04

04

04

04

05

05

05

05
6.

8.

9.

1.

2.

2.

4.

5.

7.
.0

.0

.0

.1

.1

.0

.0

.0

.0
12

01

20

09

29

17

08

28

17

Time

TMPs TMPe Flux Turbidity Filtration Time

Figure 1.5 Experimental data

Modeling and optimization of ceramic membrane microfiltration using neural networks and genetic
algorithms

30 0,050
0,045
25
0,040

mTMP [kPa/min]
0,035
20
TMP [kPa]
pw_pH [-]
0,030
15 0,025
0,020
10
0,015
0,010
5
0,005
0 0,000
04

04

04

04

04

04

04
1.

1.

1.

1.

1.

1.

1.
.1

.1

.1

.1

.1

.1

.1
09

11

13

15

17

19

21
Time

TMPs TMPe bw_pH mTMP

Figure 1.6 Detailed progresses of selected parameters

Modeling settings can be seen in table 1. First model training was conducted with all
possible input parameters. Afterwards, the networks were validated and evaluated with a
sensitivity analysis. In sensitivity analysis each input variable is varied by +/- 0.5 % of its
parameter range. All impacts on the target value are statistically evaluated so that for each
input parameter the mean absolute sensitivity and the mean sensitivity showing the positive
or negative influence of each input on the target are generated. The analysis gives insight into
relationships between the model parameters and can be interpreted by the degree of influence
and positive or negative impact on the target parameter. Herewith the input parameters can
be reduced to the necessary ones and the network performance can be optimized.
For parameter prediction only input parameters already measured can be presented to the
network. Therefore the neural networks for prediction were trained in a way that TMPs was
predicted from the values measured one cycle before (network TMPs_red) and with this
prediction, TMPe was predicted from TMPs and from input parameters of the cycle before
(network TMPe_red). TMPe was then the new input for the next prediction of TMPs+1. As
Figure 1.7 shows, with this method prediction over several cycles is possible.

Figure 1.7 Prediction of TMPs and TMPe

For optimization a new network for TMPe prediction was trained (network TMPe_opt).
This network had to contain the manipulable variables of the actual cycle (tf, Flux, Al) that
were to be optimized and the actual measurable water quality parameters (Feed_T,
Feed_Tur). As TMPs is the first parameter of the optimized filtration step of the actual cycle,
TMPs was still predicted by the parameters measured or determined one cycle before
(network TMPs_red). After prediction TMPs was delivered to the network TMPe_opt which
was used as model for the genetic optimizer (GenOpt, Atlan-tec, Germany). Here a cost
function was defined in order to optimize the malleable variables gaining minimal costs. The
Modeling and optimization of ceramic membrane microfiltration using neural networks and genetic
algorithms

results of the modified process parameters from the optimizer were given back to the model
TMPs_red with a time delay of one simulation step. The procedure is shown in figure 1.8.

Figure 1.8 Procedure of costs minimization with genetic algorithms

As backwashing always follows the same procedure constant mean costs of 0,012 €/m²
membrane surface can be assumed for effluent and chemical costs. As the process is run with
constant flux the hydraulic power Phydr is proportional to TMP.

Phydr = TMP ⋅ Flux (1)

The applied energy Ehydr is given by integration of Phydr over filtration time. Assuming a
linear TMP increase the applied hydraulic energy is

TMPe − TMPs
Ehydr = 0.7 ⋅ ⋅ Flux ⋅ tf (2)
2

assuming a constant efficiency factor of 0.7. Energy costs can be specified with 0.08 €/kWh.
Costs for aluminum dosing are given by the aluminum load Load_Al (table 1) and a price
of 0.30 €/kg. Produced water Qprod is given by

Q prod = Flux ⋅ tf (3)

and process efficiency is given by the time slice of production tprod

tf
t prod = (4)
tf + tbw

This gives the specific cost function

0.012 + 0.08 ⋅ Ehydr + 0.3 ⋅ Load _ Al

Costs = ⋅ t prod (5)
Q prod
Modeling and optimization of ceramic membrane microfiltration using neural networks and genetic
algorithms

Results and Discussion

TMP modeling and prediction
All literature on neural networks mentioned in the introduction show that modeling of
membrane processes is possible with a quite low prediction error. Therefore only one
exemplary plot of parameter prediction is shown later in figure 2.4 for costs optimization.
For TMP prediction it is of interest to interpret the sensitivity analysis and to evaluate the
addition of errors for stringing together the networks for prediction of TMPs and TMPe. This
gives an impression of the time horizon proper prediction of the process under constant
conditions is possible.
Figure 2.1 and figure 2.2 show the results of sensitivity analysis for the networks
TMPs_red, TMPe_red and also for the network TMPe_opt used later on for optimization. For
some relationships figure 1.6 may be useful for better understanding.
After first training the input parameters were reduced to the necessary ones by sensitivity
analysis resulting in the networks TMPs_red and TMPe_red. For both networks the
parameters tf_1, tbw_1, Feed_pH_1, which was mostly constant at pH7, and one information
on turbidity in the cycle before (Feed_Tur_1 and Load_Tur_1 respectively) could be
neglected. From linear regression TMP increase was described by a linear equation from
TMPs to TMPe over filtration time tf and with the slope mTMP. Using all four parameters for
network training, the TMP increase would be overdetermined. As tf was the parameter of this
linear equation with the lowest variability it was the input parameter with the lowest
sensitivity and could therefore be neglected. Also backwash time tbw only varied between
two options with longer backwash time for CEB due to soaking with chemical solution.
More important parameters to describe the backwash procedure were the backwash
efficiency BEC_1 and the period tbw_same_pH_1 between two equal backwashes and CEB
respectively. The input tbw_same_pH_1 shows negative sensitivity. High values for
tbw_same_pH_1 arise from the longer intervals between two CEBs leading to lower TMPs
and TMPe in the next cycle. If BEC_1 increases e.g. after CEB, TMPs will be lower in the
cycle afterwards, if BEC_1 decreases, irreversible fouling occurs leading to an increasing
TMPs.
Turbidity only showed few peaks over the year. This resulted in the low sensitivity for
this input parameter though it was not negligible. Both for Load_Tur_1 and Load_Al_1 a
positive sensitivity could be found as for higher loads on the membrane a higher pressure
was needed. This is also the reason for positive sensitivity of Feed_P_1. From experimental
results it appeared that a higher coagulant dosage is favorable for the process (Lerch, 2005).
This relates to the negative sensitivity of aluminum concentration Al and Al_1. The high
dependency on aluminum dosage and load due to flocculation shows that an optimal
coagulation process could improve the membrane process. Therefore in further studies the
coagulation process will be considered in more detail.
Modeling and optimization of ceramic membrane microfiltration using neural networks and genetic
algorithms

30%
TMPs_red TMPe_red TMPe_opt

Mean absolute sensitivity [%]

20%

10%

0%

Flux_1

bw_ph_1
tbw_same_pH_1

Feed_T
Al
TMPs

tf

Flux
mTMP_1
TMPs_1
TMPe_1
DTMPs_1
TMPirr_1
BEC_1

Feed_T_1
Feed_P_1
Al_1
Load_Al_1
Load_Tur_1

Feed_Tur_1

Feed_Tur
Figure 2.1 Sensitivity analysis results for degree of influence

0,8
TMPs_red TMPe_red TMPe_opt
0,6

0,4
Mean sensitivity

0,2

-0,2

-0,4

-0,6
Flux_1

bw_ph_1
tbw_same_pH_1

Feed_T
Al
TMPs

tf

Flux
mTMP_1
TMPs_1
TMPe_1
DTMPs_1
TMPirr_1
BEC_1

Feed_T_1
Feed_P_1
Al_1
Load_Al_1
Load_Tur_1

Feed_Tur_1

Feed_Tur

Figure 2.2 Sensitivity analysis results for positive or negative influences

To describe irreversible fouling by TMPirr was only relevant for prediction of TMPs
though the change in TMPs from one cycle to the other (DTMPs_1) was more important. For
increasing TMPs the value for DTMPs showed a negative impact. Most backwashes were
conducted with filtrate. This led to higher TMPs in the next cycle which gives negative
DTMPs. The negative sensitivity of mTMP_1 on TMPs is explainable from the same
relationship as mTMP decreases with increasing TMPs during periods with filtrate
backwashes. If TMPs increases and mTMP decreases TMPe must relatively decrease leading
to the positive sensitivity or mTMP_1 for prediction of TMPe.
The model TMPe_opt for costs optimization shows a negative sensitivity for tf. This is
explainable from the course of experiments. The process was operated with high filtration
time when flux was low (figure 1.5). The low flux led to a comparatively lower TMPe.
Modeling and optimization of ceramic membrane microfiltration using neural networks and genetic
algorithms

For parameter prediction the networks were stringed together as explained above. The
errors that arose are shown in figure 2.3. The deviation increased from about 1.1 % with a
slightly higher error for TMPs prediction in form of a saturation curve. After Cycle+6 the
error increased a little stronger. Prediction horizon for 7 cycles to the future (Cycle+6) would
be about 7 h to 14 h depending on the filtration time applied in the different test periods.
Mean absolute deviation [%]

4
3,5
3
2,5
2
1,5
1
0,5 TMPs TMPe
0
0

+1

+2

+4

+5

+6

+7

+8

9
e+

e+
le

le

cle

cle
le

le

le

le
yc

cl

cl
yc

yc

yc

yc

yc
Cy

Cy

Cy

Cy
C

C
Cycle

Figure 2.3 Deviation increase for parameter prediction over 10 cycles

Costs optimization
For cost optimization first a precisely predicted dataset should be found. Therefore the
networks for TMPs_red and TMPe_opt were stringed together like for parameter prediction
with the difference that the input parameters were always the actual ones or the ones of the
cycle before as trained. The calculated parameters BEC, DTMPs, TMPirr, mTMP, Load_Al
and Load_Tur here were calculated from the prediction of TMPs and TMPe. Figure 2.3
shows the result for prediction of 10 cycles to the future exemplary for prediction of TMPs
with a reduced number of datasets for faster calculation performance. A mean average error
of 1.8 % for TMPs_red and of 1.5 % for TMPe_opt arose for prediction of 10 cycles to the
future. For optimization datasets with very low deviation from 120 to 131 were chosen.

80 80
Period 2 Period 3
tf = 90 min tf = 90 min
Flux = 200 l/m²h Flux = 250 l/m²h
60 60
Al = 3.5 mg/l Al = 3.5 mg/l
TMPs+9 [kPa]

Deviation [%]

40 Period 1 40
tf = 120 min changed
Flux1 = 120 l/m²h backwash
Flux2 = 80l/m²h program
20 20
Al = 1.5 mg/l

0 0

-20 -20
0 100 200 300
Datasets
TMPs measured cycle+9 TMPs predicted cycle+9
Deviation Deviation +/-10%

Figure 2.4 Prediction of 10 cycles to the future

Modeling and optimization of ceramic membrane microfiltration using neural networks and genetic
algorithms

55
50

Cost reduction [%]

TMP [kPa] 45
40
35
30
25
20
09.04.2005 10.04.2005 10.04.2005 10.04.2005 10.04.2005 10.04.2005
19:12 00:00 04:48 09:36 14:24 19:12
Time
TMPs TMPe TMPs_opt TMPe_opt Costs reduction

Figure 2.5 Costs optimization and development of process parameters

The mean operational costs calculated from experimental results within this period
amounted 4.5 €-Cent. As figure 2.5 shows, with optimization savings could be about 30 %.
The results are transferable to other data ranges. The genetic optimizer is usually used for
online optimization. As only historical data were available this study shows the effectiveness
of these genetic algorithms. For further studies the genetic optimizer will be applied online.
Costs optimization was done by productivity increase. Genetic optimization proposed to
operate the plant with maximum flux of 250 l/m²h and high filtration time of 120 min.
Aluminum dosage was optimized to about 3.4 mg/l which is near the dosage that was found
to be optimal during the experiments. Due to higher Flux the TMPe increased. For the last
predictions the increase of TMPe from one cycle to the next is equal to the operational
results. TMPs stayed on the same level as without optimization so that backwashes are
effective enough to run the optimized process as stable as without optimization. The break in
point 3 is an artifact from the calculation procedure. The network for TMPs prediction first
needs starting points for the calculated parameters, DTMPs, TMPirr, mTMP, Load_Al and
Load_Tur. After the second loop all parameters could be calculated from the predictions.
In this optimization the influence of CEB is not considered as the intervals of CEB were
equal to membrane plant operation. This must be done in a further step as irreversible fouling
has to be reduced to a minimum.

Conclusions
Due to the high complexity of membrane processes models are useful for parameter
prediction to obtain insight in mechanisms and for process optimization. Most promising
way for modeling is the application of neural networks. For optimization a new approach of
combination of neural networks with genetic algorithms was used. Genetic Algorithms are a
powerful tool to find the optimal solution for a multidimensional optimization problem by an
algorithm related to Darwin’s survival of the fittest.
In this study neural networks were used for parameter prediction over a time horizon of 7
to 14 hours with high accuracy. For this, characteristic values were calculated to describe a
cycle consisting of backwash and following filtration. The data derived from one year pilot
plant study of ceramic membrane microfiltration. For this process models for transmembrane
pressure (TMP) prediction at filtration start and end could be trained that also considered
Modeling and optimization of ceramic membrane microfiltration using neural networks and genetic
algorithms

influences of coagulation pretreatment and chemical enhanced backwash (CEB). Sensitivity

analysis was performed to gain closer insight into relationships between relevant parameters.
The influences of coagulation and CEB offer a great potential for further improvements of
model precision and description of mechanisms.
Already the experimental results showed an economic application alternative of ceramic
membranes to more common polymer ones. By application of genetic algorithms it could be
shown that a reduction of operational costs is possible. Up to now membrane process is
operated with constant process adjustments. With genetic algorithms it is possible to find
more optimal values for manipulable variables which in case of this study were filtration
time, flux and aluminum dosage. For the presented optimization problem it could be shown
that processing with maximum filtration time and flux were possible. This causes a big
increase of productivity and therefore operational costs reduction. For this study only
selected case studies for optimization were possible as only historical data were available.
But the results of this study show that the application of neural networks with genetic
algorithms is a very promising and powerful tool for online optimization of membrane
processes also in combination with pretreatment steps.

References
Baxter, C.W., Zhang, Q., Stanley, S.J., Shariff, R., Tupas, R-R.T., Stark, H.L. (2001). Drinking water
quality and treatment: the use of artificeal neural networks, Canadian Journal of Civil
Engineering, 28(Suppl 1), 26-35
Bishop, C.M. (1995). Neural networks for pattern recognition, Oxford University Press
Cabassud, M., Delgrange-Vincent, N., Cabassud, C., Durand-Bourlier, L., Lainé, J.M. (2002). Neural
networks: a tool to improve UF plant productivity, Desalination, 145 223-231
Delgrange, N., Cabassud, C., Cabassud, M., Durand-Bourlier, L., Lainé, J.M. (1998). Neural networks
for prediction of ultrafiltration transmembrane pressure – application to drinking water production,
Journal of Membrane Science, 150, 111-123
Froese T. (1997). Verfahren zur Erkennung von fehlerhaften Vorhersagen in einer neuromodell-
gestützten oder neuronalen Prozessregelung“, European Patent Application EP 0 762 245 A1
Gagnon, C., Grandjean, H.P.A., Thibault, J. (1997), Modeling of coagulant dosage in a water treatment
plant, Artificial Intelligence in Engineering, 11, 401-404
Lerch, A., Panglisch, S., Gimbel, R. (2005). Research experiences in direct potable water treatment
using coagulation/ultrafiltration, Water Science and Technology, 51 (6-7), 221-229
Loi-Brügger, A., Panglisch, S., Buchter, P., Gimbel, R., Hattori, K., Yonekawa, H., Tomita, Y. (2006).
One year pilot study on economical surface water treatment with new ceramic membranes,
Industrial Water Conference Proceedings, Frankfurt, Germany
Maier, H.R., Morgan, N., Chow, C.W.K. (2004). Use of artificial neural networks for predicting
optimal alum doses and treated water quality parameters, Environmental Modeling & Software,
19, 485-494
Murthy, Z.V.P., Vengal, J.C. (2006). Optimization of a reverse osmosis system using genetic
algorithms, Separation Science and Technology, 41, 647-663
Oh, H.K., Yu, M.J., Gwon, E.M., Koo, J.Y., Kim, S.G., Koizumi, A. (2004). KNT-artifical neural
network model for flux prediction of ultrafiltration membrane producing drinking water, Water
Science and Technology, 50(8), 103-110
Shetty, G.R., Chellam, S. (2003). Prediction of membrane fouling during municipal drinking water
nanofiltration using artificial neural networks, Journal of Membrane Science, 217, 69-86
Teodosiu, C., Pastravanu, O., Macoveanu, M. (2000). Neural network models for ultrafiltration and
backwashing, Water Research, 34(18), 4371-4380
Veerapaneni, S., Budd, G., Horsley, M., Freeman, S. (2004). Use of Neural Networks for prediction
performance of MF/UF membrane systems (Extended abstract), CD-ROM Proceedings, American
Water Works Association Annual Conference, Orlando, USA
Modeling and optimization of ceramic membrane microfiltration using neural networks and genetic
algorithms

Yuen, C.C., Aatmeeyata, Gupta, S.K., Ray, A.K. (2000). Multi-objective optimization of membrane
separation modules using genetic algorithms, Journal of Membrane Science, 176, 177-196