ARTICLE

International Journal of Water Sciences

Water Turbidity Modelling During

Water Treatment Processes Using
Artificial Neural Networks
Regular Paper

Adam Rak1,*

1 Opole University of Technology, Faculty of Civil Engineering, Opole, Poland

* Corresponding author E-mail: [email protected]

Received 25 Feb 2013; Accepted 21 Jun 2013

DOI: 10.5772/56782
2013 Rak; licensee InTech. This is an open access article distributed under the terms of the Creative
Commons Attribution License (http://creativecommons.org/licenses/by/3.0), which permits unrestricted use,
distribution, and reproduction in any medium, provided the original work is properly cited.

AbstractArtificialneuralnetworksareincreasinglybeing
used in the research and analysis of unit and technical
processes related to water treatment.An artificial neural
network model was created to predict the turbidity of
treated water in a newly operating water treatment
system for surface and retention water at the Sosnwka
reservoir,Poland.
To model water turbidity during the water treatment
processforaselectedsystem,aflexibleBayesianmodelof
neural networks, Gaussian processes and mixtures that
demonstrate Markov chains of Monte Carlo methods
(FBM) was applied. A model was created to predict the
turbidity of water, pretreated in a specific technical
system. The physicochemical parameters of the water
drawn from the reservoir and those of the treated water
constitutedtheinputsforthemodel.
Two selection criteria for the final neural network
parameters were assumed. The first was the root mean
squared error (RMSE) and the second, the correlation
factor, R, between the projected and observed raw water
turbidity values. The calculated correlation factor was
R = 0.84. The results achieved proved that neural
networkscanbeappliedtopredictthequalityfactorsfor
waterpretreatedinaspecifictechnicalsystem.
www.intechopen.com

Keywords Artificial Neural Networks, Technological

System,WaterTreatment,WaterTurbidity

1.Introduction

A basic problem with taking water from rivers and

mountaintorrentsforpipelinedistributionistheperiodic
excessive turbidity of the water. This turbidity is caused
by sudden intensive rainfalls or spring thaws, which
increase the water level in the rivers and the amount
of suspended and other solids carried. The amounts of
these solids depend on the climatic, geological,
geomorphological, hydrobiological and biological
conditions; they may also result from human activities.
Theperiodofhighwaterturbiditymaylastfromseveral
hours to a month. Changes in the turbidity of
mountainous rivers can happen very quickly and the
turbiditymayreachalevelofafewthousandNTU1[12].
The PNEN ISO 7027 standard [11] provides the
followingdefinitionofturbidity.Turbidityisadecrease
of liquid transparency caused by the presence of

1NTUNephelometricTurbidityUnit

Int. j. Using
water Artificial
sci., 2013,
Vol. 2,
3:2013
Adam Rak: Water Turbidity Modelling During Water Treatment Processes
Neural
Networks

undissolvedsubstances.Turbidityisalsodefinedas,An
abilitytoabsorbsunlightduetothepresenceinwaterof
veryfinepowder,suspendedsolidsorcolloidalsolidsof
mineral or organic source [11]. Turbidity is understood
as the optical properties of fine suspended solids in a
watersamplecausingsunlightdispersion[5].Turbidityis
one of the basic factors for assessing water quality.
Althoughitisnotcausedbyharmfulpollutantsanddoes
notinfluencepeopleshealth,itdegradestheaestheticsof
potable water. In accordance with a decree of the Polish
Minister of Health, the NTU level of potable water shall
not exceed 1, which is acceptable to consumers and
withoutabnormalchanges[4].

The turbidity of mountain river water depends on the

water level it is higher for higher water levels than for
lower ones. Turbidity is also affected by the bottom
sediments and the disturbance of suspended solids
transported in fast flowing water. It can be seen that an
increase in the turbidity of the inflowing river water
results in an increase in the turbidity of the water in the
reservoir. In general, the waters coming from rivers and
mountain torrents are of a relatively good quality, but
where the water is very turbid, reservoirs are used to
store it. Ideally, a system would include both water
storage and water storage and balancing reservoirs.
Dependingonthelocationofthewaterstoragereservoir,
bothrawwaterandtreatedwatercanberetained.Using
water storage reservoirs allows the water intake to be
closediftheturbidityincreasesoriftheincomingwateris
contaminated. Both of these conditions make the proper
treatmentofwaterdifficult.Waterisstoredinthereservoir
soitcanbeavailablefordifferentcircumstances,e.g.when
turbidity increases. If the reservoir is located before the
waterintakeinthesystem,thenamonitoringsystemmust
beinstalledtocontrolthewaterbeingtaken.Sucasystemis
used to detect contamination. Water storage reservoirs,
dependingontheirplacementinthesystem,canstoreraw
or treated water. Water storage and balancing reservoirs
mayalsobeusedtobalancewaterquality.

More and more often, integrated (hybrid) processes are

being used to treat water. In these hybrid processes,
standard water treatment technologies are combined
with unconventional methods resulting in higher levels
of contamination removal. An example would be a
system where coagulation is combined with membrane
filtration [1]. Despite the significant increase in the
development of these hybrid processes in recent years
and the methods of defining their process parameters
(coagulantdose,powerconsumption,rinsingmethods),
there are still many unresolved issues related to the
expected treated water quality. These depend on the
methodsappliedtocarryingouttheexperimentandthe
modelling method used [8]. Artificial neural networks
(ANNs)canbeappliedforthispurpose.

Int. j. water sci., 2013, Vol. 2, 3:2013

These mathematical models are adaptive systems that

change their structure based on internal or external
information that flows through the network during the
learningphases.Thesystemscanusealimitedamountof
data, which could be time consuming and expensive to
generate, to determine the quantitative and qualitative
parametersoftechnicalprocesses.Informationtransferin
the AANs imitates the human nervous system [16]. The
neurons from which the network is built are the data
processingunits.TheANNcalculatestheoutgoingvalues
onthebasisoftheinformationgivenattheincomingside
of the network. There are many types of ANN, but
currently the most popular kind is the multilayer
perceptron(MLP),withoneincominglayer,one(ormore)
hiddenlayer(s)andasingleoutgoinglayer.

Determining the number of hidden layers is not easy.

There are arbitrary and less arbitrary methods by which
to select them correctly. Typically, the best results are
obtainedwhenthesevaluesareselectedempirically[16].
After determining the number of layers, the weight and
threshold values for all the neurons should be selected.
Theyshouldbeselectedtoprovidetheminimumerrorof
network activity that gives the highest prediction. Most
often, the network error of the configuration of some
weightandthresholdparametersforthecorrectsolution
is assessed by comparing the output values with those
given as examples (learning data). The basis for
theautomatic modification of weights and ranges,
allowing for minimal error are collected examples (cases),
i.e., input data with a solution fraction of the correct
learningdatathataretypicallyhistoricalcases.Thesedata
are involved in the networks learning, a process which is
equivalenttoadjustingtheparametersofthemodelinthe
discussed cases. Prediction operations are made on the
basis of examples of input data of testing series and are
performedbysocalledlearningalgorithms.

The learning of neural networks is not a mechanistic

processandrequiresalotofconcentrationonthepartof
theexperimenter.Themaindifficultyistoavoidtoogreat
afitofthemodeltothedatasothatitdoesnotrepresent
all the details of individual cases. The data is in fact
certainly contaminated by various noises, so one cannot
expect that the best model is the one that will represent
eachcaseindividually.Onthecontrary,thebestmodelis
theonethatshowstheoverallshapeoftheapproximated
dependenceandavoidssmall(ofteninsignificant)defects
ofinputdata.Networkswithhigherweightsusuallyhave
a tendency to exaggerate the fitting of the model to the
data and in the literature this problem is called over
fitting. Such networks typically have small errors, but
poorpredictions.

TheattractivenessoftheANNapplicationisrelatedtothe
possibility of approximating any nonlinearity. It is not

www.intechopen.com

necessary to know the function describing the modelled

variabletocreatethenetworkmodel.Moreover,theANN
easily adapts to variable environmental conditions [3].
Modelling with an ANN application is recognized as a
black box type of approach and it is not possible to
determine a priori the optimal network architecture. The
trial and error method is used to achieve a suitable
network structure to solve the specific engineering
problem.

There are many known examples of mathematical

simulations which involve an ANN for predicting the
values of various parameters. Water treatment involving
membraneprocessesisonefieldinwhichmodellingwith
anANNisverypopular.ByusingANNmodelsbasedon
quantitative equations describing correlations between
processvariables,itispossibletoenvisagethelevelofthe
removal of natural compounds on polyamide
nanofiltrationmembranesandthroughareverseosmosis
process [17]. The level of removal of a substance from
topsoilandthephenomenonofultrafiltrationmembrane
blockingwerepredictedusinganapplicationofthemost
popular network learning algorithm, i.e., the method of
backward error propagation [2]. The Siwo et al. [15]
monograph presents a very interesting approach to the
subject of water collection prediction in which the
stochasticmethodsofwatercollectionfromcitymainsare
comparedwiththeresultsoftheANNmodels.Giventhe
high variability in the water collection time series, the
ANN turned out to be a very promising tool for short
term predictions. Modelling energy return from a heat
converterusedinawatertreatmentstationwasproposed
byHernadezetal.[6].Inthis,thethermodynamicmodel
was compared with the ANN model. Both models
provided sufficiently similar measurement and
simulationresults.

2.Descriptionoftheresearchobjective

In the 1990s, in order to supply the city of Jelenia Gra

withwater,theSosnwkawaterreservoirwasbuilt(Table
No 1). The Sosnwka reservoir stores waters from the
catchments of the Czerwonka and Soniak rivers as well
as from a small tributary, the Sosonwka. The overall
catchmentareais15.3km2[14].Thewatertreatmentplant
(WTP) has been in operation since 2007. Since then,
treatmenthasbeencarriedoutonacontinuousbasisusing
the following processes straining, primary ozonation,
coagulation, pH correction, flocculation, rapid filtration
through anthracitesand filters, secondary ozonation,
sorption on active carbon, final pH and water hardness
correctionanddisinfectionofthetreatedwater[13].

Water from Sosnwka reservoir is taken from two levels

and is transferred under natural flow to the separating
chamber through two 600 mm diameter pipes. From the

www.intechopen.com

chamber, the water is directed straight onto two rotary

screens, each with a mesh of 1mm x 1 mm. Drum
screens, 800 mm in diameter, are regularly rinsed with
waterandthescreeningsaredirectedtoasettlementtank
containing the washings. After passing through the
screens, the water flows into the two primary ozonation
chambers. Ozone generated from dried air is introduced
intothewaterusingporousdiffusers.Thecapacityofthe
chambers ensures a nominal time of contact, t, in the
ozonechamberofgreaterthan5min.Ozoneisgenerated
bythreesetsofgeneratorseachproducing1500gO3.h1.
From the primary ozonation chambers, the water is
transferredbygravitythroughan800mmdiameterpipe
viaastaticmixerintothereactionchamber.Coagulant,in
the form of a 14% solution of aluminium sulphate
[Al2(SO4)3], is introduced before water enters the static
mixer. Limewater is introduced into the same reaction
chamber through 80 mm diameter pipes, with valves to
control the lime doses depending on the water acidity.
Flocculent is added at the outflow zone of the chamber.
Then the water flows through rapid filters. Filtration is
carried out using twolayer anthracitesand beds. After
the filtration process the water is directed into the
secondary ozonation chamber and then onto the active
carbonfilters.

Following the active carbon filtration process, the water

passes into two filtrate tanks each of 210 m3 capacity.
Static mixers are mounted in pipes before the tanks.
Reagents (sodium carbonate and magnesium chloride)
are dosed in those mixers for final water quality
correction. The overall system in operation allows for
various combinations of the component technical
systems, depending on the quality of the water taken
duringthedifferentperiodsofthehydrologicalyear.

Takingintoaccountthequalitychangesintherawwater
stored in the reservoir, the various technical operational
systemsareseparatelypresentedinTableNo2.

Depending on the quality of the water taken from the

reservoir and its temperature, this system can be
modified. Thus in technical system W1A, there is no
secondary ozonation. Where the pH of the raw water is
high, the water treatment process can be carried out
withoutthepHcorrectionsystemW1B.

Technical system W2 involves the seven processes

indicated.WherethepHoftherawwaterislow,technical
system W3 is used. This is the same as system W2, but
withpHcorrectionadded

3.Researchmethodology

The following quality indicators of the raw and treated

waterintheSosnwkaWTParecontinuouslymonitored

Adam Rak: Water Turbidity Modelling During Water Treatment Processes Using Artificial Neural Networks

automatically temperature, pH, turbidity and

conductivity.Moreover,duringtheresearchperiod,from
11/2007to10/2008,physicalwateranalyseswerecarried
out every few days. The analysis measured such
parameters as water temperature, turbidity, colour, pH,
general hardness, basicity, iron, manganese, chloride,

Indicators

ammonia and nitrate nitrogen, oxidability, dissolved

oxygen, conductivity and phosphates. The quality
composition of the water taken for treatment was
balanced and typical for the specific season of the year.
Table No 3 below presents minimum, maximum and
averagevaluesofselectedqualityindicators.

Unit

Value

SurfaceofindirectcatchmentareaF1

km2

15,3

SurfaceofreplenishcatchmentareaF2

km2

35,5

Totalcatchmentareasurface

km

50,8

MaxsurfaceofcatchmentareaF

ha

178

Waterdepthatthedam

13,5

Averagewaterdepthatthereservoir

8,15

Averageannualflowofindirectcatchmentarea

QAV[m s ]

0,192

Lowaverageflow

QLAV[m3s1]

0,038

QB[m3s1]

0,046

Untouchableflow(biological)QB

Annualaverageoutflow(VAV)

m a

6054912

Annualuntouchableoutflow

m3a1

1821519

Waterintakeforconsumptionpurposes
QWTP(2008)
Planned water intake for consumption purposes
QWTP(plannedamount)
AppliedvolumeofthereservoirV

m h
m3a1
m3h1
m3a1
m3

380,0
3328800
1000,0
9000000
11000000

Indicators of water exchange per annum excl

waterintake;WW1
Waterintake9041m3d1;WW2
Plannedwaterintake24657m3d1;WW3
SchindlerindexWS1forF1
SchindlerindexWS2forF1+F2
AverageSchindlerindexWSaver.
Methodofcatchmentareamanagement

%
%
%
WS1
WS2
WAV
%forestcover

16,6
46,8
98,4
1,39
4,62
3,00
65

Table1.Basicindicatorsdescribingthereservoiranditscatchmentarea

Unitprocesses

Passagethrougha1mmx1mmscreen
Primaryozonation
Coagulationwithaluminiumsulphate
Flocculation
CorrectionofpHwithlimewater
Filtrationthroughanthracitesandbed
Secondaryozonation
Filtrationthroughactivecarbon
Finalcorrectionofwaterquality
Disinfection

W1
+
+
+
+
+
+
+
+
+
+

Technicalsystems
W1A
W1B
W2
+
+
+
+
+
+
+
+

+
+

+
+
+
+
+
+
+
+
+
+
+
+
+
+

W3
+
+

+
+
+
+
+
+

Table2.Unitprocessesoftheseparatetechnicalsystems

Int. j. water sci., 2013, Vol. 2, 3:2013

www.intechopen.com

Parametersofthe
analysedwater

Unit

Reservoir
SMax
S

SMin

Temperature
Turbidity
Colour
pH
Generalhardness
Ammonianitrogen
Nitratenitrogen
Chlorides
Oxidability
Conductivity
Dissolvedoxygen

C
NTU
mgPtdm3
pH
mvaldm3
mgNdm3
mgNdm3
mgCldm3
mgO2dm3
Scm1
mgO2dm3
o

4.0
1.0
5.0
7.2
0.62
0.01
2.0
4.0
1.0
98.28
9.4

22.0
12.0
14.9
8.4
1.03
0.11
5.0
6.8
5.22
122.50
12.4

11.7
5.2
8.21
7.63
0.85
0.05
2.44
4.95
3.458
110.88
11.18

6.12
2.47
2.791
0.366
0.072
0.004
1.300
1.321
1.532
6.692
1.172

Table3.Characteristicvaluesofselectedqualityparametersofthetreatedwater

Theresultsoftherawandtreatedwateranalysesallowed
forthelevelofreductionofthemonitoredparametersto
be determined. The possibility of providing continuous
monitoring allowed for analyses of the changes in
parameterssuchastemperature,pHandturbidityinone
daycyclestobecarriedout.

4.Researchresults

During the research period, water was treated following

the processes of chain W2 with a water flow of between
172and202m3h1,whichwasalsothespeedoffiltration
intheanthracitesandfilters.Thewaterflowthroughthe
active carbon filters was between 4.9 and 5.7 m3h1. The
waterreceivedthefollowingtreatments:
during primary ozonation, ozone between 1 and 2
mgO3dm3,
duringsecondaryozonation,ozone1mgO3dm3,
forthecorrectionoftreatedwater,sodiumcarbonate
1.0mgNa2CO3dm3,
for the stabilization of treated water, magnesium
chloride1.0mgMgCl2dm3.

Water quality during the research period was balanced

and typical for the specific season of the year. The water
temperature varied from 4oC during winter to 22oC
during summer. Rainfalldistribution wasdetermined on
the basis of IMGW data [7].At the time of the research,
theannualamountofrainfallforthecatchmentareawas
1011.8mm.

Theturbidityofthewatertakenfromthewaterreservoir
ranged from 1 to 12 NTU (Fig. 1), the colour from 5 to
14.9 mg Ptdm3, pH was in the range of 7.2 to 8.4 and
conductivity from 98 to 122 Scm1. A database of the
dailyvaluesofthemonitoredparameterswascreatedfor

www.intechopen.com

theentireperiodoftheresearch.Theeffectivenessofthe
reduction of the water turbidity was determined on the
basis of the raw water turbidity measurements taken
from the reservoir (Ms) and the treated water turbidity
measurements(Mp),asperthecorrelation[8]:

Rm 1

Mp
Ms

(1)

During the research period, the average water turbidity

reductionindex,Rm,was0.63,withastandarddeviation,
,of0.21.AccordingtothechangesinRm(aspresentedin
Fig. 2), it appears that the technical system was not
operating stably. This was particularly the case during
thaws and heavy daily rainfalls. The flow of water into
the reservoir from the mountainous catchment area
during thaws was about 1.5 ms1. The concomitant
turbidity was as high as 27 NTU. These conditions
resultedinadecreaseoftheRmto0.2.Atthesetimes,the
treated water turbidity was higher than 1 NTU,
periodicallyreaching4NTU.Itshouldalsobementioned
thatduringtheresearchperiodproblemsoccurredinthe
watertreatmentprocess.Thiswasnoticeableinthespring
when the turbidity of the treated water (4.09 NTU after
treatment) was higher than that of the raw water (3.0
NTU).

The results for the turbidity of the raw water and the
water treated using technical system W2 were compiled
daily in a database. The optimum prediction model has,
at any given moment, to be able to determine the
projectedeffectsofthewatertreatmentbasedon:
information coming from the monitoring of the
hydrologicalconditionsofthecatchmentarea,
meteorologicalobservations,
thephysicalandchemicalanalysesofthewater.

Adam Rak: Water Turbidity Modelling During Water Treatment Processes Using Artificial Neural Networks


Figure1.TheturbiditiesofrawwaterandtreatedwaterusingtheW2technicalsystemintheexaminationperiod(11/200710/2008)

Figure 2. Changes in the turbidity reduction index (Rm) compared with the daily rainfall in the catchment area in the examination
period(11/200710/2008)

The turbidity of the treated water is a function of the

meteorological phenomena occurring in the catchment
area during the period preceding the analyses and the
changes in the monitored indices of the water reservoir.
To model the changes in the treated water turbidity
achieved by technical system W2, a flexible Bayesian
model of neural networks, Gaussian processes and
mixtures that demonstrates Markov chains of Monte
Carlo methods was used. The model was executed in a
UNIX/Linux environment, version 19990313 [10].
Neural network models are described in detail in the
annex of Neal [9]. This programme is a tool for using
Bayesian models to predict parameters by applying
neuralnetworks.Italsoincludesapplicationssupporting
MarkovchainsofMonteCarlomethods(MCMC).Ittakes
samplesfromaseriesofvariables,usingsimpleformulas,
including Bayesian ones, for the previously assumed
probability. Neural networks software is used for
Bayesian models to learn about and predict results for
defined problems. The models are based on networks
6

Int. j. water sci., 2013, Vol. 2, 3:2013

with any number of hidden layers and with a prior

distributionofabroadselectionofparametersandhyper
parameters of the network. In contrast, the software for
Gaussian models supports the classification and
prediction of processes which are similar to the neural
networks with an infinite number of hidden units. The
advantage of the Bayesian learning, for both types of
models, is the automatic determination and control of
hyperparameters without the need for validation. This
avoids excessive adjustments involving large networks
anduncertaintyaboutthepredictionsobtained.

TheneuralanalysiswascarriedoutonanMLPmodel.The
structure of the MLP model was based on numerical
analyses in which the target variables were continuous
ones, such as treated water turbidity and five input
variables raw water turbidity, water flow to reservoir,
waterretentionlevel,dailyrainfallandtemperatureofthe
waterinthereservoir.Thenetworklearningwasbasedon
the collected historical data 366 fully described cases
www.intechopen.com

from1November2007to31October2008.Theverification
of each structure of the numeric model was carried out for
the same research data. The parameters of the network
architectureweredefinedforlevelsassuringthelowesterror
values in the predictions by controlling the rejection rate,
which was set at a level close to 0.5. The selected hyper
parametersoptimizedthenetworklearning processes.The
numerical simulation was carried out for 100 iterations
afterrejectingthefirst20%oftheburninsteps.

Inordertoevaluatetheassumedparametersoftheneural
networks model, the RMSE was determined using the
followingfunction:

1
N

RMSE

SP SR

(2)

where N is the number of data, SP the projected turbidity

valuesafterwatertreatmentandSRthemeasuredturbidity
values(observed)ofthetreatedwater.

The second evaluation criterion was to define the

correlation factor, R, between the projected and the
observed water turbidities. The value of the linear
correlation, on the basis of the nelement sample, was
calculatedfromtheformula:

S
i 1

S
i 1

Pi

S P S Ri S R
N

Pi S P S Ri S R

(3)

i 1

where S R istheobservedaveragevalueoftheturbidity
ofthetreatedwaterand S P thepredictedaveragevalue
oftheturbidityofthetreatedwater.

The correlation coefficient R expresses a linear

dependence between the two variables. The closer the
correlation coefficient is to 1, the stronger the linear
dependence. Usually, the following dependence of
thecorrelationcoefficientisadopted:
R<0,2nolineardependence,
R=0,20,4weakdependence,
R=0,40,7moderatedependence,
R=0,70,9strongdependence,
R>0,9verystrongdependence

The data sets for the quality of the raw water and the
water in the reservoir which had been used for the
analyses and learning were divided into oneweek time
series. A MLP neural networks model with a defined
architecture was created to predict the turbidity of the
treatedwater.Thedatafortheanalyseswerethetime(the
week of the measurement) and the daily rainfall; the
www.intechopen.com

amount and temperature of the water stored in the

reservoir each day (the explanatory variables) were
assigned.Thesevariableswereusedtocreateanonlinear
neuralregressionmodel.Therestrictionsimposedonthe
model consisted of limiting the conditions of the
predictablevarianttestsoftheexplanatoryvariablesand
defining upper and lower bounds for the projected
variables.(Itwasassumedthattherangeoftheprojected
conditions and the responses to these would not exceed
the measured values). The incoming layer was made up
of those neurons derived from the learning and tested
series. Adequate information for each model was
delivered to those neurons from all the explanatory
variablesintroducedintotheanalyses.Theoutgoinglayer
consisted of the modelled data of the projected turbidity
ofthetreatedwater.Thehiddenlayerswerecomposedof
hundredsofneuronlayerswithidenticalnumbersofcells
as in the incoming and outgoing layers. The number of
hidden layers was arbitrarily defined during the
optimization of the learning process. Worse prediction
possibilities were discovered for a different number of
hidden layers, after comparing the expected values against
the measured data. The remaining parameters of the
structural networks were determined during the networks
learning process, in accordance with the procedures for
programme operations using a Bayesian analysis of an
ANN. The iterative sequence was established for 100 steps
guaranteeing the energy stability of the model (Markov
chainequilibriumdistributionoftheMCMC).

TheMLPmodelgeneratedforpredictingtheturbidityof
the treated water was marked as per the specification:
netspec sun 5 100 20/ 5:50 5:50 5:50 100 modelspec
sun class dataspec sun 5 1 20/ metn2.txt@1:366.
metn2.txt@1:948846. The architecture of the assumed
neuralnetworkscanbedescribedasfollows:
oneincominglayerwithfiveexplanatoryvariables,
hundredsofhiddenlayers,
one outgoing layer with 20 units within the range
ofthedefinedvaluesbasedonthemeasureddata,
numberoflearningcycles366,
numberofeditedpredictions948,846.

For the polarization of the input data, a standard

deviation of 100 was assumed. The range of data in the
incominglayerwas5:50,inthehiddenlayers5:50andin
theoutgoinglayer5:50.

It was assumed that the neural network architecture for

the analysed time series was to be marked with the
followingsequence:
abbreviationofnetworkname:MLP,
numberofexplanatoryvariables:numberofincoming
variables,
numberofhiddenlayers,
numberofunitsfromtherangeoftheprojectedindex
numberofoutgoinglayers.

Adam Rak: Water Turbidity Modelling During Water Treatment Processes Using Artificial Neural Networks

Taking the above into account, the architecture of the

MLPmodelwasdescribedasfollows:MLP1:5100120.

Afteranalysingthedatasetwiththeneuralnetworkand
afterdiscoveringthebalanceofthesimulatedvaluesand
the high compatibility of the expected values with the
values observed in the next step, a simulation of the
occurrence probability of specific water turbidities was
performed for specific combinations defined within the
range of the observed variables. The results of the
effectiveness analyses of the MLP model for the projected
turbidityofthetreatedwaterarepresentedinFigs.3and4.

5.Discussionofresearchresults

As a result of developments in information technology

systems,itisnowpossibletomeasureprocesstechnology
dataandrecord,visualize,presentandarchivetheminan
electronicformthankstothedevelopmentofsupervisory
controlanddataacquisition(SCADA)typesoftware.This
system has become very popular for monitoring and
controlling processes in WTPs.A SCADA system allows
for the automatic control of process equipment and the
registration and archiving of information coming from
measuring devices. Ongoing analyses of changes in
controlledparametersallowprocessdecisionstobemade
well in advance. Such decisions have positive effects on
theeconomicaspectsoftheprocess.TheSCADAsystem
mustbeequippedwithpredictionfunctions,whichhave
someinfluenceonthechoiceoftheoptimumexploitation
methodforawatertreatmentsystem.

Fromtheresearchandanalysescarriedout,itcanbeseen
thattheproposedmodel,basedonanANN,forpredicting
oneparameterofthetreatedwatercansupportthecontrol
ofawatertreatmentprocessinaWTP.Beingabletopredict

thevalueoftheturbidityofthewaterallowsforadvanced
technical and logistical actions to be taken to change the
method of using the process equipment in the WTP. For
example, it allows for additional unit processes, such as
coagulation, and changes in the chemical agents applied
in the water treatment process to be made that is, to
switchfromsystemW2tosystemW1.

On the basis of the results obtained, it appears that the

turbidity of the water treated over a period of 261 days
exceeded the value of 1 NTU (Fig. 5). The technical
system tested, W2, ensures the achievement of a level of
water turbidity below 1 NTU only for a period of 105
days.Itisapparentfromthisthatanadditionaltechnical
system,W1,isrequired.

This study presents analyses of possible applications of an

ANN type MLP model for predicting one of the quality
indicators of treated water. For the numerical analyses, the
regressive MLP neural network model was applied. The
continuous dependent variable was the turbidity of the
treated water and there were five explanatory variables
raw water turbidity, water level in the reservoir (retention
level),dailyrainfallinthecatchmentarea,volumeofwater
flowing into the reservoir and the reservoir water
temperature.Verificationofthenumericalmodelwascarried
out using the same sets of results as obtained from the
research.Theparametersofthearchitectureweredefinedto
assurethelowesterrorvaluesachievedinprediction.

Thereweretwocriteriaassumedtoselectthefinalneural
network parameters. The first was the RMSE and the
second was the correlation factor, R, between the
projected and observed values of the treated water
turbidity. The calculated value of RMSE was 0.49 NTU
andthatforRwas0.84.

Figure3.CorrelationbetweenprojectedandobservedvaluesoftheturbidityofwatertreatedusingtechnicalsystemW2
8

Int. j. water sci., 2013, Vol. 2, 3:2013

www.intechopen.com


Figure 4. Changes in the measured turbidities of the raw and treated water, and the projected turbidity of the treated water in the
examinationperiod(11/200710/2008)

Figure5.Dailyvaluesofturbidityofthetreatedwaterintheexaminationperiod(11/200710/2008)

Figure6.DiagramofthestepsforpredictingwaterqualityindicatorswithanANNapplication

www.intechopen.com

Adam Rak: Water Turbidity Modelling During Water Treatment Processes Using Artificial Neural Networks

The effectiveness of the neural network in predicting

water contamination indicators in the reservoir was
relativelygood.Fromtheanalysescarriedout,itappears
that appropriate definition of the parameters
characterizing the water inflow (the explanatory
variables) increases the accuracy of the projected water
quality indicators. The selected model complements the
algorithmoftheprojectedinformaticssystemsaswellas
those of the modernized systems, which are applied to
manage and control the water intake and treatment
processes of water used for drinking, industrial and
agriculturalpurposes.Creatingamodeltopredictquality
indicators requires research and analyses to be carried
out for each individual objective of the water
management. The result of this research was the
determination of the various water quality indicators
which define the characteristics of the tested water and
determine the expected water parameters. Predicting the
selected water quality indicators using neural networks
was carried out in stages. A flowchart of the process is
presentedinFig.6.Asaresultofresearchandanalysis,a
turbidity prediction model was proposed as one of the
indicators of water quality to be evaluated in the water
treatment process. The proposed method can be used to
project other indicators, such as water colour and pH.
This model can be helpfulin projecting additional
information algorithms for the management system of
water management processes in reservoirs and water
treatmentforspecifiedpurposes.

6.References

[1] Adamski W., MajewskaNowak K. (2010).

Application of multifunction reactors for water
purification.Ochronarodowiska,32(1),38.
[2] AlAbriM.,HilalN.(2008).Artificialneuralnetwork
simulationofcombinedhumicsubstancecoagulation
and membrane filtration. Chemical Engineering
Journal,141(1),2734.
[3]Cieak W., Siwo Z., Cieak J. (2006).Application for
the prediction of time series of shortlived water
intake in selected water mains. Ochrona rodowiska,
28(1),3944.
[4]Decree.(2010).DecreeoftheMinisterofHealthdated
20thApril2010.Relatingtoqualityofwaterintended
forconsumptionbyhumans.
[5] Hermanowicz W., Doaska W., Dojlido J.,
KoziorowskiB.(1976).Physicalandchemicaltesting
ofwaterandsewage.Arkady,Warsaw,Poland,8592.
[6] Hernandez J., Romero R., Juarez D., Escobar R.,
Siqueiros J. (2009). A neural network approach and
thermodynamicmodelofwasteenergyrecoveryina
heat transformer in a water purification process.
Desalination,243,273285.

10 Int. j. water sci., 2013, Vol. 2, 3:2013

[7] IMGW. (2009). Instytut Meteorologii i Gospodarki

Wodnej (IMGW) o/Wrocaw. Meteorology data; total
daily rainfalls measured at the station located in
Karpaczfortheperiod01.11.200731.10.2008.
[8]KabschKorbutowicz M., Kutyowska M. (2010).
Application of artificial neural networks for
predicting of water turbidity changes in
coagulation/ultrafiltration
process.
Ochrona
rodowiska,32(2),1520.
[9] NealR.(1996).BayesianLearningforNeuralNetworks.
SpringerVerlag,NewYork,(Availableat

http://www.cs.utoronto.ca/~carl/).
[10]Neal R. (1999). Flexible Bayesian models on neural
networks, Gaussian processes, and mixtures v 2000
0813. University of Toronto, Canada, pp. 182.
(Availableathttp://www.cs.utoronto.ca/~radford/).
[11]Norma PNG01300. (1997). Standard PNG
01300:1997 Hydrogeology. Basic terms and
definitions (withdrawn). Polish Committee for
Standardization(PKN),Warsaw,Poland,
[12]Paweek J. (1993). Turbidity of water in rivers and
mountain torrents from the perspective of water
treatment for water mains purposes. Ochrona
rodowiska,51(4),6973.
[13]Rak A. (2010). Prediction of criterion water quality
indicators in reservoir for water treatment plant. In:
Supply of water, water quality and protection modern
subjects.Vol.I.PublishedbyWydawnictwoISBN97883
89696339,PZITSo/Pozna,2010,Poland,pp.4360.
[14]RakA.,KonarczakT.(2008).Pilotstudiesofmethods
of determining surface water treatment technology
typical for water from mountain catchment areas
storedinawaterreservoir.WaterSupplyandWater
Quality.PublishedbyPZITS641657.
[15]SiwoZ.,omotowskiJ.CieakW.,LicznarP.,Cieak
J.(2008).Analysisandpredictionofwaterdistribution
in water mains. Polska Akademia Nauk, Komitet
Inynierii Ldowej i Wodnej, Instytut Podstawowych
ProblemwTechniki,Warsaw,Poland.
[16]TadeusiewiczR.(1993).Neuralnetworks.Akademicka
OficynaWydawniczaRM,Warsaw,Poland.
[17]YangaliGuintanillaV.,VerliefdeA.,KimT.,Sadmani
A, Kennedy M., Amy G. (2009). Artificial neural
network models based on QSAR for predicting
rejection of neutral organic compounds by
polyamide nanofiltration and reverse osmosis
membranes.JournalofMembraneScience,342,251262.

www.intechopen.com