ProcediaEngineering70(2014)399–408
12thInternationalConferenceonComputingandControlfortheWaterIndustry,CCWI2013Usingartificialneuralnetworkmodelstoassesswaterqualityinwaterdistributionnetworks
AbstractThepurposeoftheresearchistoassesschlorineconcentrationinWDSusingstatisticalmodelsbasedonANNincombinationwithMonte-Carlo.Thisapproachoffersadvantagesincontrasttothegenerallyusemethodsformodelingofchlorinedecayindrinkingwatersystemsuntilnow.Themodelwastestedononespecificlocationusingthehydraulicandwaterqualityparameterssuchasflow,pH,temperature,etc.Themodelallowsforecastingchlorineconcentrationatselectednodesofthewatersupplysystem.TheresultsobtainedintheseselectednodesallowthentocomparethechlorineconcentrationwithEPANETinthesystemunderassessment
Keywords:WaterDistributionSystemsChlorinedecayArtificialNeuralNetworkMonteCarloMethod
1.Introduction
ThepurposeofaWaterDistributionSystem(WDS)istomakewateravailabletocustomerswithatleastacceptablepressure,flow,continuityandwaterquality.Waterqualitycanbemeasureintermsoftaste,odor,appearanceandchlorineconcentrationbetweenothersparameters.MaintainingwaterqualitythroughtheWDSuntilthepointofconsumptionisoneofthemostchallengingtaskfacedbythewaterutilityindustries(Clementet.al.,2004),takinginconsiderationthecomponentsoftheWDS,suchaspipematerials,tanks,valvesetc.andotherrisksrelatedtowaterdistribution.InsidetheWaterTreatmentPlant(WTP)thereisacombinationofprocessesfor
drinkingwatertreatment.PrincipalprocessesofaconventionalWTPinclude;aeration,coagulation,sedimentation,filtrationanddisinfection.Adisinfectantresidualshouldbemaintainedthroughoutthedistributionsystematalltimes..Althoughitisrecognizedthatexcessivelevelsofdisinfectantmayresultintasteandodorproblems,itisthereforerecommendedthatadisinfectantresidualbemaintainedandmonitoreddailythroughouttheentiresystem.Themostcommon
disinfectantusedinwatermanagementischlorine.TheconceptofResidualChlorineConcentrationisassociatedwithdisinfectiondurability.Thereis,however,anotherproblemregardingdisinfectioninaWDS.Itisaphenomenonknownaschlorinedecay;chlorinereactswithothercomponentsalongthesystemanditsconcentrationdecrease(CastroandNeves,2003).Knowingthephysic-chemicalaspectsbehindchlorinedecayisimportantinordertodevelopastrategycapableofdisinfectingaWDSandatthesametime,preservingwaterqualityuntilthepointofuse,withoutusingmoredisinfectantthannecessary(Bowdenet.al.,2006).TheobjectiveofthisresearchistodevelopanArtificialNeuralNetwork(ANN)modelincombinationwithMonteCarloMethodthatcansimulateresidualchlorinedecayatselectednodesunderaspecialzoneoftheWDSinthedistrictofBrno-Kohoutovice,CzechRepublic,withtheadvantagesofasimplefunctionalformandgoodaccuracy.Inaddition,itcanalsobeemployedtoestimateresidualchlorinedecayintherestofpointsinsidethenetworkandremarkthe
affectedareasinwhichhighorlowlevelsofchlorinearepresentedinthesystembyusingthecomputationalmodelEPANET2.0.TheANNmodelwasdevelopedandtestedfortheselectedpressurezone1.3.2ZemníVDJKohoutovice.2.Methods
DifferenttopologiesoffeedforwardANNsusingthebackpropagationlearningalgorithmwerestudiedtoapproachthebehaviorofchlorinedecayforvaryinglevelsofchlorineresidualinseveralnodesinsidethewaterdistributionsystemofthedistrictofBrno-Kohoutovice,inaddition,somephysic-chemicalinputparameters(e.g.pH,Temperature,turbidityandflow)werealsoassessedastheycanaffectchlorinedecay.ThesystematicmodelingprocedureproposedandimplementedinthisresearchcanbeseeninFig.1.ThemainstepsformodelingchlorineresidualinaWDSusingANNinvolve:Datapreparation,Inputselection,MonteCarlosimulationformissingvalues,Datadivisionandselectionofsubsets,Modelcreation,ModelcalibrationandPerformanceevaluation
AnotherimportantstepforthecreationofANNmodelsforchlorinedecaypredictioninaWDSisthecreationofthehydraulicmodel.Thehydraulicmodelwasalsocalibratedinthebestpossiblewaysincethewaterqualitystudiesrequiredhighlevelcalibrationtoavoidmisleadingdataorerrorintheresearch(RossmanL.1999).
ANNuseshistoricaldataforpredictionofparameters.WhencreatingANNmodels,somedatamaybemissingfromtheoriginaldatabase.Modelersusuallyreplacethemissingdatawiththeaverageofthesampleorsimplydeleteorignorethecompleterow,causingthelossofimportantdata.TheMonteCarlo(MC)methodcanbeusedtogenerateadatabaseofeachparametersaffectingchlorinedecayintheseveralnodesstudiedinsidetheWDS.MonteCarlosimulationcanbeperformedtofulfillthemissingvalues(ifany)intheoriginaldatabase,asitprovideflexibility,managetheuncertaintyandevenprovidemoreaccurateresultsthatsimpledescriptivestatistics(e.g.theaveragevalue).Theobjectiveistocreateabigdatabaseforeachinputandoutputparameter.TheresultsobtainedfromMCmethodcanbeagainanalyzedwithdescriptivestatistics(average,standarddeviationandconfidenceinterval).Thisanalysiscanbedoneagainforeachinputandoutparameterofthemodel.
2.1.DatapreparationforANNmodelsandhydraulicmodelHistoricaldataofseveralparametersthatinfluencechlorinedecayshouldbegatheredtobesuccessfulintheapplicationofANNmodels.UsuallyutilitiesmeasureparameterssuchaspH,temperature,turbidity,color,manganese,iron,conductivity,e.coli,coliformbacteriaetc.Caremustbetakeninhowthedatawascollected,theformatgiven,thesupplierofthedataanditmustbeverifiedwhethertheinformationiscurrentandaccurate.Seasonsalsoinfluenceinthevaluedata,sotheapproachcanbetakenintwoways.Firstdividethedataforeachseasonandrunthestudyfordifferentseasonsasineachseason,thevariablessuchastemperatureandpHhave
401
greatchangesorthesecondway,createabigdatabaseandruntheanalysisforallthedataavailable.ANNwillfindtherelationshipbetweenthechanges,butabigamountofdataisrequiredtogetbetterresults.
2.2.SelectionofinputsandoutputsfortheANNmodelSuccessfulapplicationofartificialneuralnetworkmodelrequiresproperinputdataselection.ThebetterwaytochoosetheinputsfortheANNmodelistominimizethesizeofthenetworkandatthesametimemaintainingacceptableperformance(LingireddyandBrion,2005).Thebetterwaytochoosetheinputsinpracticeisrelatedtotwoprimaryconsiderations:
0Priorknowledgeabouttheprocess
0Availabilityandqualityoftherequireddatainthetrainingset
Recentstudies(Gibbset.al.2006,RaoandAlvarruiz,2007)haveshownphysic-chemistryparameterssuchas
temperature,pH,turbidity,andnaturalorganicmatter(NOM)betweenotherthat
affectthechlorinedecay,alsohydraulicparametersasflow,pressureandpropertiesofthepipelineaspipematerial,diameterandageofpipesshouldalsotakeintoaccountasaninfluencetothechlorinedecay.
2.3.ConstructionoftheinputdatabaseusingMonte-CarlomethodInitialChlorine,pH,piperoughness,TurbiditybetweenothersparametersoftheinitialconditionsofthewaterdistributionsystemweregeneratedbyMonteCarlosimulationtechniquethatrequirestheuseofrandomnumbergenerator.Itgeneratesthenumbersthatfollowanormaldistributionanduniformdistributiondependingoftheparametersimulated.Generatedfactorsarethentobecomparedwiththeactualfactortocheckthesignificanceofthesimulatedparameterssuchaschlorineaddedtowater,pH,flow,turbidity,temperature,etc.ForanalyzingthefactorsaffectingchlorinedecaythegeneralMonteCarlostepsaremodifiedasgivenbelow:
0Domainofpossibleinputs–VariesfromminimumtomaximumvaluesofChlorineaddedtowater(initialchlorine),Flow,pH,Temperature,Turbidity,asperhistoricaldatareceivedfromwaterutility.
0RandomNumberGenerator–TheSoftwareSTATISTICA10willbeusedtogeneraterandomnumberswithinthedomain.
ArtificialNeuralNetwork–ToaggregatetheresultsArtificialNeuralNetworkisplottedinordertocalculatethechlorinedecay
Aninitialsimulationwasperformedstartingwiththehydraulicparametersfollowedbythephysic-chemicalparameters.Anormaldistributionwasfollowforthegenerationoftherandomnumber.Fortheanalysisofthemeasureddata,thesoftwareStatistica10fromStatsoftusesafunctioncalledDistributionFitting,thisoptionallowsverifyingwhetherthemeasurevaluesfollowanormaldistributionandaftertheconfirmationitwasrunasimulationusingtheMonte-CarloMethodproposed
3.Resultsanddiscussion
ThegoalistodeterminethefactorsinfluencingchlorinedecayforthecasestudyinthepressurezoneofBrno,Kohoutovice.Thesefactorsarebasedonlocalmeasurementsofresidualchlorine.FordeterminingfactorsaffectingchlorinedecayinWDS,underdifferentparameterconditions,valuesofhistoricaldataarerequired.Initial
Chlorine,pH,Flow,TemperatureandturbiditybetweenotherfactorsareusedinthisstudyasinputforforecastingchlorinedecayinBrno,Kohoutovice-pressurezone,CzechRepublicCaseStudy.
3.1.ResultsfromthehydraulicmodelcalibrationThebasemodelhadnotbeenproperlycalibratedbutinthisstudyitwasperformed
calibrationandqualitycontrolreviewoftheresultingmodel.Thismodelcoveredonly1hourintimebuttheroughnesscoefficientsanddemandswerealreadycalculated.A
quasi-dynamicanalysiswasperformed,i.e.thesimulationtookplaceinthetimestep-inthiscasewas1hourtimestepfortwodays(48hrs)usingEPANET2.0.SeeFig.2(b).Themostcommonparameterstoanalyzeandmeasureforhydraulicmodelcalibrationare;pressure,flowandwaterlevelintanks(Rossman,1999).4.Conclusion
TheuseofANNforevaluationofhistoricaldataandchlorinedecaypredictionwasassessedinthreepointsinsidethepressurezoneinKohoutovice,CzechRepublic.
Initialchlorine,pH,flow,turbidityandresidualchlorineinthreenodesinsidethepressurezonewereusedasdatasetinthismodel.Becausepriorinformationisavailableaboutsomeparameters,thenthisinformationwasusedtoassesschlorineresidualinNodes1,2and3.Continualdistributionsareusedtoquantify,evaluateandfittheseparameters.Basicstatisticssuchasmean,standarddeviationandconfidenceintervalswereusedforevaluationoftheseparameterscalculatedwithMonteCarlomethod.3000readingsweregenerated.TheuseoftheMonteCarlocalculationsincombinationwithartificialneuralnetworkhavebeenproventobeapowerfultooltoperformchlorineresidualpredictioninthreenodesinsidetheKohoutovicepressurezone.SeveralANNtopologieshaverespondedaccuratelytothevaluesofthetrainingdatasetandalsototestingandvalidationdatasetofvaluecalculatedusingtheMCmethod.Networktopologies(Subset1)canbefullyusedtopredictthevaluesofchlorineresidual
concentrationinNode1.Resultsshowthattheproblemdealtwithinthispaperisindeedverycomplexbecausechlorinedecayisdependingonseveralparameterssuchastemperature,pH,turbiditybetweenothers.FortrainingtheANNperformedwasveryaccurate,eventhoughitcouldnotbeachievedthesameresultsinSubsets4to7in
testingandvalidationphase,asalreadywrittenintheevaluationofthedataobtainedsection.Therefore,thetrainingdatasetshouldbechosentobestrepresentthe
expectedchlorineresidualintheselectedNodesinsidethepressurezoneKohoutovice.Thepresentfindingssuggestthat:
•ANNsiscapabletopredictfreechlorineatKohoutovicepressurezoneusinghistoricaldataanddatageneratedbyMCmethod.
•ThekeyparametersInitialchlorine,flowandtemperaturehavethemostinfluenceinthechlorinedecaypredictioninthepressurezone1.3.1Kohoutovice.Recommendationsfortheuseofthismodelarefollowing:
•ThepresentmodelcanonlybeusedforchlorinedecaypredictioninKohoutovicepressurezone.
•Thismethodologycanbeusetoassesschlorinedecaybehaviorindifferentpressurezonesofanywater
distributionsystem,followingtheprocedureshowninFig.1.
•Creationofacompletedatabasebasedonthemeasurementsoftheparametersaffectingchlorinedecayinthe
sameplacesinsidethepressurezone.
•Temperaturemeasuredinseverallocationshasahighinfluenceinchlorinedecay.Idealcalibrationshouldtake
intoaccountthebehaviorofchlorineatdifferenttemperaturevaluesindifferentseason.
•Createalargedatabasewhichincludeparametersfromalltheseasonandstudydifferentmodelsubsetforeachseason.
References
ClementJ.,PowellJ.,BrandtM.2004,Predectivemodelsforwaterqualityindistributionsystems.IWAPublishing,pp.106
BowdenG.,NixonJ.,DandyG.,MaierH.,HolmesM.2006,Forecastingchlorineresidualsinawaterdistributionsystemusingageneralregressionneuralnetwork.MathematicalandComputerModelling.44,469-484.
LingireddyS.,BrionG.,2005,Artificialneuralnetworksinwatersupplyengineering.AmericanSocietyofCivilEngineers,pp173.
Rao,Z.,AlvarruizF.,2007,Useofanartificialneuralnetworktocapturethedomainknowledgeofaconventionalhydraulicsimulationmodel.JournalofHydroinformatics.9,15-24.
AbdiH.,ValentinD.,EdelmanB.,1999,Neuralnetworks.ThousandOaks,California.Sage.PowellJ.,HallamN.,WestJ.,ForsterC.,SimmsJ.2000,Factorswhichcontrolbulkchlorinedecayrates.WaterResearch.34,117-126
GibbsM.,MorganN.,MaierH.,DandyG.,NixonJ.,HolmesM.2006,Investigationintotherelationshipbetweenchlorinedecayandwaterdistributionparametersusingdatadrivenmethods.MathematicalandComputerModelling.44,485-498.
RodriguezM.,SérodesJ.1998,Assessingempiricallinearandnon-linearmodellingofresidualchlorineinurbandrinkingwatersystems.EnvironmentalModelling.14,93-102.MayR.,DandyG.,MaierH.,NixonJ.2008,ApplicationofpartialmutualinformationvariableselectiontoANNforecastingofwaterqualityinwaterdistributionsystems.EnvironmentalModelling.23,1289-1299.
VasconcelosJ.,BoulosP.1996,Characterizationandmodelingofchlorinedecayindistributionsystems.AmericanWaterWorksAssociation,pp.402.
RossmanL.,MaysL.,(ed.).1999,WaterDistributionSystemsHandbook.ComputerModels/Epanet,Chapter12.McGraw-Hill,NewYork.Pp.12.1-12.23
CastroP.,NevesM.2003,Chlorinedecayinwaterdistributionsystemscasestudy,
LousadaNetwork.ElectronicJournalofEnvironmental,AgriculturalandFoodChemistry.2,261-266.
Izquierdo,J.,PérezR.,IglesiasP.2004,Mathematicalmodelsandmethodsinthewaterindustry.MathematicalandComputerModelling.39,1353-1374.
因篇幅问题不能全部显示，请点此查看更多更全内容