基于聚类神经网络的西南季风降雨长期预报(IJITCS-V6-N7-1)

I.J.InformationTechnologyandComputerScience,2014,07,1-8PublishedOnlineJune2014inMECS()DOI:10.5815/ijitcs.2014.07.01Copyright©2014MECSI.J.InformationTechnologyandComputerScience,2014,07,1-8LongRangeForecastonSouthWestMonsoonRainfallusingArtificialNeuralNetworksbasedonClusteringApproachMayaL.PaiDepartmentofMathematics,AmritaVishwaVidyapeetham,Cochin,682024,IndiaEmail:[e-mail]erApplications,CUSAT,Cochin,682022,IndiaEmail:[e-mail]icalOceanography,CUSAT,Cochin,682016,IndiaEmail:[e-mail]—ThepurposeofthisstudyistoforecastSouthwestIndianMonsoonrainfallbasedonseasurfacetemperature,sealevelpressure,humidityandzonal(u)andmeridional(v)winds.WiththeaforementionedparametersgivenasinputtoanArtificialNeuralNetwork(ANN),therainfallwithin10x10gridsofsouthwestIndianregionsispredictedbymeansofoneofthemostefficientclusteringmethods,namelytheKohonenSelf-OrganizingMaps(SOM).TheANNistrainedwithinputparametersspanningfor36years(1960-1995)andtestedandvalidatedforaperiodof9years(1996-2004).Itisfurtherusedtopredicttherainfallfor6years(2005-2010).TheresultsshowreasonablygoodaccuracyforthesummermonsoonperiodsJune,July,AugustandSeptember(JJAS)ofthevalidationyears.IndexTerms—SouthWestMonsoon,Clustering,ArtificialNeuralNetworks,Self-OrganizingMap.I.INTRODUCTIONMonsoonisanoutstandingtropicalphenomenonoftheIndiansub-continent.TheforecastoftheIndiansummermonsoonrainfall(JJAS)hasbeenverycrucialandadvantageousforfarmers.Thesouth-westwinds(oftenknownasthesouthwestmonsoon)blowingfromtheIndianOceanontotheIndianlandmassduringthemonthsofJunethroughSeptemberisgenerallyrainbearingwindsthatbringrainfalltomostpartsofthesubcontinent.Theysplitintotwobranches,namely,theArabianSeaBranchandtheBayofBengalBranchnearthesouthernmostendoftheIndianPeninsula.Ontheotherhand,oceanographically,theIndianOceanistheleastexploredofthemajoroceans.TheIndianmonsoondependsonmanypre-monsoonfactorsoftheIndianOcean[1-3].TheempiricalforecastingoftheIndianmonsoonhasbeenachievedusingacombinationofclimaticparameters,includingtheatmosphericpressure,thewind,theSeaSurfaceTemperature(SST),thesnowcoverandthephaseoftheElNiño–SouthernOscillationENSO[4,5].Regressionmodelsbasedontheseandotherempiricalcorrelationshavebeenabletopredict60%–80%ofthetotalseasonalIndianrainfallbythemonthofMayprecedingthesummermonsoon[6].SSThasbeenrecognizedasanimportantoceanicparameterbecauseitdirectlyinfluencestheair-seaexchangeofheat.TheotherparametersofinterestaretheSeaLevelPressure(SLP),theHumidity,andtheU-andV-winds.TheinitialworkonAsianmonsoonpredictionwasconductedbyWalker[7]followedbyseveralattempts[8-10]leadingtothedevelopmentofbettermodelstowardslongrangeforecastofsummermonsoonrainfalloverIndia.Forinstance,theparametricandpowerregressionmodelsused[10]gavereasonablyaccurateresults.ThesemodelsareusedbytheIndiaMeteorologicalDepartment(IMD)forlongrangeforecastsforIndia.Butthesestatisticalmodelshavesomelimitations.Soattemptsweremadetodevelopbetter,alternatetechniquesforlongrangeforecastsofIndianSummerMonsoonRainfall(ISMR).The8parameterHybridPrincipalComponentModelwasdeveloped[11]byusinga30-year(1958-87)trainingperiodanda10-year(1988-97)verificationperiod.AnartificialintelligenceapproachforregionalrainfallforecastingforOrissa(Indianstate),onmonthlyandseasonaltimescaleswasattempted[1].Inthatstudy,thepossiblerelationbetweenregionalrainfalloverOrissaandthelargescaleclimateindiceslikeEL-NiñoSouthernoscillation(ENSO),EquatorialIndianOceanOscillation(EQUINOO)andalocalclimateindexofocean–landtemperaturecontrastwerefirststudiedandthenusedtoforecastmonsoonrainfall.ThetimeseriesofallIndiasummermonsoonrainfallwasgeneratedbyareaweightingtherainfallat306raingaugesacrossthecountry[12,13]andtheempiricalmodelingapproacheswereusedtoforecastISMR.[14]givesageneraloverviewofforecastingmodelsforISMR.Later,[15]and[16]presentedreviewsonsuchempiricalmodels.2LongRangeForecastonSouthWestMonsoonRainfallusingArtificialNeuralNetworksbasedonClusteringApproachCopyright©2014MECSI.J.InformationTechnologyandComputerScience,2014,07,1-8Theauthors[17]exploretherecentapplicationsofNeuralNetworks(NN)andArtificialIntelligence(AI)andprovidesanoverviewofthefield;theyhavediscussedthecriticalroleofNNandAIindifferentareas.Neuralnetworksthatmodelcomplexrelationshipsbetweeninputsandoutputsortofindpatternsindatahavebeenemployedinthepastforrainfallprediction.ANNshavethecapabilityofcapturingcomplexnon-linearityintimeseriesandalsoinprediction.Inthiscontext,theusefulnessofNNforrealtimenumeralrecognitioninvolved150onlinenumeralswrittenindifferentstylesby10differentpersonsandobtainedaccuracyrangingfrom97%to100%fordifferentresolutionoftheinputvectors[18].Ofcourse,NNtechniquelearnsthedynamicswithinthetimeseriesdata[19].Intheearlytwentiethcentury,ANN’swereusedtopredictISMR[19-21].Thetimeseriesapproachwasusedtopredictfuturevaluesby[19].Whereasin[20],theauthorshavepredictedtheIndianmonsoonrainfallwiththehelpofsomepredictorsandcompa