基于支持向量機的非線性時間序列預(yù)測方法研究

1、江南大學(xué)碩士學(xué)位論文基于支持向量機的非線性時間序列預(yù)測方法研究姓名：王洪波申請學(xué)位級別：碩士專業(yè)：檢測技術(shù)與自動化裝置指導(dǎo)教師：崔寶同朱啟兵20080601AbstractAbstractAsthemaintechniqueofthequantitativeforecast，themethodoftimeseriespredictionisusedalmostinallfieldsofforecastThetraditionalmeth

2、odsoftimeseriespredictionmainlyfocusonlineartimeseriesandweaknonlineartimeseries，SOtheylackeffectivenesswhentheyfacecomplexnonlineartimeseries(evenchaostimeseries)SupportVectorMachine(SVM)isakindofnovelmachinelearningmet

3、hods，theoreticallybasedonstatisticlearningtheoryItemploysthecriteriaofstructuralriskminimizationandprovidesaframeworkforthesmallsamples，nonlinearityandhighdimensionproblemsSVMhasbeenappliedinnonlineartimesedespredictioni

4、nrecentyearsFocusingonraisingaccuracyofSVMpredictionmodel，thispaperstudiesnoise—reduction，nonstationaryprocessingandmodelparametersoptimizationofnonlineartimeseriesAlthougllwidelyappliedinnonlineartimeseriesanalysis，then

5、oisereductionmethodvialocalprojectionhasthesubjectivityofselectingtheneighborhood，whichgreatlyaffectstheperformanceAnewmethodbylocalprojectionusingadaptiveneighborhoodselectionisstudiedFirst，onedimensionaltimeseriesareem

6、beddedintoahi曲dimensionalphasespaceaccordingtotime—delaytheoryTheneighborhoodsizeforeachcandidatephasepointinphasespaceisincreasedbyaddingneighboringpointonebyoneTheoptimalneighborhoodsizeforthephasepointisdeterminedduri

7、ngthedirection’Svariationofthemostsignificanteigenvectorofneighborhoodassizeincreasing，andthenthenoiseiseliminatedthroughlocalgeometricprojectionExperimentresultsshowthatadaptiveneighborhoodselectioncanimprovethenoisered

8、uctionperformanceoflocalprojectionmethodNon。stationarytimeserieshasperiodicityandrandomnessSOthatitisdifficulttoconstructthemodelofaccurateforecastwithsinglemethodAhybridforecastingmethodbasedonEmpiricalModeDecomposition

9、(EMD)andLeastSquareSupportVectorMachine(LSSVM)ispresentedinthispaperFirstlythenon—stationarytimeseriesisadaptivelydecomposedintoaseriesofstationaryintrinsicmodefunctions(IM聊indifferentscalespaceusingEMDThentherightparame

10、terandkernelfunctionsarechosentobuilddifferentLSSVMrespectivelytoeveryIMEFinallytheseforecastingresultsofeachIMFarecombinedtoobtainfinalforecastingresultThismethodissuccessfullyappliedtononstationarytrendpredictionofmech

11、anicalvibrationPhasespacereconstructionandSVMparametersoptimizationaretwoimportantaspectsforimprovingpredictionperformanceandaresolvedseparatelytraditionallyThispaperproposesajointoptimizationalgorithmbasedonHybridPSOThi

12、smethodcombinesthediscretePSOwiththecontinuousvaluedPSOtosimultaneouslyoptimizethephasespacereconstructionandtheSVMparameterssettingTheexperimentalresultsshowedtheproposedapproachcanraisepredictionaccuracyKeywords：Nonlin