Page 74 - 2022年第53卷第12期
P. 74
46(3):1099 - 1108.
[21] LIJ,SONGSS,KANGY,etal.Predictionofurbandomesticwaterconsumptionconsideringuncertainty[J].
JournalofWaterResourcesPlanningandManagement,2021,147(3):05020028.
[22] GUOTL,SONGSS,MAW J.Pointandintervalforecastingofgroundwaterdepthusingnonlinearmodels[J].
WaterResourcesResearch ,2021,57(12):e2021WR030209.
[23] 王晓东,鞠邦国,刘颖明,等.基于 QR - NFGLSTM与核密度估计的风电功率概率预测[J].太阳能学报,
2022,43(2):479 - 485.
[24] DRAGOMIRETSKIYK,ZOSSOD.Variationalmodedecomposition[J].IEEETransactionsonSignalProcessing,
2014,62(3):531 - 544.
[25] 王若恒.基于 LSTM的风电功率区间预测研究[D].武汉:华中科技大学,2018.
Anewstepwisedecompositionensemblemodelbasedontwo - stageparticle
swarm optimizationalgorithm fortherunoffprediction
1,2
1,2
1,2
1,2
GUOTianli ,SONGSongbai ,ZHANGTe ,WANGHuimin
(1.CollegeofWaterResourcesandArchitecturalEngineering,NorthwestA&FUniversity,Yangling 712100,China;
2.KeyLaboratoryofAgriculturalSoilandWaterEngineeringinAridandSemiaridAreas,MinistryofEducation,
NorthwestA&FUniversity,Yangling 712100,China)
Abstract:Thetraditionaldecompositionensemblerunoffpredictionmodelfirstlydecomposestheentirerunoff
seriesintoseveralsubseries ,andthendividesthesubseriesintotrainingandvalidationperiodsformodeling,
whichwronglytreatsthepredictordataofvalidationperiodasknowndataandisdifficulttobeappliedtoactualrun
offforecasting.Moreover ,thepredictionresultsofsuchmodelsareonlydefinitevalues,whichisdifficulttode
scribethepredictionuncertaintycausedbytherandomnessandvolatilityofrunoffseries.Tosolvetheaboveprob
lems ,thisstudyproposesastepwisedecompositionensemble(VMD - SVM- KDE)modelcombiningvariablemode
decompositionmethod ,supportvectormachinemodelandkerneldensityestimationmethod,whichperformsboth
pointpredictionandintervalprediction,andproposesatwo - stageparticleswarm optimization(TSCPSO)algo
rithm.ThemonthlyrunoffseriesoftheYellowRiverBasinisusedtoevaluatethemodelperformance ,andthe
2
studyresultsshowthat:(1)theVMD - SVM- KDEmodelimprovesthecoefficientofdetermination(R)andNash
efficiencycoefficient(NSE)valuesofthesingleSVM- KDEmodelfromtherangeof0.145to0.630totherangeof
0.872to0.921,andreducestheintervalaveragedeviation(INAD)valuesfromtherangeof0.046to95.844tothe
rangeof0.005to0.034 ,indicatingthattheVMD - SVM- KDEmodelsignificantlyimprovesthepointpredictionand
intervalpredictionperformanceofasingleSVM- KDEmodel ;(2)comparedwiththetraditionalone - stagePSOal
2
gorithm ,theTSCPSOalgorithmimprovestheR andNSEvaluesofthesinglemodelfrom therangeof0.145to
0.480totherangeof0.309to0.630 ,andreducestheINADvaluefromtherangeof48.813to95.844totherange
2
of0.046to0.195 ,andalsoimprovestheR andNSEvaluesofthedecompositionensemblemodelfromtherangeof
0.872to0.912totherangeof0.876to0.921 ,andreducestheINADvaluesfromtherangeof0.007to0.034tothe
rangeof0.005to0.014 ,indicatingthattheTSCPSOoptimizationalgorithmovercomestheoverfittingproblemofsup
portvectormachinemodelsandeffectivelyimprovesthepredictionaccuracyofthesingleanddecompositionensemble
models ;(3)theVMD - SVM- KDE - TSCPSOmodeladdressedthemistakesoftraditionaldecompositionensemble
modelsthatforecastfactordataofvalidationperiod ,andhashigheraccuracyofpointpredictionandintervalpredic
2
tionwithR andNSEvaluesofabout0.9andtheINADvaluesrangingfrom0.005to0.014.TheVMD - SVM - KDE -
TSCPSOmodelcanprovideabasisforpracticalforecastingofnon - stationaryandnon - linearhydrologicalseries.
Keywords:runoffprediction;intervalprediction;decompositionensemblemodel;two - stageparticleswarm op
timizationalgorithm ;variablemodedecomposition;supportvectormachine
(责任编辑:于福亮)
— 1 4 6 —
6