Page 129 - 2024年第55卷第9期
P. 129
[32] YAOZY,WANGZC,WANGDW,etal.AnensembleCNN - LSTM andGRUadaptiveweightingmodelbased
improvedsparrowsearchalgorithm forpredictingrunoffusinghistoricalmeteorologicalandrunoffdataasinput [J].
JournalofHydrology(Amsterdam),2023,625:129977.
Streamflowpost - processingbasedondistributedhydrologicalfluxesand
spatio - temporaldeeplearningalgorithm
1
1
1
2
WUYao,XUYueping,LIULi,HEKeqi
(1.InstituteofWaterScienceandEngineering,CollegeofCivilEngineeringandArchitecture,
ZhejiangUniversity,Hangzhou 310058,China;
2.EarthandClimateSciences,NicholasSchooloftheEnvironment,DukeUniversity,Durham,NC 27708,USA)
Abstract:Accuratesimulationofstreamflowisacrucialprerequisiteforwaterresourcesmanagementandregional
integratedpolicymaking.Inordertoimprovetheaccuracyofstreamflowsimulation,thisstudytakesYonganxiRiv
erBasininTaizhou,ZhejiangProvinceasthestudyarea.ACNN - LSTM spatio - temporalpost - processingmodel
bycouplingCNNwithLSTMisproposedbasedonthemeasureddailydischargedataatBaizhi ’aoStationfrom2010
to2019andhydrologicalfluxessimulatedbytheGrid - HBV model.Weconstructtwopost - processingmodels,
namelyCNN - LSTM withsingleflux(s - CNN - LSTM)andCNN - LSTM withdoublefluxes(bi - CNN - LSTM).Their
performanceiscomparedandanalyzedwithabenchmarkmodel (s - LSTM).TheresultsshowthattheNSEofthe
Grid - HBVmodelduringthecalibrationandvalidationperiodsare0.78and0.81 ,respectively,indicatinganover
allgoodrunoffsimulation.However ,thereareunderestimationinmediumandhighflowandoverestimationinlow
flowsimulations.Afterpost - processing ,theNSEofs - LSTM inthetwostudyperiodsare0.87and0.85,withan
increaseof11.2% and5.8%,andtheNSEofs - CNN - LSTM are0.90and0.89,withanincreaseof14.6% and
10.9%.TheNSEofbi - CNN - LSTMinthetwostudyperiodsbothreach0.92 ,withanincreaseof17.2% and14.2%.
Comparedtothes - LSTM model,thebi - CNN - LSTM modelpresentsafurtherenhancementof6.0% and8.4% in
accuracy.Inaddition ,thebi - CNN - LSTM modelcanmarkedlyimprovethedefectsoforiginalsimulationinthe
high ,mediumandlowflows.Forfourtypicalfloodevents,thebi - CNN - LSTMmodelhasthebestpost - processing
effect ,whichreducesthefloodpeakerrorby36.6% onaverage,thes - LSTMmodelandthes - CNN - LSTMmodel
reducesthefloodpeakerrorby19.3% and30.3% onaverage.Insummary ,theCNN - LSTM modelbasedondis
tributedhydrologicalfluxeshasagoodabilityofstreamflowpost - processing ,whichcansignificantlyimprovethe
streamflowsimulationsofhydrologicalmodels.
Keywords:post - processing;CNN - LSTM;deeplearning;gridHBVhydrologicalmodel;JiaoRiverbasin
(责任编辑:耿庆斋)
— 1 1 4 —
3