Page 77 - 《水资源与水工程学报》2024年第5期
P. 77
!35 "!5 # & ' ( ) & * + , - Vol.35No.5
2024 $ 10 % JournalofWaterResources&WaterEngineering Oct.,2024
DOI:10.11705/j.issn.1672-643X.2024.05.09
'x CNN-LSTM-Attention ýwv³²?ë²
ýöA^q{a}
1,2
1,2
1
1
nvw , xyz , { W , n|} , ~ 3
(1. ùÄe. AR[c¶.T, 56 ùÄ 450001;2. 56ªAWXYZ[AXhÚ¯!~©è^_`,
56 ùÄ 450001;3. Q!hK\KÐUtThK\KQo, ;; Ut 264000)
H I: @úÐ.lYVð~XhiUV{çUYZÿ, V5UV@Ê, /ÁW%èÙ
(CNN)、 c`ÕGHèÙ(LSTM) äÓ¦2¸_(Attention) { CNN-LSTM-Attention ¥èYV, c¼yø
ð, °èû&¡!~¥èYQ'( CMIP6 Q{ BCCCSM2MRáõYQÎCÁ¼yÓ, lÁUV{çU
YZä" , |¼yÐ.lYV{YZÿ。°±²o:CNN-LSTM-Attention YVéV5UV²É
£{çUYZk±, YZÿµøÁ,Ð.lYV, øXkû¿(NSE) @ 0.883, µ>J45(RMSE)
@ 2.317, ´µpð45(MAE) @ 1.098; ºáõ~Ói, V5UVé 2025—2050 v{vçU1µÈÉ×n
~mÌÍÃʦ»e, Æ,é SSP12.6 Ói, çU1~mäʦ»。%<*@Ð.lYVé9A
¨¿ï1'YZÑV{lúÙ¶hÉ, Î@UVáÑ{ADEÈ&RäÃúÙ-.BCD½。
JKL: çUYZC" ; Ð.lYV;CNN-LSTM-Attention; áõ~; V5UV
MNFOP:TV213.4 QRSTU:A QVWP:1672643X(2024)05007309
RunoffsimulationandfuturemultiscenariopredictionintheQinheRiver
BasinbasedontheCNN-LSTM -AttentionModel
1,2
3
1
1
1,2
ZHANGShuqi,ZUOQiting ,ZANGChao ,ZHANGLekai,BAYinji
(1.SchoolofWaterConservancyandTransportation,ZhengzhouUniversity,Zhengzhou450001,China;2.Henan
InternationalJointLaboratoryofWaterCycleSimulationandEnvironmentalProtection ,Zhengzhou450001,China;
3.YantaiCenterofCoastalZoneGeologicalSurvey,ChinaGeologicalSurvey,Yantai264000,China)
Abstract:Toenhancetheaccuracyofdeeplearningmodelsinsimulatingwatershedrunoffunderchan
gingenvironmentalconditions,acoupledmodelofconvolutionalneuralnetwork(CNN),longshortterm
memory (LSTM)andAttentionmechanismwasconstructedforthestudyoftheQinheRiverBasin.Inte
gratedwithmultipleoptimizationalgorithmsandmultiplescenariosinBCC-CSM2-MRclimatemodel
fromtheCoupledModelIntercomparisonProjectPhase6(CMIP6),thismodelwasappliedtowatershed
runoffsimulationandprediction.Itssimulationaccuracywasthencomparedwiththatofvariousdeep
learningmodels.TheresultsdemonstratethattheCNN-LSTM-Attentionmodelexhibitssuperiorper
formanceinsimulatingrunoffintheQinheRiverBasin,withNashSutcliffeefficiencycoefficient(NSE)
of0.883,rootmeansquareerror(RMSE)of2.317,andmeanabsoluteerror(MAE)of1.098,outper
formingotherdeeplearningmodels.Notably,theannualrunoffoftheQinheRiverBasinfrom2025to
2050showsaslowdecreasingtrendwithsignificantfluctuationsunderdifferentclimatechangescenarios ,
especiallyintheSSP12.6scenario.Thisstudyprovidesnewinsightsintotheapplicationofdeeplearn
ingmodelsinintelligentsimulationofhuman-waterrelationshipsandoffersareferentialvalueforsubse
!"#$:20240327; %$:20240714
'()*: ! " # $ % & ' ( ) * (2021YFC3200201); Q ! § » - & + Ö Ù 5 6 % < T Ö Ù R S % < ) *
(2024HENYB01); Q!hK\KÐ)*(DD20220885)
+,-.: c¨è(1999—), H, 56ùÄ9, u;%<=, ÞßfAB.CADE%<。
/0+,: çT[(1989—), H, ;;;9, u;, §»P, %<>?@+,DEçè\K[ß
çè| 。