Page 122 - 《水资源与水工程学报》2025年第1期
P. 122
!36 "!1 # & ' ( ) & * + , - Vol.36No.1
2025 $ 2 % JournalofWaterResources&WaterEngineering Feb.,2025
DOI:10.11705/j.issn.1672-643X.2025.01.13
Ôãä嵦W BiTCN-Attention-LSSVMfOMN
1,2,3 1,3 1,3 1,3 1,2,3
, qV , |y , , ,
(1. ¼Fõa0 Zb0G, ¼F Á¤ 650201;2. E9a0 Zþ
á9!?@pqr, ¡Á FU 210098;
3. ¼Ff4ºZbCDáâ¦CDAB4, ¼F Á¤ 650201)
> ?: ¯§4Æp0iº¦ÿÏ@íêÏcÖ{©ªô©éÞstª:?WÑ{ª³, #¯}
.%ÔÏ@øÌµ¶·(BiTCN)、 Ú'p´(Attention) z_I¤¡Ôp(LSSVM) {è,<½Jðñ
¿/iº。uiºb BiTCNX¨、 PfÔ©*Ï@íê4{ô©éÞst, £ç Attention p´xiº'Ú
cðñs{sîª:Â7, À° BiTCN-Attention æ~ÿP{Â7Àç LSSVMiº4ÆÇðñ, zPÖ 2 f
Ò{íêþµ¶¯iº{ðñ{«。ò µ¶S¤: c LSSVM、CNN-LSSVM TCN-LSSVM6,BiTCN-At
tention-LSSVMiºðñ{Á(³´ü3z», ¦è,<ñ&Z¦ðñ490#¢é{iºß~í·Ñ;
BiTCNc Attention {®/ȯ¢Ý4+Ï@íê4{éÞst, ° BiTCN-Attention +{Â7Àç
LSSVM4ÆÇðñ*t[Ý{ðñÑÈ, íêþÃ^ÿP{é¯iº{0È'。
1@A: è,<ñ&Z¦; ½Jðñ; ÔÏ@øÌµ¶·; Ú'p´; z_I¤¡Ôp
BC<DE:TV698.11 FGHIJ:A FKLE:1672643X(2025)01011811
BiTCN-Attention-LSSVM Modelingforseepage
predictionofearth -rockdams
1,3
1,3
1,3
FUShuyan 1,2,3 ,YANGShiyong ,CHENDehui ,WANGZixuan ,OUBin 1,2,3
(1.CollegeofWaterConservancy,YunnanAgriculturalUniversity,Kunming650201,China;2.NationalKeyLaboratoryof
WaterDisasterPrevention,HohaiUniversity,Nanjing210098,China;3.ResearchCenterforSmartManagementand
MaintenanceofSmallandMediumSizedWaterConservancyProjectsinYunnanProvince ,Kunming650201,China)
Abstract:Conventionalmachinelearningmodelsoftenfailthetaskofidentifyinglongterm dependencies
andlocalimportancewhendealingwithtimeseriesdata.Toaddressthisproblem ,weproposedacoupled
modelforseepagepredictionofearthandrockdamsbasedonbidirectionaltemporalconvolutionalnetwork
(BiTCN),attentionmechanism (Attention),andleastsquaressupportvectormachine(LSSVM).The
modelusesBiTCNtocapturethelongterm dependenciesinthetimeseriesdatafrom bothforwardand
backwarddirections ,introducestheattentionmechanismtohelpthemodelfocusonthekeylocalfeatures
relatedtotheprediction,andinputsthedeeplyprocessedBiTCN-AttentionfeaturesintotheLSSVMmodel
forbetterpredictionresults.Toanalyzethepredictioneffect,themodelwasusedtoprocesstwodifferent
datasets.TheresultsofcaseanalysisshowthatcomparedwithLSSVM,CNN-LSSVMandTCN-LSSVM,
theevaluationindexesofBiTCN-Attention-LSSVM modelarethebest,andtheproposedmodelshows
higheraccuracyandstabilityinthepredictionofpiezometerwaterlevelintheearth-rockdam.Thecombi
nationofBiTCNandAttentioncanbetterextracttheinterdependencerelationshipintimeseriesdata.Input
tingthefeaturesextractedbyBiTCN-AttentionintoLSSVM forpredictioncanachievebetterprediction
performance.Afterthedatasetisexpanded ,thelearningabilityofthemodeliseffectivelyimproved.
Keywords:piezometerwaterlevelinearth-rockdam;seepageprediction;bidirectionaltemporalconvolution
!"#$:20240726; %$:20241031
'()*: !"wx#0%»(3(52069029、52369026);“ }¨}Ç” Zc¡¢89#$%»]x(3(2023490411);
¼Ffõ%&ABo/'(r¬(3(202401BD070001071)
+,-.: .((1976—), Æ, ¼F:/N, OP, QR, ÈPSTU, ABÔCD'0cZß89。
qr+,: A (1983—), K, ±FÁ¯N, OP, QR, ÈPSTU, ABÔZC®X89b¾c½¾0-。