Page 26 - 2023年第54卷第10期
P. 26
[22] 蔡杰.基于 MLP - ARX模型盾构自动掘进技术研究与应用 [J].隧道建设,2022,42(10):1797 - 1803.
[23] ISABONAJ,IMOIZEAL,OJOS,etal.Developmentofamultilayerperceptronneuralnetworkforoptimalpredic
tivemodelinginurbanmicrocellularradioenvironments[J].AppliedSciences - Basel,2022,12(11):5713.
[24] ZHANGY,CHENGP,MALIKOP,etal.Anartificialneuralnetworkbasedadaptivepower - system stabilizer
[J].IEEETransactionsonEnergyConversion,1993,8(1):71 - 77.
[25] 王雨虹,王淑月,王志中,等.基于改进蝗虫算法优化长短时记忆神经网络的多参数瓦斯浓度预测模型研
究[J].传感技术学报,2121,34(9):1196 - 1230.
[26] LUOJ,CHENH L,ZHANGQ,etal.Animprovedgrasshopperoptimizationalgorithmwithapplicationtofinan
cialstressprediction[J].AppliedMathematicalModelling,2018,64:654 - 668.
IGOA - MLPdynamicpredictionmodelforsimulationparametersofhighcore
rockfilldam constructionundertransferlearningframework
1
1
1
1
2
LFei,ZHONGDenghua,YUJia,ZHANGJun,ZHANGYunuo
(1.StateKeyLaboratoryofHydraulicEngineeringSimulationandSafety,TianjinUniversity,Tianjin 300072,China;
2.CollegeofWaterResourcesandCivilEngineering,ChinaAgriculturalUniversity,Beijing 100083,China)
Abstract:Forconstructionsimulationofhighcorerockfilldam,theparametersarethekeytoensuringitsaccura
cy.However,existingparameterpredictionmethodsusedhistoricaldataandignorethedifferencesbetweenthe
constructionprocessesofdifferentlayers ,andthereisofteninsufficientormissingdataatthebeginningofanew
layer.Inaddition ,theparametersareaffectedbymanyfactorssuchasmeteorologicalconditionsandoperating
stateofthemachine.Tosolvetheaboveproblems ,thispapertakesadvantageofthetransferlearning’ scapability
ofmodelingwithsmallsamplesthroughknowledgetransferandconsidersthequantitativeinfluenceofvariousfac
tors.Animprovedmulti - layerperceptrondynamicpredictionmodel (IGOA - MLP)isproposedforconstruction
simulationparametersofhighcorerockfilldamundertheframeworkoftransferlearning.Firstly ,theIGOA - MLP
predictionmodelisestablishedthatconsideringtheinfluenceofmultiplefactors.Thegrasshopperoptimizationalgo
rithmisimproved (IGOA)bynonlinearreductionfactorandCauchy - Gaussianhybridmutationmode,andtheeffi
cientglobaloptimalsearchcapabilityofIGOAisutilizedtooptimizethehyperparametersofmulti - layerperceptron
(MLP).Secondly,thetransferlearningstrategyisintroducedtorealizetheknowledgetransferbetweenthehistori
calandnewconditionsandsolvetheproblemofinsufficientormissingdatainthenewconditions.Thetrainingset
isdividedintosourcedomainandtargetdomain ,andanadaptivelayerisaddedtothehiddenlayerofMLPtore
presentthedifferencebetweensourcedomaindataandtargetdomaindata.Acasestudyshowsthatcomparedwith
othermachinelearningmethodssuchasMLPmodelandIGOA - MLPmodelwithouttransferlearning,themeanab
solutepercentageerror (MAPE)oftheproposedmethodisreducedby54.68% and40.57%,respectively.Itis
provedthattheproposedmodelcanpredicttheparametersofconstructionsimulationmoreaccuratelyandprovidea
reliabledatabasisforsimulation.
Keywords:transferlearning;highcorerockfilldam;constructionsimulation;multi - layerperceptronoptimized
byimprovedgrasshopperoptimizationalgorithm;parameterprediction
(责任编辑:李福田)
— 1 1 2 —
6