Page 130 - 2022年第53卷第6期
P. 130
2021,2(5):1 - 4.
[20] 王俊凯,乔非,祝军,等.基于支持向量机的烧结能耗及性能指标预测模型[J].同济大学学报(自然科学
版),2014,42(8):1256 - 1260.
[21] CHOK,VANMERRI?NBOERB,GULCEHREC,etal.LearningphraserepresentationsusingRNNencoder - de
coderforstatisticalmachinetranslation [J].arXive - prints,2014,arXiv:1406.1078.
[22] 姚德臣,李博阳,刘恒畅,等.基于注意力 GRU算法的滚动轴承剩余寿命预测[J].振动与冲击,2021,
40(17):116 - 123.
[23] 毕扬,郑波,张亚武,等.基于 MIC与 BiGRU的水电机组振动趋势预测 [J].水利学报,2021,52(5):
612 - 621.
[24] HUANGS,TANGJ,DAIJ,etal.Signalstatusrecognitionbasedon1DCNNanditsfeatureextractionmechanism
analysis [J].Sensors,2019,19(9):2018.
[25] ZHUANGZ,DINGJ,TANAC,etal.FaultdetectionofHigh - Speedtrainwheelsetbearingbasedonimpulse -
envelopemanifold [J].ShockandVibration,2017,2017:2104720.
[26] SHIJ,DINGZ,LEEWJ,etal.HybridforecastingmodelforVery - Shortterm windpowerforecastingbasedon
greyrelationalanalysisandwindspeeddistributionfeatures [J].IEEETransactionsonSmartGrid,2014,5(1):
521 - 526.
[27] HUANGC,WANGL.Gaussianprocessregression - basedmodellingoflithium - ionbatterytemperature - dependent
open - circuit - voltage [J].ElectronicsLetters,2017,53(17):1214 - 1216.
[28] GAOW,KARBASIM,HASANIPANAH M,etal.DevelopingGPRmodelforforecastingtherockfragmentation
insurfacemines [J].EngineeringwithComputers,2018,34(2):339 - 345.
Degradationtrendassessmentandpredictionofpumpedstorageunitbasedon
deepauto - encodercompressionandmultiscalefeatureextraction
2
1
1
1
2
3
CHENPeng,WUYifan,CAIShuang,YANGBin,ZHANGHaiku,LIChaoshun
(1.SchoolofCivilandHydraulicEngineering,HuazhongUniversityofScienceandTechnology,Wuhan 430074,China;
2.DatangTibetEnergyDevelopmentCompanyLimited,Chengdu 610072,China;
3.DatangHydropowerScience&TechnologyResearchInstituteCo.,Ltd.,Chengdu 610074,China)
Abstract:Theharshoperatingenvironmentbringschallengestothesafeoperationofthepumpedstorageunit
(PSU).ThedegradationtrendassessmentandpredictionofPSUcaneffectivelyreflectoperationstateoftheunit.
However ,thereisalargeamountofredundantandinterferinginformationinoperatingparametersoftheunit,
whichseriouslyaffectsthereliabilityoftheassessment.Besides,itishardtoachieveaccuratepredictionforcom
plexdegradationtrendsequences.Tosolveaboveproblems ,adegradationtrendassessmentandpredictionmodel
ofPSUisproposedonthebasisofdeepauto - encodercompressionandmultiscalefeatureextraction.Firstly,the
healthymodelisbuiltusingdeepauto - encoder(DAE)andmultilayerperceptron(MLP)toreducethefittinger
ror ,whereDAE isadoptedtocondensethecriticalinformation.Secondly,thedegradationtrendisgenerated
basedonthehealthymodel.Finally,amultiscalefeatureextractionnetworkisconstructedbycombiningthead
vantagesofone - dimensionalconvolutionalneuralnetwork(1DCNN)andbi - directionalgatedrecurrentunit(BiG
RU )fortheaccurateprediction.Comparedwithothermodels,theproposedhealthymodelachievedthelowestfit
tingerrorandthemultiscalefeatureextractionnetworkhasthehighestpredictionaccuracy.
Keywords:pumpedstorageunit;degradationtrendassessmentandprediction;deepauto - encoder(DAE);
multiscalefeatureextraction;one - dimensionalconvolutionalneuralnetwork(1DCNN);bi - directionalgatedre
currentunit (BiGRU)
(责任编辑:王 婧)
— 7 5 —
6