第47卷第4期电网技术Vol.47No.42023年4月PowerSystemTechnologyApr.2023文章编号：1000-3673（2023）04-1540-08中图分类号：TM614文献标志码：A学科代码：470·40基于格拉姆角场与改进的风电功率预测方法张淑清1，杜灵韵1，王册浩1，姜安琦1，徐丽华2（1．教育部智能控制系统与智能装备工程研究中心(燕山大学)，河北省秦皇岛市066004；2．国网冀北电力有限公司唐山供电公司，河北省唐山市063000）WindPowerForecastingMethodBasedonGAFandImprovedCNN-ResNetZHANGShuqing1,DULingyun1,WANGCehao1,JIANGAnqi1,XULihua2(1.EngineeringResearchCenterforIntelligentControlSystemandIntelligentEquipment(YanshanUniversity),MinistryofEducation,Qinhuangdao066004,HebeiProvince,China;2.TangshanPowerSupplyCompanyofNorthHebeiElectricPowerCo.,Ltd.,Tangshan063000,HebeiProvince,China)ABSTRACT:Theuncertaintyofwindpowerbringsaboutgreatdifficultiestothepowersystemscheduling,soitisparticularlysignificanttoaccuratelyforecastthewindpowervariations.Undertheinfluenceoftherapiddevelopmentoftheimageprocessinginrecentyears,itispossibletoimprovetheaccuracyofclassificationandcomputationbyencodingthetimeseriesintotwo-dimensionalimages,whichallowstheneuralnetworkstorecognizeandlearnfromthedata"visually".Therefore,awindpowerforecastbasedontheGAFandtheimprovedCNN-ResNetisproposed.Firstly,theGramianAngularFields(GAF)isusedtoconverttheone-dimensionalhistoricalwindpowerdataintotwo-dimensionalimages.Thecorrelationandfeaturesofthetime-seriesareextractedbytheConvolutionalNeuralNetwork(CNN).Secondly,theResidualNeuralNetwork(ResNet)isusedtoextractthefeaturesoftheotherrelevantdatatowindpower,andthedegradationproblemissolvedwhilethenetworkdepthisincreasedtoimprovetheaccuracyofprediction.Then,thetwoneuralnetworksarefusedtoconstructadual-inputneuralnetworkstructure.Finally,thismethodissuccessfullyappliedtothedatasetsoftheMahuangshanNO.1windfarminNingxiaHuiAutonomousRegioninChinaandtheGaliciawindfarminSpain.ComparedwiththeResNet,theCNN-MLP,theGRU,theBP,theLSTMandtheBiLSTMnetworkmodels,theforecastmethodoftheGAFandtheimprovedCNN-ResNetproposedinthispaper...
