第41卷第1期2023年1月MACHINERY&ELECTRONICSVol.41No.1Jan.2023收稿日期:20220908作者简介:吴平雄(1997-),男,福建福州人,硕士研究生,研究方向为风电功率预测;肖迎群(1975-),男,湖南邵阳人,博士,副教授,研究方向为高维数据分析、机器学习。基于组合数据清洗与NLConvLSTM模型的多步风电功率预测吴平雄1,肖迎群2,张苏2,林兴宇1(1.贵州大学电气工程学院,贵州贵阳550025;2.贵州理工学院大数据学院,贵州贵阳550003)摘要:针对风电数据在采集与传输过程中会产生大量缺失值和异常值,采用DBSCAN算法和最优组内差分法(OIV)组合筛删异常值,随机森林(RF)算法填补缺失值,提升数据准确性;并建立基于以ConvL-STM为单元的编码预测(EF)网络的风电多气象输入多步预测模型,为了更好利用气象特征信息,在Conv-LSTM模型的输入侧添加具有自注意力机制的非局部(NL)模块增强数据特征表现,从而搭建组合数据清洗方法的NLConvLSTM多步风电功率预测模型。实验结果表明,该方法能够进一步提高风电功率多步预测精度和稳定性。关键词:多步预测;风电功率预测;ConvLSTM;数据清洗;非局部操作中图分类号:TM614文献标志码:A文章编号:10012257(2023)01001307MultistepWindPowerPredictionBasedonCombinedDataCleaningandNLConvLSTMModelWUPingxiong1,XIAOYingqun2,ZHANGSu2,LINXingyu1(1.SchoolofElectricalEngineering,GuizhouUniversity,Guiyang550025,China;2.SchoolofBigData,GuizhouInstituteofTechnology,Guiyang550003,China)Abstract:Consideringthatalargenumberofmissingvaluesandoutlierswillbegeneratedduringthecollectionandtransmissionofwindpowerdata,DBSCANalgorithmandoptimalinterclassvariance(OIV)methodareusedtofilteroutoutliers,andrandomforestalgorithmisusedtofillinthemissingdatatoim-provetheaccuracyofthedata.TheencodingforecastingnetworkbasedonConvLSTMcellisamultistepforecastmodelforwindpowerwithmultimeteorologicalinput.Inordertomakebetteruseofmete-orologicalfeatureinformation,anonlocalmodulewithselfattentionmechanismisaddedtotheinputsideoftheencodingforecastingnetworktoenhancetherepresentationofdatafeatures,soastobuildtheNLConvLSTMmultistepwindpowerpredictionmodel,acombiningdatacleaningmethod.Theexperi-mentalresultsshowthatthismethodcanfurtherimprovetheaccuracyandstabilityofmultistepforecas-tingofwindpowe...
