文档名：基于时空注意力机制的台区多用户短期负荷预测
摘要：针对在低压台区海量高波动用户负荷预测场景下,传统探索单个用户时间特征的负荷预测方法存在无法学习用户之间的空间相关性、无法实现多用户共同预测的问题,该文提出一种基于时空注意力机制的Transformer负荷预测模型(STformer),提供精准的台区多用户短期负荷预测.首先,改进传统Transformer模型,嵌入序列分解模块、自相关计算模块和空间注意力模块.其中,序列分解模块可以将波动较大的用户负荷曲线分解为相对平稳的多个子序列,有助于更好地提取负荷曲线的时间依赖性和周期因子;自相关计算是一种改进的注意力机制,可以挖掘多个历史同时期子序列的时间相关性;空间注意力机制可以提取台区多用户之间的动态空间相关性.然后,利用蒙特卡洛随机失活方法(MCdropout)将STformer拓展到台区多用户负荷概率预测.最后,采用真实台区多用户负荷数据集进行验证,与多种负荷预测模型进行对比,证明STformer模型可有效提高短期多用户负荷点预测和概率预测的精确性和鲁棒性.
Abstract：Withalargenumberofcustomer-sidedistributedpowersourcesenteringthenetworkfromlow-voltagedistributionstationsandthewidespreaduseofdevicessuchassmartmetersonthecustomerside,aloadforecastingmodelformultipleusersisneededtofacilitatepointforecastingandprobabilistictasksforalargenumberofusersefficientlyandaccurately.Traditionalmethodsofcustomerloadforecastingmodelthetemporalcharacteristicsofindividualcustomersandareunabletolearntheproblemsofspatialcorrelationbetweencustomersandtheinabilitytoachieveforecastsformultiplecustomers.Customersinthesameregionsharethesamegeographicspace,weatherconditions,holidayinformation,tariffpolicies,andothercomprehensivefactors,andthereisoftenacertainamountofspatialandtemporalcorrelationbetweencustomers'electricityconsumptionbehavior.Ifthisspatial-temporalcorrelationcanbefullyexplored,itwillhaveextremelypositiveimplicationsformodelingshort-termcustomerloads.Asmallbodyofliteraturehasalreadyexploredtheinitialexplorationofcustomerloadforecasting,takingspatial-temporalcorrelationintoaccount.However,theexistingspatio-temporalmethodscanonlyprovidedeterministicforecasts,notprobabilisticones.Toaddresstheseissues,thispaperproposesamulti-customershort-termloadforecastingmodelforstationareas.Learningspatial-temporalcorrelationinformationfromhistoricalloaddatacanperformaccuratemulti-usershort-termloadpointforecastsandprobabilisticforecastsforstationareas.Firstly,threemodulesareembeddedforeachencoderanddecoderbyimprovingthestandardTransformerself-attentionmechanism:sequencedecompositionmodule,autocorrelationcalculationmodule,andspatialattentionmoduletoeffectivelyextractthedynamicspatio-temporaldependenciesamonghighlyvolatileresidentialusers.Amongthem,thesequencedecompositionmodulecandecomposehighlyvolatilesubscriberloadcurvesintorelativelysmoothmultiplesub-series,whichhelpstoextractbetterthetimedependenceandperiodfactorofloadcurves;theautocorrelationcalculationisanimprovedattentionmechanismthatcanminethetimedependenceofmultiplehistoricalcontemporaneoussub-series;andthespatialattentionmechanismcanextractthedynamicspatialsupportamongmultipleusersinastationarea.TheSTformermodelisthenextendedtothefieldofprobabilisticforecastingusingaMonteCarlostochasticdeactivationmethod(MCdropout).ThismethoddoesnotrequireadditionalmodificationstoSTformerbutallowsSTformertooutputbothpointpredictionandprobabilisticpredictionresults.Finally,theSTformermodelwithMCdropoutisusedtoforecastthestationcustomerload,andbothpointandprobabilisticforecastsareoutput.Inthispaper,themodel'svalidityisverifiedusingone-hour-aheadloadforecastingandday-aheadloadforecastingusingaccuratestationcustomerloaddatafromaprovinceinthesoutheast.TheproposedSTformermodelhasaMAPEof4.44%foreachuserand2.21%forthetotalloadinstationareaA.Theprobabilisticforecastevaluationindexpinballis0.3701;theaveragerelativeerrorMPEforeachuserand3.25%forthetotalloadinstationareaAis6.21%.is3.25%,andtheprobabilisticforecastassessmentindexpinballis0.5942.Thispaperalsocomparestheeffectsofdifferentmodulesontheexperimentalresultsthroughablationexperiments.ThispaperalsoverifiesthechangeinmodelinferencespeedbroughtaboutbytheadditionofFFT,comparingtherunningmemoryandtimeoftheautocorrelation-basedmodelwiththatoftheself-attentive-basedmodelduringthetrainingphase.Thefollowingconclusionscanbedrawnfromthesimulationanalysis:(1)Comparedwithotherbaselines,theSTformermodelproposedinthispaperextractsthetemporalvariationpatternofusersthroughthetemporalattentionmechanismandthespatialdependencybetweenmultipleusersthroughthespatialattentionmechanism,whichultimatelyachievesthebestpredictionresultsinallscenarios.(2)EachmoduleofSTformercontributestotheimprovementofpredictionaccuracyandmodelrobustness.ThespatialattentionmodulehasthegreatestimpactonthepredictionaccuracyofSTformer,andtheFouriertransformmethodoftheautocorrelatedmodelreducesthecomputationalcomplexityandthusacceleratesthecomputationalspeedofthemodel.(3)ThepredictionintervalsoftheproposedSTformermodelwithMCdropouthavereliablecoverageofthetruevaluesandprovidenarrowerpredictionintervals,especiallyatsomepeaksandtroughs,whicharecriticalforthetemperatureoperationofpowersystems.
作者：赵洪山 吴雨晨 温开云 孙承妍 薛阳 Author：ZhaoHongshan WuYuchen WenKaiyun SunChengyan XueYang
作者单位：河北省分布式储能与微网重点实验室(华北电力大学)保定071003中国电力科学研究院有限公司北京100180
刊名：电工技术学报
Journal：TransactionsofChinaElectrotechnicalSociety
年，卷(期)：2024, 39(7)
分类号：TM743
关键词：多用户负荷预测 时空相关性 Transformer模型
Keywords：Multi-customersloadforecasting spatial-temporalcorrelation Transformermodel
机标分类号：TP391TM715TP183
在线出版日期：2024年4月12日
基金项目：国家电网公司总部科技项目基于时空注意力机制的台区多用户短期负荷预测[
期刊论文] 电工技术学报--2024, 39(7)赵洪山 吴雨晨 温开云 孙承妍 薛阳针对在低压台区海量高波动用户负荷预测场景下,传统探索单个用户时间特征的负荷预测方法存在无法学习用户之间的空间相关性、无法实现多用户共同预测的问题,该文提出一种基于时空注意力机制的Transformer负荷预测模型(...参考文献和引证文献
参考文献
引证文献
本文读者也读过
相似文献
相关博文
基于时空注意力机制的台区多用户短期负荷预测 Short-Term Load Forecasting for Multiple Customers in A Station Area Based on Spatial-Temporal Attention Mechanism
基于时空注意力机制的台区多用户短期负荷预测.pdf
- 文件大小:
- 1.07 MB
- 下载次数:
- 60
-
高速下载
|
|
