ElectricalAutomation《电气自动化》2023年第45卷第1期电力系统及其自动化PowerSystem&Automation基于深度回归森林的短期电力负荷预测黄文思,陆鑫,陈婧,林超,薛迎卫,施炜炜(国网信通亿力科技有限责任公司,福建福州350003)摘要:为减轻深度学习算法对于网络超参数的依赖,提出了基于深度回归森林的短期电力负荷预测方法。所提方法利用深度森林的默认超参数构建多粒度扫描过程和级联森林过程的森林模型。首先,通过多粒度扫描过程有效学习样本的内在特征并提取序列数据的时序特征;然后,将所有特征向量用作级联森林过程的输入,筛选最终特征向量;最后,利用训练数据的特征对预测样本进行预测。结果表明,所提方法能够有效地减轻超参数配置对深度学习模型的影响,预测结果比较精确。关键词:深度回归森林;短期负荷预测;多粒度扫描过程;级联森林过程;数据挖掘DOI:10.3969/j.issn.1000-3886.2023.01.005[中图分类号]TP181[文献标志码]A[文章编号]1000-3886(2023)01-0018-04Short-termElectricalLoadForecastingBasedonDeepRegressionForestHuangWensi,LuXin,ChenJing,LinChao,XueYingwei,ShiWeiwei(StateGridInfo-TelecomGreatPowerScienceandTechnologyCo.,Ltd.,FujianFuzhou350003,China)Abstract:Inordertoreducethedependenceofdeeplearningalgorithmonnetworkhyperparameters,ashort-termpowerloadforecastingmethodbasedondeepregressionforestwasproposed.Theproposedmethodusedthedefaultsuperparametersofdeepforesttoconstructtheforestmodelofmultigranularityscanningprocessandcascadeforestprocess.Firstly,theinternalfeaturesofsampleswereeffectivelylearnedthroughthemultigranularityscanningprocess,andthetemporalfeaturesofsequencedatawereextracted.Then,allfeaturevectorswereusedastheinputofcascadeforestprocesstofilterthefinalfeaturevector.Finally,thecharacteristicsoftrainingdatawereusedtopredictthepredictionsamples.Theresultsshowthattheproposedmethodcaneffectivelyreducetheinfluenceofhyperparameterconfigurationonthedeeplearningmodel,andthepredictionresultsaremoreaccurate.Keywords:deepregressionforest;short-termloadforecasting;multi-grainedscanningprocedure;cascadeforestprocedure;datamining定稿日期:2022-10-200引言负荷预测对于指导电...