基于孤立森林和GRU-CNN-Attention的超短期电力负荷预测  

Ultra⁃short⁃term power load forecast based on isolated forests and GRU⁃CNN⁃Attention

作　　者：陈雷[1] 刘林虎 闫川川 孙艳镯 于凯[1] 周超[1] CHEN Lei;LIU Linhu;YAN Chuanchuan;SUN Yanzhuo;YU Kai;ZHOU Chao(Department of Telecommunications,Northeast Petroleum University Qinhuangdao,Qinhuangdao 066004,China;Shougang Jingtang Iron and Steel United Co.,Ltd.,Tangshan 063205,China)

机构地区：[1]东北石油大学秦皇岛校区电信系,河北秦皇岛066004 [2]首钢京唐钢铁联合有限责任公司,河北唐山063205

出　　处：《电子设计工程》2023年第20期140-144,149,共6页Electronic Design Engineering

基　　金：黑龙江省属高等学校基本科研任务费项目(2018QNQ-07);秦皇岛市科技局科学技术研究与发展计划项目(201902A016)。

摘　　要：针对目前超短期负荷预测存在的速度慢、精度低等问题,在广泛应用的门控循环单元(GRU)神经网络预测方法的基础上加入卷积神经网络(CNN)和注意力(Attention)机制模型。使用孤立森林算法筛选出不良负荷数据,将负荷数据通过GRU神经网络按时序提取负荷特征,利用CNN按每日负荷规律进行特征提取,并引入Attention机制提高预测精度。选用10天负荷数据进行分析,在同等条件下与GRU、GRU-CNN等方法进行对比,平均绝对百分比误差分别降低0.28%、0.272%,证明所述模型能应用在超短期负荷预测中。In view of the current problems of slow speed and low accuracy of ultra⁃short⁃term load prediction,Convolutional Neural Network(CNN)and Attention mechanism models are added to the widely used method of Gated Recurrent Unit(GRU)neural network prediction.The isolated forest algorithm was used to screen out the bad load data,to extract the load features in time series through the GRU neural network,and the CNN was used to extract the features according to the daily load rule,and the Attention mechanism was introduced to improve the prediction accuracy.The 10-day load data was selected for analysis,compared with the GRU and GRU-CNN methods under the same conditions,and the average absolute percentage error was reduced by 0.28%and 0.272%,respectively,proving that the model can be applied in ultra⁃short⁃term load prediction.

关 键 词：超短期负荷预测 孤立森林 门控循环单元 卷积神经网络 注意力机制 

分 类 号：TM714[电气工程—电力系统及自动化]