基于自适应图注意力网络的短期用户负荷预测  

Short-term user load prediction based on an adaptive graph attention network

作　　者：黄冬梅 陈欢 王宁 吴志坚 胡伟 孙园 HUANG Dongmei;CHEN Huan;WANG Ning;WU Zhijian;HU Wei;SUN Yuan(College of Electronic and Information Engineering,Shanghai University of Electric Power,Shanghai 201306,China;College of Electrical Engineering,Shanghai University of Electric Power,Shanghai 200090,China;Suzhou Power Supply Branch,State Grid Jiangsu Electric Power Co.,Ltd,Suzhou 215004,China;College of Economics and Management,Shanghai University of Electric Power,Shanghai 200090,China;College of Mathematics and Physics,Shanghai University of Electric Power,Shanghai 201306,China)

机构地区：[1]上海电力大学电子与信息工程学院,上海201306 [2]上海电力大学电气工程学院,上海200090 [3]国网江苏省电力有限公司苏州供电分公司,江苏苏州215004 [4]上海电力大学经济与管理学院,上海200090 [5]上海电力大学数理学院,上海201306

出　　处：《电力系统保护与控制》2023年第20期140-149,共10页Power System Protection and Control

基　　金：国家社会科学基金项目资助(19BGL003);上海市科委地方院校能力建设项目资助(20020500700)。

摘　　要：为提高短期用户负荷预测精度,提出了一种基于自适应图注意力网络(adaptive graph attention network,AGAT)的短期用户负荷预测模型。首先,针对用户负荷存在规模小、波动性强的问题,通过门控机制结合多个大小不同的扩张卷积核来构造时序特征提取层,从多个尺度上提取用户负荷的高维时序特征。同时,考虑到不同用户负荷间潜在的动态相关性,使用马氏距离构造动态图学习层,生成动态图邻接矩阵。然后,采用图注意力网络根据动态图邻接矩阵将用户负荷的高维时序特征进行信息汇聚。最后,通过全连接层输出用户负荷预测值。为验证AGAT模型的有效性,采用UCI电力负荷数据集进行预测实验,分别与多种基线模型比较。实验结果表明,所提模型预测指标优于各基线模型,有助于提高短期用户负荷预测精度。To improve the accuracy of short-term user load prediction,a model based on an adaptive graph attention network(AGAT)is proposed.First,to solve the problem of small scale and strong volatility of user load,a gating mechanism is used to construct a time sequence feature extraction layer combined with multiple expanded convolution kernels of different sizes,and the high-dimensional time sequence features of user load are extracted at multiple scales.Considering the potential dynamic correlation between different user loads,Mahalanobis distance is used to construct the dynamic graph learning layer and generate the dynamic graph adjacency matrix.Then the graph attention network is used to gather the information of the high dimensional time sequence features of the user load according to the dynamic graph adjacency matrix.Finally,the predicted user load is output through the fully connected layer.To verify the validity of the AGAT model,the UCI power load dataset is used for prediction experiments.The experimental results show that the prediction indices of the proposed model are better than those of various baseline models.This is helpful for the improvement of the accuracy of short-term user load prediction.

关 键 词：短期用户负荷预测 自适应图注意力网络 时序特征提取 动态图学习 图神经网络 

分 类 号：TM715[电气工程—电力系统及自动化] TP183[自动化与计算机技术—控制理论与控制工程]