基于GCN-Informer模型的电力负荷短期预测  

Short Term Forecasting of Power Load Based on GCN-Informer Model

作　　者：付泉泳 秦骁 张导 吴纯模 莫婷 胡豁然 FU Quanyong;QIN Xiao;ZHANG Dao;WU Chunmo;MO ting;HU Huoran(State Grid Chongqing Information&Telecommunication Company,Chongqing 401121,China)

机构地区：[1]国网重庆市电力公司信息通信分公司,重庆401121

出　　处：《电力大数据》2024年第7期22-34,共13页Power Systems and Big Data

摘　　要：电力负荷短期预测旨在指导电力调度计划的制定、实现供需动态平衡。深度学习在电力负荷短期预测中得到了广泛应用,但是现有的短期预测模型主要利用历史负荷数据,较少考虑多个用户用电行为之间的相互影响。该文利用图卷积网络(graph convolution network,GCN)处理非欧几里德数据的高效性和Informer模型预测长时序列的准确性,提出了一种集成GCN和Informer的预测模型——GCN-Informer。该模型首先基于用户群组之间的相关性构建图卷积网络、提取用户的用电模式,然后利用Informer模型提取用户的历史负荷特征,通过二者集成,学习历史负荷序列的局部隐藏相关性和不同用户用电模式之间的全局潜在相关性,在降低处理复杂程度的同时提高预测精度。利用重庆市某区的用电数据对GCN-Informer模型和6种基准模型进行了对比实验。结果表明,在均方根误差方面,GCN-Informer模型比Transformer模型降低了2%~10%,比长短期记忆网络模型降低了2%-8%;在回归效果方面,GCN-Informer模型的决定系数的平均值为95.42%,是所有模型中最高的。Short-term forecasting(STF)of power load aims to guide the formulation of power dispatch plans and achieve dynamic balance between supply and demand.Deep learning(DL)has been widely applied in power load forecasting.However,existing short-term forecasting models mainly rely on historical load data and rarely consider the mutual influence between multiple user electric power consumption behaviors.This paper proposes an integrated GCN-Informer prediction model,which combines the efficiency of graph convolution network(GCN)in processing non Euclidean data and the accuracy of Informer model in predicting long-term sequences.The proposed model first constructs a graph convolutional network based on the correlation between user groups,extracts users'electric power consumption patterns,and then uses the Informer model to extract users'historical load characteristics.These two features are used to learn the local hidden correlation of historical load sequences and the global potential correlation between different users’electric power consumption patterns,reducing the processing complexity while improving prediction accuracy.An experimental comparison is conducted between the GCN-Informer model and six benchmark models using power load data from a certain district in Chongqing.The results show that in terms of root mean square error,the GCN-Informer model reduces by 2%~10%compared to the Transformer model and by 2%~8%compared to the long short-term memory model.In terms of regression performance,the average R 2 value of the GCN-Informer model is 95.42%,which is the highest among all models.

关 键 词：短期预测 深度学习 图卷积网络 Informer模型 

分 类 号：TM715[电气工程—电力系统及自动化]