基于图数据分割的子图集成学习方法及在电价预测领域的应用  

A Subgraph Ensemble Learning Method Based on Graph Data Segmentation and Its Application in Electricity Price Forecasting

作　　者：杨家俊 余涛[1] 余盛灿 陈鑫沛 吴毓峰 卢冠华 YANG Jiajun;YU Tao;YU Shengcan;CHEN Xinpei;WU Yufeng;LU Guanhua(School of Electrical Power,South China University of Technology,Guangzhou 510640,Guangdong Province,China)

机构地区：[1]华南理工大学电力学院,广东省广州市510640

出　　处：《电力信息与通信技术》2023年第8期59-67,共9页Electric Power Information and Communication Technology

基　　金：国家自然科学基金委员会–国家电网公司智能电网联合基金(U2066212)。

摘　　要：目前的日前电价预测模型往往在欧式空间下进行建模,而许多研究表明图神经网络技术在各领域都具有优良的性能,但存在难以叠加多层以及鲁棒性不强等问题,因此,为进一步提升电价预测精度及图神经网络算法性能,提出基于图数据分割的子图集成学习方法,算法首先通过对区域电价多源信息进行图数据建模,形成具有边信息和节点信息的电价图数据,然后借鉴集成学习的思想,通过将电价图数据进行分割,形成多个子图数据,利用图卷积对每个子图进行图学习,最后将每个子图学习结果进行聚合,形成一层多子图学习层,所提方法适用于不同的图卷积核以及不同的下游任务场景,最后为日前电价预测任务构建预测模型。利用美国电力市场的运营数据进行算例分析,通过与对照算法对比及不同的图卷积核对比,证明所提算法具有更好的预测精准度。In the electricity market,accurate day-ahead electricity price forecast can bring greater benefits to suppliers.In order to further improve the accuracy of electricity price forecast,combined with the graph neural network technology that has excellent performance in various fields,this paper proposes a subgraph ensemble learning method based on graph data segmentation.The algorithm firstly use the regional electricity price and multi-source information to model the graph data,which has edge information and node information,then split the graph data into subgraphs,drawing on the idea of ensemble learning,use graph convolution to each subgraph,and then the subgraph learning results are aggregated to multi-subgraph learning layer.This method is suitable for different graph convolution kernels and different downstream task scenarios,and finally a prediction model is built for the day-ahead electricity price prediction task.And using the operational data of the US power market to conduct an example analysis,it is proved that the algorithm has better prediction accuracy anduniversality by comparing with the control algorithm and with different graph convolution kernels.

关 键 词：图神经网络 图数据分割 集成学习 子图分解 子图聚合 电价预测 

分 类 号：F407.6[经济管理—产业经济]