基于长短期记忆网络模型的联邦学习居民负荷预测  

Resident Load Forecasting Using Federal Learning Based on LSTM Model

作　　者：朱嵩阳 张歌 贾愉靖 白晓清[1] ZHU Songyang;ZHANG Ge;JIA Yujing;BAI Xiaoqing(School of Electrical Engineering,Guangxi University,Nanning 530000,Guangxi Zhuang Autonomous Region,China;Xinxiang Power Supply Company of Henan Electric Power Company,Xinxiang 453000,Henan Province,China)

机构地区：[1]广西大学电气工程学院,广西壮族自治区南宁市530000 [2]河南省电力公司新乡供电公司,河南省新乡市453000

出　　处：《现代电力》2025年第1期129-136,共8页Modern Electric Power

基　　金：国家自然科学基金项目(51967001)。

摘　　要：居民生活用电量在全社会用电量中占比达到15%以上,且用户数量巨大、分布广。对居民负荷进行准确预测有助于需求侧资源整合,满足需求侧响应的要求。现有居民负荷预测方法多为集中式,在服务器和客户端之间需要进行大量数据交换,导致通信成本增加,并引发信息安全问题。基于联邦学习框架,采用长短期记忆网络对居民负荷预测方法进行研究。利用真实居民负荷数据进行仿真计算分析,结果表明,基于联邦学习的居民负荷预测准确率和计算效率优于集中式。此外,将FedAvg、FedAdagrad、FedYogi三种联邦学习策略进行比较,采用具有自适应优化功能的FedAdagrad联邦学习策略对居民负荷预测的准确率更高,收敛性更强。Residential electricity consumption constitutes over 15%of the total social electricity consumption,a significant number of users spread across various locations.Accurate prediction of the residential load is conducive to integrating demand-side resources and meeting the response requirements on the demand side.Most of the existing residential load forecasting methods are centralized,requiring extensive data exchange between the server and the client.This results in increased communication costs and information security problems.Based on the federal learning framework,the long-term and short-term memory network is utilized to study the residential load forecasting method.The simulation results indicate that the residential load forecasting based on federal learning outperforms centralized learning in real calculation efficiency.In addition,we compared our algorithm with the FedAvg,FedAdagrad and FedYogi federal learning strategies,and found that the FedAdagrad federal learning strategy with adaptive optimization function achieves higher accuracy and better convergence in residential load forecasting.

关 键 词：居民用户 集中式 联邦学习 负荷预测 长短期记忆网络 

分 类 号：TM73[电气工程—电力系统及自动化]