基于Bi-SRNN的联邦学习区域电力短期负荷预测模型  

The Federated Learning Power Short Term Load Forecasting Model Based on Bi-SRNN

作　　者：袁郁 杨超 郑伟铭 林俊鹏 陈新 YUAN Yu;YANG Chao;ZHENG Weiming;LIN Junpeng;CHEN Xin(School of Automation,Guangdong University of Technology,Guangzhou 510006,Guangdong,China;Guangdong Key Laboratory of IoT Information Technology,Guangzhou 510006,Guangdong,China)

机构地区：[1]广东工业大学自动化学院,广东广州510006 [2]广东省物联网信息技术重点实验室,广东广州510006

出　　处：《电网与清洁能源》2023年第10期45-55,共11页Power System and Clean Energy

基　　金：国家自然科学基金项目(U1911401)~~。

摘　　要：随着配电网终端需求多样化和清洁能源的大规模接入,对区域电力负荷的准确预测变得至关重要。在电力市场化改革背景下,客户端倾向于用电信息保存在本地以确保隐私安全。利用天气数据和历史负荷数据,提出面向区域客户端隐私保护的联邦学习双向叠加循环神经网络负荷预测框架。根据短期电力负荷长序列数据之间的强关联性建立基于双向叠加循环神经网络的负荷预测模型。利用联邦平均算法构建基于联邦学习的区域负荷预测框架,将多个利用不同区域客户端负荷数据训练得到的双向叠加循环神经网络的模型进行融合,反复迭代获得全局模型。采用某市96组实时区域电力负荷公开的数据集,对该模型在区域客户端不共享负荷数据条件下的训练效果进行测试,结果表明,所构建模型具有较低的训练耗时和较高的预测精度。With the diversification of distribution network terminal demand and the large scale integration of clean energy,accurate prediction of regional power loads has become crucial.However,in the context of power market reform,clients tend to store their electricity information locally to ensure privacy and security.This paper proposes a federated learning load forecasting framework for regional clients’power load data protection using weather data and historical load data.According to the strong correlation between short-term power loads long sequence data,a load forecasting model based on two-way superimposed recurrent neural network is established,and a subregional load forecasting framework based on federated learning is constructed by using the federal average algorithm.The global model is obtained by fusing multiple two-way superimposed recurrent neural network models trained using load data from clients in different regions frequently.Finally,using 96 sets of real-time regional power load publicly available datasets from a certain city,the effectiveness of the above model in training without sharing load data among regional clients is verified.The results show that the constructed model had lower training time and higher prediction accuracy.

关 键 词：双向叠加循环神经网络 负荷预测 联邦学习 联邦平均算法 

分 类 号：TM714[电气工程—电力系统及自动化]