隐私保护下融合联邦学习和LSTM的少数据综合能源多元负荷预测  

Integrated energy multivariate load forecasting combining federated learning with LSTM in privacy-protected and low-data environments

作　　者：陈志鹏 张勇 高海荣 孙晓燕 胡荷娟 CHEN Zhipeng;ZHANG Yong;GAO Hairong;SUN Xiaoyan;HU Hejuan(School of Information and Control Engineering,China University of Mining and Technology,Xuzhou 221116,China;Department of Traffic Engineering,Shandong Transport Vocational College,Weifang 261206,China)

机构地区：[1]中国矿业大学信息与控制工程学院,江苏徐州221116 [2]山东交通职业学院交通工程系,山东潍坊261206

出　　处：《智能系统学报》2024年第3期565-574,共10页CAAI Transactions on Intelligent Systems

基　　金：国家重点研发计划项目(2022YFE0199000);国家自然科学基金项目(62133015).

摘　　要：对于用能数据不足的综合能源系统,借助相似系统的丰富数据可以为其建立高精度的多元负荷预测模型,然而,受数据安全等因素的限制,很多系统并不愿意共享自身数据。联邦学习为处理隐私保护下的少数据综合能源多元负荷预测问题提供了一个重要的思路,但是现有方法依然存在相似参与方识别精度不高等不足。鉴于此,本文提出一种融合联邦学习和长短期记忆网络(long short-term memory,LSTM)的少数据综合能源多元负荷预测方法(multitask learning based on shared dot product confidentiality under federated learning,MT-SDPFL)。首先,给出一种基于共享向量点积保密协议的相似参与方识别方法,用来从诸多可用的综合能源系统中选出最为相似的参与方;接着,使用参数共享联邦学习算法对选中的各参与方联合训练,结合LSTM和finetune技术建立每个参与方的多元负荷预测模型。将所提方法应用于多个实际能源系统,实验结果表明,该方法可以在数据稀疏的情况下取得高精度的多源负荷预测结果。For an integrated energy system with insufficient energy consumption data,a high-precision multivariate load forecasting model can be established using data from similar systems.However,due to the limitations of data security and other factors,many systems are unwilling to share data.Federated learning provides an important idea to deal with the problem of multivariate energy load forecast based on a small amount of data under privacy protection.However,the existing methods still exhibit deficiencies,such as low accuracy in identifying similar parties.In this view,a few-data multitask learning based on shared dot product confidentiality under federated learning(MT-SDP-FL)is proposed,combining federated learning and long short-term memory(LSTM).A similar party identification method using a shared vector dot product confidentiality protocol is proposed to select the most similar parties from many available integrated energy systems.Then,the parameter sharing federated learning algorithm is used to jointly train the selected participants,combining the LSTM and fine-tuning technology to establish the multivariate load prediction model for each participant.The proposed method is applied to several energy systems,and the experimental outcomes show that the proposed method can achieve high-precision multi-source load forecasting results in the circumstance of sparse data.

关 键 词：多元负荷预测 综合能源系统 联邦学习 隐私保护 神经网络 少数据 时序数据预测 点积协议 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]