面向跨季度多时段特征双向聚类与时序迁移的多任务短期电网负荷预测  

Multi-task Short-term Power Load Forecasting With Multi-seasonal Time-slot Tiered Bi-directional Cluster and Temporal Transfer Learning

作　　者：潘艳霞 刘国瑞[1] 任建婧 赵堃 谭沛然 马容婷 郝玲 何建樑 PAN Yanxia;LIU Guorui;REN Jianjing;ZHAO Kun;TAN Peiran;MA Rongting;HAO Ling;HE Jianliang(State Grid Shanxi Electric Power Company,Taiyuan 030025,Shanxi Province,China;Marketing Service Center,State Grid Shanxi Electric Power Company,Taiyuan 030025,Shanxi Province,China;School of Mechanical Engineering,Nanjing University of Science and Technology,Nanjing 210018,Jiangsu Province,China;College of Computer and Information Engineering,Henan Normal University,Xinxiang 453007,Henan Province,China)

机构地区：[1]国网山西省电力公司,山西省太原市030025 [2]国网山西省电力公司营销服务中心,山西省太原市030025 [3]南京理工大学机械工程学院,江苏省南京市210018 [4]河南师范大学计算机与信息工程学院,河南省新乡市453007

出　　处：《电网技术》2025年第4期1479-1490,共12页Power System Technology

基　　金：国家电网有限公司科技项目“新型电力系统下计及多市场交易结算品种的电力市场价值体系评估与构建项目”。

摘　　要：在电力系统领域,由于电网用电负荷受到时段性和季节性的用电需求动态变化的影响,使得电网负荷时序数据呈现分布间歇性变化的分布漂移特性。上述现象导致一般的负荷预测模型难以有效的针对动态变化数据进行充分的信息挖掘与利用,降低了电力负荷模型预测的准确度。为此,文章提出了一种融合跨季度多时段的双向聚类与时序迁移的多任务短期电网负荷预测模型。该方法以分层处理的形式,首先通过聚类分析识别出负荷分布差异显著的时间段,利用多任务学习方法对各时间段内序列预测建模,实现信息共享的同时提升预测效果;随后利用时序迁移学习对每个子任务内数据分布差异进行适配,进一步减轻数据分布差异对建模的影响。实验结果表明,与现有主流预测方法相比,所提方法在真实电力负荷预测场景下展现出更优的预测性能,特别是在当数据分布发生显著变化的情况时,预测误差明显减小。所提方法可为电网调度和能源管理提供更可靠的支持。In the field of power systems,the temporal and seasonal variations in electricity demand caused intermittent distribution shifts in the time-series data of grid load.The above phenomenon makes it difficult for the general load forecasting model to effectively mine and utilize the information of the dynamic change data,which reduces the accuracy of the power load forecasting model.To address these problems,this paper proposes a multi-task,short-term power load forecasting model that integrates cross-quarter multi-period bidirectional clustering with temporal transfer learning.The method adopts a hierarchical processing approach:first,it identifies periods with significant load distribution differences through clustering analysis.Then it applies multi-task learning to model the sequence predictions within each period,facilitating information sharing and enhancing prediction performance.Subsequently,temporal transfer learning is used to adapt to the distribution differences within each sub-task,further mitigating the impact of these differences on the modeling process.The experimental results show that compared with the existing mainstream forecasting methods,the proposed method shows better forecasting performance in the real power load forecasting,especially when the data distribution changes significantly,the prediction error is significantly reduced,and it provides more reliable support for power grid scheduling and energy management.

关 键 词：时序双向聚类 迁移学习 协方差对齐 多任务学习 电力负荷预测 

分 类 号：TM715[电气工程—电力系统及自动化]