摘要
针对居民用电负荷识别中数据隐私保护与模型适应性协同优化的挑战,提出了一种基于联邦学习的局部适应型Transformer模型。该模型旨在解决非独立同分布数据对模型泛化能力的挑战。在Transformer高维时间序列结构中嵌入用户级自适应调节机制,并结合全局共享与局部适应的混合参数更新策略,以实现对个体用电特征的动态感知和个性化建模。为增强隐私保护能力,设计了异步联邦聚合策略,避免了用户原始数据集的集中传输,并引入了差分隐私技术以提升隐私保护能力。研究采用REFIT真实家庭用电数据集,覆盖日、周、月3种时间尺度,并在多客户端环境下验证了模型性能。结果表明,该模型在高负荷时段识别任务中表现优异,平均F1分数高于0.97,在不同时间尺度和异构客户端上均展现出良好的稳健性与泛化能力。此外,动态微调测试表明该模型能有效适应用电行为的突变,验证了其在中长期负荷预测和个性化节能服务中的应用潜力。该研究为隐私友好的居民用电负荷建模提供了新的方法支撑,并为构建多区域场景的个性化节能服务体系奠定了基础。
To address the dual challenges of data privacy protection and model adaptability in residential electricity load identification,this paper proposes a residential electricity load identification model based on federated learning and local adaption Transformer model to tackle the generalization issues caused by non-independent and identically distributed data.The model integrates user-level adaptive modulation mechanisms within the Transformer’s high-dimensional time series structure and employs a hybrid parameter update strategy combining global sharing and local adaptation,enabling dynamic perception and personalized modeling of individual electricity usage patterns.To enhance privacy protection,an asynchronous federated aggregation strategy is designed to avoid centralized transmission of raw user data,and differential privacy techniques are incorporated to further strengthen data security.Experiments are conducted on the REFIT real-world residential electricity dataset,covering daily,weekly,and monthly time scales,with multi-client setups to evaluate model performance.Results demonstrate that the model achieves excellent performance in high-load period identification,with an average F1-score exceeding 0.97,and exhibits robust generalization across different time scales and heterogeneous clients.Furthermore,dynamic fine-tuning experiments show that the model effectively adapts to abrupt changes in electricity usage behavior,confirming its potential for medium-and long-term load prediction and personalized energy-saving services.This study provides a novel methodological foundation for privacy-preserving residential load modeling and supports the development of personalized energy management systems across multiple regions.
作者
王彦华
邱洪军
WANG Yanhua;QIU Hongjun(Langfang Normal University,Langfang 065000,China;National Bureau of Statistics of China,Beijing 100826,China)
出处
《供用电》
北大核心
2025年第10期126-135,共10页
Distribution & Utilization
基金
河北省统计科学研究(计划)项目“AI+区块链背景下智慧统计体系建设研究”(2024HY21)。
关键词
居民用电识别
局部适应
差分隐私
Transformer架构
联邦学习
residential electricity identification
local adaptation
differential privacy
Transformer architecture
federated learning
