基于无监督域适应的非介入式负荷监测方法

打开文本图片集

中图分类号：TB9;TM715 文献标志码：A 文章编号：1674-5124（2025）09-0167-09

Abstract： To addressthe distribution discrepancy between synthetic and real-world data and the lack of labeled samples in the target domain for non-intrusive load monitoring （NILM），an unsupervised domain adaptation method based on feature reconstruction is proposed. The method mitigates distribution shift by disentangling domain-invariant and domain-specific featuresand enhances generalization through the integration of an external attntion mechanism，enabling label-free transfer from synthetic to real data. Experimental results show that the proposed method reduces the disaggregation error for dishwashers and microwaves by 52.5% and 88.0% ，respectively， on the UK-DALE dataset. In intra-domain transfer tasks for periodic appliances such as refrigerators， themean absolute error decreasesby 44.3% .Furtheranalysisindicates thatmodel performanceis influenced by device power characteristics and inter-domain distribution divergence，with disaggregation accuracy significantly degrading when the Jensen-Shannon divergence exceeds 0.8.This method provides an effective solution for energy disaggregation under low-label scenarios.

Keywords： non-intrusive load monitoring; load disaggregation; domain adaptation; transfer learning; JensenShannon distance

0引言

全球能源需求的持续增长、能源短缺以及环境恶化被认为是未来几十年内的重大全球性问题。（剩余12888字）