Federated Sequence-to-Sequence Learning for NILM with Heterogeneous Data
Xiangrui Li, Yunqi Wang, Yang Gao, Bo Jie · 2025
The increasing recognition of the importance of Non-Intrusive Load Monitoring (NILM) stems from its capability to improve energy awareness and offer valuable insights for the formulation of energy programs. Many existing NILM methodologies often depend on specialized equipment to capture complex high-frequency signal data, which limits their feasibility in practical applications, especially when smart meters only provide low-frequency active power data for residential use. In this study, we introduce a novel method that leverages readily available weather data to enhance the feature set of low-sampling NILM models. We formulate a federated learning (FL) model, adapted from a sequence-to-sequence model, to disaggregate extremely low sampling rate smart meter records without the need for data exchange. Our experiments demonstrate that the FL framework, particularly FedProx, effectively handles statistical heterogeneity and alleviates concerns related to overfitting. By integrating weather data, our approach markedly enhances NILM performance while maintaining the confidentiality of user data. The experiment shows our proposed sequence-to-sequence model improves the Eacc from 0.8365 to 0.8557 on FedAvg framework and from 0.8470 to 0.8564 on FedProx.