Non-Intrusive Load Monitoring (NILM), a technique that extracts appliance-level energy consumption information through analysis of aggregated electrical measurements, has become essential for smart grids and energy management applications. Given the increasing diversification of electrical appliances, real-time NILM systems require continuous integration of knowledge from new client-side appliance data to maintain monitoring effectiveness. However, current methods face challenges with inter-client knowledge conflicts and catastrophic forgetting in distributed multi-client continual learning scenarios. This study addresses these challenges by proposing a confidence-based collaborative distributed continual learning framework for NILM. A lightweight layer-wise dual-supervised autoencoder (LWDSAE) model is initially designed for smart meter deployment, supporting both load identification and confidence-based collaboration tasks. Clients with learning capabilities generate new models through one-time fine-tuning, facilitating collaboration among client models and enhancing individual client load identification performance via a confidence judgment method based on signal reconstruction deviations. Furthermore, an anomaly sample detection-driven model portfolios update method is developed to assist each client in maintaining optimal local performance under model quantity constraints. Comprehensive evaluations on two public datasets and real-world applications demonstrate that the framework achieves sustained performance improvements in distributed continual learning scenarios, consistently outperforming state-of-the-art methods.
Keywords: Non-Intrusive Load Monitoring (NILM); catastrophic forgetting; confidence-based collaboration; distributed continual learning; lightweight architecture.