A Cloud-Edge-End Collaborative Self-Training Non-Intrusive Load Identification Architecture for Smart Meters

With the advancement of smart grids and the increasing demand for precise energy management, non-intrusive load monitoring (NILM) has emerged as a critical technology in power systems. Traditional methods that depend on complex deep learning architectures impose substantial computational burdens on resource-constrained embedded devices such as smart meters and are further limited by low-bandwidth communications. To address these challenges, this paper proposes a self-training NILM architecture leveraging smart meters within a cloud-edge-device collaborative framework. Specifically, a lightweight student model is deployed on smart meters to perform preliminary load state inference by processing real-time current and voltage waveforms and extracting active power, reactive power, and key harmonic features. Concurrently, a high-precision teacher model-integrating 1D-CNN, BiLSTM, and Attention mechanisms-is established at the cloud layers to facilitate centralized training and iterative refinement via user feedback. Knowledge distillation is subsequently applied to transfer the learned representations from the teacher model to the student model, while an incremental update strategy based on MAS regularization enables adaptive learning for new electrical devices and complex load patterns. Pilot studies have demonstrated the feasibility and efficacy of the proposed system in terms of data acquisition, model updating, and load identification.