SEMG-Net: State-Event Guided Multi-Scale Gated Network for Non-Intrusive Load Monitoring in Smart Buildings
Non-intrusive load monitoring (NILM) provides a cost-effective way to obtain appliance-level electricity information from aggregate smart-meter measurements and is therefore important for energy management, demand-side response, and sustainable operation in smart buildings. However, accurate appliance-level power disaggregation remains challenging because residential load signals usually involve overlapping appliance signatures, sparse activations, heterogeneous temporal patterns, and transient switching events. To address these challenges, this paper proposes a State-Event-Guided Multi-Scale Gated Network (SEMG-Net) for NILM. The proposed framework integrates a residual temporal encoder, multi-scale dilated convolutional blocks, and a state-event-guided gating mechanism within a unified multi-task learning architecture. The shared encoder extracts hierarchical temporal representations from aggregate mains windows, while task-specific branches jointly estimate appliance power, on/off state, and switching event type. The predicted state probability, three-class event probability distribution, and shared temporal representation are jointly used to construct a continuous gate that modulates the raw power estimate, thereby directly incorporating behavioral predictions into final power estimation. Experimental results on public datasets show that SEMG-Net achieves competitive overall performance, with clear advantages in power estimation, energy consistency, and state identification, particularly for appliances with complex operating stages or transient switching behavior. The ablation results further demonstrate the benefits of multi-scale feature extraction and auxiliary supervision, as well as the effectiveness of the proposed state-event-guided power modulation mechanism.