Distributionally robust deep network based on [formula omitted]-divergence for non-intrusive load monitoring

Abstract

Non-Intrusive Load Monitoring (NILM) is a crucial technique for energy conservation as it provides detailed energy consumption data for individual household appliances, leading to targeted energy saving strategies. In practice, the startup phase of appliances is more significant than the shutdown phase, especially when predicting appliance-level consumption. When implementing Sequence-to-Point (S2P) based network models, where each appliance is trained separately, the primary source of training error comes from state transitions during startup, particularly those involving other appliances. These transitions can be mistakenly learned by the network as spurious correlations. This paper presents a method for effectively filtering such samples, and extracting features that are more resistant to spurious correlation confusion. We propose a Distributionally Robust Deep Network (DRDN) model grounded in ϕ-divergence, to address the NILM disaggregation problem. Our model incorporates an adaptive loss-scaled robust parameter to enable sample selection and improve decision-making. Our experimental results on the REDD and UK-DALE datasets demonstrate that DRDN models exhibit superior transferability and effectiveness in solving domain generalization problem compared to traditional deep learning approaches across different buildings.