Non-intrusive load monitoring (NILM), an anticipated technology in smart grids, has recently provided innovative solutions for energy savings and electrical safety. However, existing event-based NILM methods often ignore the impact of voltage fluctuations, line losses, and variations in load impedance characteristics during waveform extraction. It causes various distortions in the extracted waveforms that are difficult to collect beforehand, significantly reducing the performance of the load identification model. Therefore, we analyze waveform distortion as multi-source noise interference and propose a two-stage robust non-intrusive load identification method against unknown waveform distortions. The proposed method includes a general denoising module based on a dual-mapping mechanism and a robust load identification model based on non-local residual shrinkage modules. It achieves robust load identification under unknown noise interference by first cyclic denoising distorted waveforms and then enhancing the model's ability to capture key features. Experiments on three public datasets and one private dataset show that the proposed method maintains the best performance even under low signal-to-noise ratios. The denoising module can also be cascaded with other identification models in a plug-and-play manner to improve robustness. This is particularly desirable for edge-deployed in residential e-bike charging safety monitoring, campus dormitory prohibited-appliance detection, and office energy management.
Guo et al. (Sat,) studied this question.