Non-intrusive load monitoring based on waveform restoration and feature fusion under complex backgrounds

Abstract As a key application of advanced metering infrastructure, non-intrusive load monitoring (NILM) offers real-time feedback on appliance-level energy consumption from aggregated electrical measurements, advising on energy conservation. In daily life, multiple appliances often run together, creating a complex background. However, existing studies underestimate the interferences from complex backgrounds on the features of appliances being identified, resulting in optimistic estimates of NILM performance. Therefore, we propose a NILM method based on waveform restoration and feature fusion. Firstly, we design a denoising auto-encoder model that integrates gated convolution and deep metric learning. It effectively restores nonactive current waveforms and extracts features despite interferences from complex backgrounds. Subsequently, to better distinguish the restored waveforms, the features include both the encoded nonactive current and an image-based representation of the complete current. These features are fused using a neural tree with two hierarchically heterogeneous branches, which adaptively optimize the model structure based on performance. Experiments on the LIT-SYN public dataset and a private dataset demonstrate that the proposed method is more adaptable to complex backgrounds than existing methods.