Nonintrusive Load Monitoring Based on Spectrum Information Preserving Data Augmentation Method

Nonintrusive load monitoring (NILM) enables to decompose the aggregate signal of household's electricity consumption into appliance-level data. Data augmentation is an effective way to overcome limited data sets for NILM. However, existing data augmentation strategies may destroy the core information of the target appliance and affect the decomposition results. To this end, a new neural model spectrum information preserving for NILM (SIP4NILM) is proposed. SIP4NILM aims to protect the core information of the appliance by combining the original and augmented input spectrum weighted by preservation map. Experimental results based on the Reference Energy Disaggregation Data Set (REDD) and comparison with five existing data augmentation methods indicate that the proposed method outperform state-of-the-art models, with better generalizability, stability, improving MAE gain and SAEδ gain to unseen appliances.