A 2-D-Variation Modeling of Aggregated Load Data With Times Block for Nonintrusive Load Monitoring

Non-intrusive load monitoring (NILM) is designed to disaggregate the total load of electricity used in a household into appliance-level data. Moreover, deep learning (DL) models have been extensively studied in the context of NILM. However, existing DL models are typically studied on one-dimensional (1D) time series, which limits the representational capacity of time series data from electrical appliances. In this paper, considering the multi-periodicity features of the appliances’ power, a neural model, TimesNet for NILM (TimesNet4NILM), is proposed. In the proposed model, the one-dimensional load signals are converted into two-dimensional (2D) signals representing multiple periods, and the information across these periods is captured through the visual backbone. Hence, the temporal information of appliances within the aggregated load can be captured, enabling more detailed state detection and power decomposition. Meanwhile, a spectrum information preserving (SIP) data augmentation (DA) method is designed to address the problem that existing DA methods may destroy the core information of the target appliance. Experimental results based on the public datasets demonstrate the effectiveness and superiority of the proposed TimesNet4NILM model. Comprehensively, regarding the disaggregation accuracy metrics, improvements of 41.40% and 26.20% are achieved for MAE and SAE, respectively.