Non-intrusive load monitoring based on equipment operation state

Abstract Energy disaggregation, known as Non-Intrusive Load Monitoring (NILM), is a practical approach to providing device-level electrical information and can be applied to enhance various scenarios in smart grids. In recent years, with the emergence of large-scale energy consumption datasets, a growing multitude of deep learning methods have been employed to address energy disaggregation problems. However, these methods face challenges in resolving the disaggregation of multi-state devices and devices with overlapping operation cycles, especially during startup and shutdown periods. This paper addresses these challenges by proposing a load disaggregation framework based on equipment operation state. Diverging from existing methods, the load disaggregation task in this paper is divided into two stages. The first stage primarily employs load recognition methods to detect and identify the state transition information of the target appliance. The second stage, based on the identified information from target appliance events, utilizes power filling and power correction methods to achieve load disaggregation. Additionally, a data augmentation method is introduced to alleviate the issue of insufficient sample quantity in the dataset, which hinders model training. Experimental comparisons are conducted on widely used public datasets in the NILM field, namely REDD, and UKDALE, against various state-of-the-art methods. The proposed method achieves an average Mean Absolute Error (MAE) of 3.331 for disaggregating multiple appliances, reducing approximately 38% compared to existing methods. These results robustly validate the advancement of the proposed approach in this paper.