Non-Intrusive Load Monitoring Based on an Efficient Deep Learning Model With Local Feature Extraction
Kaile Zhou, Z ZHANG, Xinhui Lu
Abstract
Nonintrusive load monitoring (NILM) identifies individual appliance power usage within an overall power load, supporting more refined, targeted, and efficient load management. However, feature extraction from the combined power signal may be affected by the number of appliances, appliance types, noise, and other factors. This article introduces TransUNet-NILM, an improved NILM model based on TransUNet, that significantly improves the feature extraction capability by using a residual network and an attention mechanism, improving the ability to discover power variation and temporal features. A sequence-to-subsequence method is proposed to reduce the computational complexity. After testing on the REDD and U.K.-DALE datasets, our model notably improved the F1 score by 1.1% and 8.9%, respectively, compared with the suboptimal model. Experimental results show that the proposed model was both more accurate and efficient in extracting features and identifying power consumption from the aggregate power signal, offering improvements to NILM efficiency and energy management.