PSSR-Net: A Pretraining Network for NILM Based on Power Sequence Shift Recognition

Nonintrusive load monitoring (NILM) technology estimates individual appliance consumption by analyzing the total consumption collected from smart meters, enabling efficient energy management and consumption analysis. However, current deep learning (DL)-based NILM methods face generalization challenges in different households due to diverse appliance compositions and usage patterns. Additionally, these methods require extensive synchronized training data, incurring substantial data collection and storage costs. This article proposes a power sequence shift recognition network (PSSR-Net) with a novel pretraining approach: first, synthesize training data by adding appliance power sequences to background power signals with specific time shifts. Then, the DL model pretrains by predicting these shift values. Finally, the pretrained model transfers to new households through fine-tuning with limited real data. Unlike traditional methods requiring synchronized power data, our proposed PSSR-Net uses nonsynchronized data to pretrain a backbone network with strong multiappliance feature extraction capabilities, significantly reducing data collection costs while improving performance. Experimental results on two datasets demonstrate that the PSSR-Net exhibits superior generalization ability when deployed to new households, achieving a 29.6% reduction in mean absolute error (MAE) and a 46% reduction in signal aggregate error (SAE) compared to existing pretraining methods. Furthermore, the method maintains effectiveness even when using lower sampling rate data for pretraining.