SEDR-Seq2P: Lightweight Network for Industrial NILM

A novel lightweight neural network named SEDR-Seq2P has been introduced for multi-task industrial non-intrusive load monitoring (NILM). This research, accessible on arXiv (2607.28693), tackles the issue of adapting models trained on residential data to industrial environments, where performance is hindered by measurement noise and simultaneous machine operations. The authors implement a one-to-many, multi-task disaggregation framework, enabling a single network to estimate various industrial machine loads from total power consumption. Using a standardized evaluation protocol on the IMDELD dataset, they compared Seq2Seq, Seq2SubSeq, Seq2Point, GRU, and WaveNet based on energy-estimation metrics and accuracy-delay criteria. While Seq2Point shows a better accuracy-delay trade-off compared to Seq2Seq and Seq2SubSeq, GRU and WaveNet deliver superior accuracy but at significantly higher computational costs. To bridge this gap, SEDR-Seq2P enhances Seq2Point by incorporating dilated residual blocks and squeeze-and-excitation attention. Compared to the Seq2Point baseline, SEDR-Seq2P achieves roughly a 7% reduction in mean absolute error (MAE) and enhances the coefficient of determination by around 1%. This paper serves as a cross-type announcement, suggesting it may have been presented at a conference or published in a journal, and it advances the field of energy monitoring by providing a more effective solution for industrial applications.