SEDR-Seq2P Improves Industrial Energy Disaggregation Efficiency

Non-Intrusive Load Monitoring (NILM) in industrial settings faces unique challenges due to measurement noise and the simultaneous operation of many machines, which can hinder the generalization of models trained on residential data. This research addresses these issues by focusing on a multi-task disaggregation scenario where a single network estimates the loads of multiple industrial machines from aggregated power data. The study benchmarks several existing NILM architectures, including Seq2Seq, Seq2SubSeq, Seq2Point, GRU, and WaveNet, under a unified evaluation protocol. While Seq2Point offers a good balance between accuracy and delay, and GRU/WaveNet achieve higher accuracy at a much greater computational cost, there was a clear gap. To bridge this, the authors propose SEDR-Seq2P, a lightweight extension of Seq2Point incorporating dilated residual blocks and squeeze-and-excitation attention. SEDR-Seq2P demonstrates notable improvements: it reduces Mean Absolute Error (MAE) by approximately 7%, boosts the coefficient of determination by about 1%, and increases the match rate by 0.8% compared to the Seq2Point baseline. Crucially, it also reduces inference latency by roughly 58% compared to WaveNet, making it highly suitable for scalable industrial deployment where both accuracy and speed are critical.