Shape-Aware Nonintrusive Load Disaggregation via Prior-Guided State-Reused Multibranch Cross-Transformer

Nonintrusive load monitoring (NILM) is a critical technology for smart grids, enabling cost-effective extraction of appliance-level power data. Deep learning has shown promise in NILM tasks, but existing methods struggle with accurately capturing appliance power curve shapes, leading to the underestimation of disaggregated power. This article proposes a novel Shape-Aware NILM framework that uses appliance operating power patterns to achieve superior disaggregation accuracy. Our approach introduces a prior-guided state-reused network (PG-SRN) that utilizes a multibranch cross-Transformer architecture for feature fusion. PG-SRN incorporates prior information about power curve shapes through the appliance operating power patterns, resulting in more accurate disaggregation compared to existing methods. We evaluate our approach on the UK-DALE and REFIT public datasets. The results demonstrate that PG-SRN achieves higher disaggregation performance compared to state-of-the-art methods. Ablation studies further validate the effectiveness of the state reuse strategy (SR) and power pattern prior guidance (PG) for improving shape awareness.