Bi-Mamba-UNet: A Hybrid Architecture for Energy Disaggregation

Energy disaggregation, or non-intrusive load monitoring (NILM), is a single-channel blind source separation problem that aims to disaggregate aggregated power data into profiles of power consumption for individual appliances. Recent studies have shown that load disaggregation models integrating self-attention mechanisms exhibit superior performance. Nevertheless, the high computational complexity of these models poses a significant challenge for processing long-sequence data. To address this challenge, we introduce a novel hybrid Bi-Mamba-UNet architecture for the NILM task, which combines the bidirectional Mamba model with the U-Net structure. Mamba, a state-of-the-art sequence modeling architecture, achieves near-linear computational complexity through its selective state-space mechanism while maintaining strong capacity for modeling long-range dependencies. Meanwhile, the U-Net architecture effectively captures multi-scale features through its encoder-decoder design, and its skip connections help fuse features from different scales. Experimental results conducted on the UK-DALE dataset demonstrate that the proposed model outperforms baseline methods, achieving a 2.89% reduction in mean absolute error (MAE) and a 7.3% improvement in F1 score.