Non-Intrusive Load Monitoring Based on an Improved Bidirectional Transformer Self-Attention Mechanism Integrated with U-Net

With the rapid development of smart grids and the growing urgency of electricity management across society, there is an increasing demand for advanced energy monitoring technologies. Non-Intrusive Load Monitoring (NILM) is one such technology that can identify different appliance types from aggregate power data and disaggregate their individual energy consumption. However, traditional methods struggle to capture both global and local features, single-appliance training ignores inter-appliance dependencies, while simultaneous multi-appliance training is challenged by appliance switching, which degrades model performance. To address these issues, we propose BT-UNET, an algorithm that combines an enhanced bidirectional Transformer self-attention mechanism with a U-Net architecture for feature extraction, feature fusion, and multi-appliance state detection. Evaluation on the UK-DALE dataset demonstrates that BT-UNET effectively improves NILM disaggregation efficiency.