A diffusion model-based framework to enhance the robustness of non-intrusive load disaggregation

Non-intrusive load monitoring (NILM) can extract energy consumption information for each appliance in an economical and efficient way by analyzing the total power signal, which in turn promotes the construction of smart grids. However, existing NILM models require substantial labeled energy consumption data for training. Collecting long-term load data is time-consuming and challenging, and existing datasets often contain significant noise, hindering the models' ability to accurately extract load features. To address these issues, this study proposes a framework to improve model robustness and reduce data requirements. Specifically, we introduce a NILM data augmentation architecture based on a diffusion model. We optimize the diffusion model to generate multi-state, low-noise load data. Using these synthetic data, the decomposition ability of the NILM model can be improved and the need for large amounts of training data can be reduced. Furthermore, this study also designs a loss function and post-processing algorithm that is more suitable for the NILM task to enhance the model's noise resistance and decomposition stability. Experimental results demonstrate that our approach significantly improves the model's MAE, SAE, and F1 score performance in both origin-household and cross-household scenarios.