Multi-Feature Non-Intrusive Load Monitoring Method Based on Conditional Batch Normalization Combined with ResNet-KAN

Non-Intrusive Load Monitoring (NILM) plays a significant role in energy conservation, emission reduction, and smart grid development by analyzing total power signals to disaggregate the energy consumption of various electrical appliances. However, traditional methods struggle with the complexity of load signals, and although deep learning techniques have achieved notable success, challenges remain in identifying low-power loads and reducing model complexity. This paper proposes a CBN-ResNet-KAN model, which integrates Conditional Batch Normalization (CBN), Residual Networks (ResNet), and Kolmogorov-Arnold Networks (KAN). The model dynamically adjusts feature normalization via CBN, extracts deep features through ResNet, and captures non-linear relationships using KAN, enabling precise multi-feature load identification. Experimental results on the PLAID dataset demonstrate that the CBN-ResNet-KAN model achieves an identification accuracy of 97.93%, with particularly strong performance in classifying low-power devices. This research enhances the identification performance of complex load signals through model innovation, expanding the application scenarios of deep learning in the NILM field.