Self-Adaptive Loss Weighting for Improving Multi-Appliance Non-Intrusive Load Monitoring

Non-Intrusive Load Monitoring (NILM) is a technology that identifies the operational status and power consumption of individual electrical devices (typically household appliances) from aggregated smart meter readings. This paper introduces a self-adaptive loss weighting strategy aimed at improving the performance of a single deep neural network-based NILM model for disaggregating multiple appliances. The proposed technique adaptively balances the contributions of each appliance, preventing performance biases toward high-power or frequently operating devices. Additionally, we design and implement a multi-appliance adaptive Autoencoder framework to evaluate this mechanism. Experimental results on the REDD and UKDALE datasets show improvements in disaggregating the power consumption of multiple appliances.