Edge-Optimized NILM: Combining Structured Pruning and Quantization for Energy Disaggregation

Non-Intrusive Load Monitoring (NILM) enables the disaggregation of total energy consumption, measured in a single household, into the individual energy usage of domestic appliances. This paper presents a novel methodology aimed at optimizing NILM applications by integrating structured pruning and quantization techniques to enhance the efficiency and performance of deep neural networks for edge-enabled deployment. Structured pruning is employed to identify and remove non-essential weights on a layer-by-layer basis while preserving critical feature information and maintaining overall model performance. In parallel, quantization is applied further to compress the model in terms of its memory requirements. The proposed approach is evaluated using the Plegma dataset, a novel resource in NILM research that captures local consumption behaviors and device usage pat terns specific to the Mediterranean region. Experimental results demonstrate significant reduction in the size of the model without jeopardizing inference performance, highlighting the potential of edge-based NILM to contribute to energy efficiency and transition in Mediterranean environments.