Data-Driven Sparsification and Multi-Resolution Analysis-Based Framework for Load Identification

The present work focuses on certain pivotal aspects regarding load identification for non-intrusive load monitoring (NILM) using convolutional neural networks (CNN). Primarily, image sets of load signatures are generated via multi-resolution analysis leading to improved image sets compared to the recent relevant works. Secondly, to avoid manual settings for deep learning optimization algorithm hyper-parameters, Bayesian optimization is performed. Lastly, a data-driven strategy based on Taylor's score is considered to significantly compress the CNN architecture. The proposed overall strategy leads to high classification performance and reduced memory footprint.