Comprehensive Analysis of Convolutional Neural Network Models for Non-Instructive Load Monitoring

Abstract

Non-Instructive Load Monitoring (NILM) schemes have become more popular in recent years with the availability of smart meters. It provides energy use data to utilities and per-appliance energy consumption details to end users. This study carries out a comprehensive analysis of existing Convolutional Neural Network (CNN) architectures that have been used for NILM. Nevertheless, it provides an unbiased comparison of the existing architectures thereby helping to select the best performing model for NILM applications. The commonly used CNN disaggregation models were categorized into distinctive groups based on their architectures which considered structure of the Neural Network (NN) and outputs. It considers regression-based sequence to sequence and sequence to point mapping, classification-based sequence to point hard association and soft association-based mapping. The CNN models are improved and modified to bring them onto a common platform for comparison. Thereafter, a rigorous comparison was performed using indices which included accuracy, precision, F-measure and recall. The results reveal interesting relationships between architectures, appliances and measures.