Explaining the Decisions of Deep Learning Models for Load Disaggregation (NILM) Based on XAI

Non-Intrusive Load Monitoring (NILM) techniques estimate the consumption of individual appliances in a household or facility, based on readings of a centralized meter. Lately, deep learning techniques have demonstrated outstanding performance for NILM predictions. Nevertheless, a possible problem is that users and consumers may find it hard to trust the results of such algorithms if they do not fully understand the reasons for their outputs. In this light, this work presents a method that explains and justifies the outputs of NILM convolutional neural network (CNN) classifiers, using Explainable Artificial Intelligence (XAI) techniques. The method operates as follows: a CNN model for NILM is used to estimate which appliances are activated in the system based on the total power consumption. Then, a XAI technique uses this model and its outputs to explain and justify the prediction of this model. Thereby, the NILM CNN classifier outputs are both accurate and are more interpretable, allowing users to make informed and trustworthy decisions. These ideas are demonstrated on the REDD dataset using a convolutional neural network classifier and two state-of-the-art XAI techniques.