NILM: Multivariate DNN performance analysis with high frequency features

Abstract

In recent years we have seen deep neural networks (DNNs) appear in almost every signal processing problem. Non Intrusive Load Monitoring (NILM) was not an exception. A detailed evaluation of the supervised deep learning approach can provide powerful insights for future applications on the matter. In this work we improve a state of the art NILM system based on DNN, by including high frequency features and modifying the autoencoders latent space dimension. Moreover, we introduce a novel dataset for evaluating NILM systems. This paper presents a contribution that adds relevant features as a multivariate input to the DNNs, based on high frequency measurements of the power. Furthermore, a thorough evaluation of the generalization capabilities of these models is presented, comparing results from public databases and those acquired locally in Latin America (LATAM), an underrepresented region on the NILM problem. The data and software generated are left of public access.