Machine learning for the analysis of energy systems

Abstract

The present study focuses on evaluating the potential of machine learning models, specifically autoencoders, for providing a simplified, denoised representation of energy systems data by extracting their most important features. Using electricity consumption data from 28 schools in Miyako city in Japan, we show that the extracted features can be effectively used for detecting abnormal consumptions behaviors precisely in time, as well as for clustering purposes from the features space (latent space), which avoids the curse of dimensionality for data with high dimensions.