How do machines predict energy use? Comparing machine learning approaches for modeling household energy demand in the United States

Abstract

This paper illustrates the use of different machine learning techniques to estimate household energy demand. To demonstrate the performance of the techniques, we discuss how the different machine learning algorithms select a model or models of energy usage and we explore how well the models predict usage. Our study employs a high-dimensional dataset of housing, socioeconomic, and behavioral characteristics, provided by the U.S. Energy Information Administration's ongoing Residential Energy Consumption Survey. In addition to discussing the machine learning models, we estimate energy price elasticities, which are important indicators of how sensitive households are to changes in residential energy prices. Given the broad set of data in the survey, we compare and contrast various machine learning techniques to see which model provides the best overall fit to the data. We find that a random forest algorithm performs better than the other machine learning approaches, which include a step-wise Akaike Information Criterion, partial least squares, ordinary least squares, k-nearest neighbors, penalized regression, and gradient boosting methods. Finally, we discuss how machine learning can be used to inform residential energy policies and predict household energy consumption.