Estimating Electricity Consumption of Buildings Using Information Theory and Machine Learning Methods

Abstract

Estimating electricity consumption of buildings is an essential task in building energy management to enable decision makers to compare different energy efficiency measures and control strategies. With the growth in available building energy data, many opportunities have been emerged to apply Machine Learning (ML) methods for prediction of building energy performance. The objective of this paper is to (1) analyze and evaluate the influence of different building features on electricity consumption, and (2) evaluate the predictive performance of different ML methods in estimating electricity consumption of buildings with various occupancy types, surface area, locations, and climates. To this end, feature importance assessment is performed using Mutual Information (MI) from information theory. Moreover, four frequently used ML methods, including decision tree, random forest, multilayer perceptron, and gradient boosting are implemented to predict building electricity consumption using the latest version of Commercial Buildings Energy Consumption Survey (CBECS) data. The primary contributions that this research adds to the body of knowledge are (1) the application of information theory to evaluate the influence of different building features on electricity consumption, and (2) the development of new ML models to predict electricity consumption of buildings with different characteristics such as occupancy, surface area, location, and climate. Based on MI analysis square footage, number of employees, cooling equipment type, and principal building activity had the highest influence on building total electricity consumption. Moreover, random forest had the best performance with the coefficient of determination of 0.79 among other ML methods.