Comparative Study on Machine Learning for Urban Building Energy Analysis

Abstract

There has been an increasing interest in applying machine learning methods in urban energy assessment. This research implemented six statistical learning methods in estimating domestic gas and electricity using both physical and socio-economic explanatory variables in London. The input variables include dwelling types, household tenure, household composition, council tax band, population age groups, etc. Six machine learning methods are two linear approaches (full linear and Lasso) and four non-parametric methods (MARS multivariate adaptive regression spline, SVM support vector machine, bagging MARS, and boosting). The results indicate that all the four non-parametric models outperform two linear models. The SVM models perform the best among these models for both gas and electricity. The bagging MARS performs only a little worse than the SVM for gas use prediction. The Lasso model has similar predictive capability to the full linear model in this case.