Building Energy Consumption Forecasting: A Comparison of Gradient Boosting Models

Abstract

Abstract: Building energy consumption forecasting is essential for improving the sustainability of buildings in the context of addressing climate change. Accurate building load predictions are useful for energy efficient building design selection and demand-side management initiatives. Using historical building energy consumption data has allowed researchers to develop machine learning models to improve the accuracy of such predictions, beyond inefficient traditional approaches otherwise used by the building sector. This work examines gradient boosting machine learning models, namely LightGBM, CatBoost, and XGBoost, for the purpose of comparing their performance on a select dataset. These gradient boosting models are popular in Kaggle machine learning contest solutions but have not been compared formally for the application of building energy consumption predictions. This work applies the three gradient boosting algorithms to a synthesized dataset for a large office building in Chicago. Preliminary results from the presented comparison demonstrate that XGBoost performs better than LightGBM and CatBoost when trained on the selected dataset.