Modelling and Comparison of Machine-Learning Algorithms for Energy Consumption Prediction in Smart Buildings

Abstract

One-third of global energy demand is attributed to consumption in buildings, with HVAC and lighting systems as the primary contributors. This study presents the development and comparison of several machine-learning algorithms for predicting energy consumption in a building simulated using EnergyPlus and following the Team Data Science Process (TDSP) methodology. Feature-selection techniques (feature selection and feature importance) were applied to identify the most influential variables. Five predictive models were trained: MLP, SVR, XGBoost, Random Forest and Keras Regressor. Results demonstrate that the MLP model achieved the highest accuracy, while XGBoost showed greater stability. Additionally, traditional statistical models (ARIMA and SARIMAX) were compared to machine-learning models for multi-horizon prediction.