A comparison between different machine learning techniques for predicting heating energy consumption for residential buildings in a cold climate

Since Russia invaded Ukraine in 2022, the security and sustainability of energy supply have been seriously highlighted. Approximately 90% of an urban context is residential buildings that demand a large amount of heating energy; therefore, predicting energy consumption is essential for successful energy supply and decision-making. This study aims to evaluate machine learning models for predicting the heating energy consumption for residential buildings in a cold climate, focusing on natural gas consumption for space heating and domestic hot water. Linking the building’s physical characteristics to socio-cultural and occupant behavioral characteristics, a novel dataset was developed in which 44 independent relevant variables were analyzed. The results indicate that XGBoost achieved the best performance with an MAE of 2.00, MSE of 2.61, RMSE of 1.61, and R2 of 0.90, followed by RF with an MAE of 1.32, MSE of 2.59, RMSE of 1.61, and R2 of 0.89, while ANN and LR showed lower performance. The feature importance analysis method identified the key variables significantly affecting heating energy consumption; therefore, among the building physics variables, space heating system (HVAC), total unit area, conditioned unit area, building age, and type of thermal insulation were the most effective predictors. Accordingly, among the socio-cultural and occupant behaviors, blocking the cooler channel in the cold seasons was the most effective variable. These findings can guide energy policymakers in designing sustainable heating strategies and assist architects and residents in optimizing energy use for cost savings and efficiency in cold climates.