Building Energy Efficiency: Using Ensemble Learning to Forecast the Load on Heating, Ventilation, and Air Conditioning Systems

Currently, many countries are prioritizing energy efficiency in buildings as a crucial component of the global energy transition to decrease dependence on foreign energy and mitigate environmental impacts caused by wasteful energy consumption. Efficient management of energy consumption in buildings helps improve occupant comfort and optimize the use of various consumer products. Previous work in load forecasting has often used slack feature selection methods, limiting their ability to identify the most relevant variables and thus affecting their performance and generalization. Furthermore, few studies have explored the combined potential of hyperparameter tuning and exhaustive variable selection, particularly with ensemble models recognized for their robustness. While complex approaches such as deep neural networks have been used, their tradeoffs between accuracy, overfitting, interpretability, and computational cost remain poorly evaluated. Here, the aim is to develop an innovative approach combining exhaustive feature selection and fine‐tuned hyperparameter optimization to significantly improve the accuracy of load forecasts in future smart urban environments while ensuring interpretability and computational efficiency. Multiple ensemble machine learning approaches, each with distinct model architectures, training methodologies, and performance characteristics, are experimented, including linear regression, decision tree regressor, gradient boosting machine, AdaBoost, XGBoost, light gradient boosting, and CatBoost, as well as a dense neural network with eight key building features: relative compactness, surface area, wall area, roof area, overall height, orientation, glazing area, and glazing area distribution. Findings indicate that XGBoost performs best for heating load predictions, closely followed by gradient boosting machine. For cooling load forecasts, the order is reversed. The dense neural network ranks third in cooling load prediction accuracy. The study also reveals that XGBoost and gradient boosting machine maintain high accuracy and robustness even with reduced input features, providing insights into the most influential building parameters for HVAC (heating, ventilation, and air conditioning) load prediction. The models achieved an improvement of up to 31% in MSE (mean squared error) for XGBoost and 26.2% for GBM (gradient boosting machine), demonstrating their robustness and efficiency. In summary, the proposed methodology is an innovative hybrid framework that combines exhaustive feature selection with an ensemble learning method. It proves to be practical and relatively straightforward to implement, particularly for initial load estimations during the early stages, aiding sustainable building design.