Predicting the total heating and cooling energy consumption with ultra high-performance fiber reinforced concrete facade panels using explainable machine learning models

Yaren Aydın, Celal Cakiroglu, Gebrail Bekdaş, Laith Abualigah, Zong Woo Geem,
Predicting the total heating and cooling energy consumption with ultra high-performance fiber reinforced concrete facade panels using explainable machine learning models,
Energy Reports,
Volume 14,
2025,
Pages 3254-3275,
ISSN 2352-4847,
https://doi.org/10.1016/j.egyr.2025.10.006.
(https://www.sciencedirect.com/science/article/pii/S235248472500575X)
Abstract: As energy demand rises, energy-efficient buildings are crucial. This study analyzes a dataset of 27,259 points with 40 input values and 1 output value (sum of heating load and cooling load) using machine learning to predict the total heating and cooling load analysis of buildings utilizing ultra-high-performance fiber-reinforced concrete (UHP-FRC) panels for their façade. The dataset comes from energy analysis based on probabilistic simulation. The study utilized and compared 16 different ML-based algorithms. The results showed an R² exceeding 0.99, proving accurate energy consumption predictions. XGBoost, Random Forest, Gradient Boosting Regressor, Histogram Gradient Boosting Regressor, Decision Tree Regressor, and Stacking showed very high accuracy in predicting total heating and cooling loads. Specifically, the coefficient of determination (R²) values of these models were above 0.99, indicating that the predicted values matched the actual measurements almost perfectly. This result demonstrates that the proposed methods can reliably predict energy consumption. Furthermore, the low error values (Mean Absolute Error (MAE)= 25.46, Root Mean Square Error (RMSE)= 53.87, Normalized Mean Absolute Error (NMAE)= 0.0012, Normalized Root Mean Square Error (NRMSE)= 0.0027) of the models prove that the predictions are statistically robust. These results show that ML-based approaches can be used as an effective tool in energy efficiency analysis and sustainable design processes in buildings. Additionally, the key features affecting energy consumption were determined using the SHAP algorithm. SHAP analysis revealed that total floor area and average annual temperature are the most influential inputs in the predictions. Furthermore, the developed graphical user interface (GUI) ensures that the obtained models can be practically used in engineering applications.
Keywords: Energy performance; Machine learning; Shapley additive explanations; Extreme gradient boosting