Predicting Building Heating and Cooling Loads Using Hyperparameter Optimization and SHAP Based Explainable Machine Learning
This study aimed to predict building heating and cooling loads from geometric and glazing characteristics and to explain the resulting model predictions. The UCI Energy Efficiency dataset was used, comprising 768 building configurations, eight input features, and two continuous targets. First, 14 regression algorithms were compared using test metrics and tenfold cross-validation. The hyperparameters of six algorithms were then optimized using RandomizedSearchCV with fivefold cross-validation. Performance was assessed using RMSE, MAE, MAPE, and R². Histogram Gradient Boosting achieved the lowest test RMSE for both targets in the initial comparison, whereas Gradient Boosting performed best after optimization. For heating load, the optimized model achieved an RMSE of 0.34048, an MAE of 0.24123, a MAPE of 1.14%, and an R² of 0.99889. The corresponding cooling-load values were 0.81253, 0.51096, 1.84%, and 0.99287. SHAP analysis identified relative compactness, surface area, and overall height as the leading contributors to heating predictions, while roof area, surface area, and relative compactness were most prominent for cooling. The results indicate that optimized boosting models can predict thermal loads with low error within the simulated configurations examined. Strong dependence among geometric features requires their SHAP contributions to be interpreted jointly. By combining predictive performance and model explanations within a common evaluation framework, the study provides an application of explainable machine learning to building energy assessment.
Authors
- Süleyman Dal (ORCID: https://orcid.org/0000-0002-4564-8076)
Institutions
- Batman University (TR)
Publication Details
- Journal
- European Journal of Technic
- Published
- 2026-11-14
- DOI
- https://doi.org/10.36222/ejt.2074512
- Primary Topic
- Building Energy and Comfort Optimization
- Type
- article
- Field-Weighted Citation Impact
- 0.00