Modeling residential building heating and cooling loads with interpretable symbolic regression

Improving building energy performance is essential for advancing sustainable architecture, particularly through accurate and interpretable predictions of heating and cooling loads. This study employs Symbolic Regression (SR) to model energy demands in residential buildings using the Energy Efficiency Dataset, which contains 768 EnergyPlus simulations. Symbolic models for Heating Load and Cooling Load are constructed based on architectural features such as compactness, surface area, and glazing characteristics. Multiple SR algorithms are evaluated and compared against traditional machine learning methods. Results demonstrate that SR achieves high predictive accuracy while generating concise, human-readable equations that enhance model transparency. The comparative evaluation, conducted against Linear Regression, Random Forest, and Gradient Boosting, relies on Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), and the coefficient of determination (R²) as key performance metrics, positioning Symbolic Regression as both a competitive and interpretable alternative. The interpretable nature of these symbolic models makes them especially well-suited for engineering applications that demand regulatory compliance, design validation, and stakeholder communication. This study fills a critical gap in the literature by demonstrating that Symbolic Regression can simultaneously deliver high predictive accuracy and interpretability in building energy modeling, addressing the limitations of conventional black-box machine learning approaches.