Interpretable analytical surrogate modeling for building energy prediction via physically informed feature reduction

Accurate prediction of building energy demand is essential for efficient HVAC system design and energy-aware building operation. Although machine learning models can achieve high predictive accuracy, their limited interpretability restricts their practical use in engineering workflows. This study proposes an interpretable analytical surrogate modeling framework that integrates data-driven learning with explicit analytical formulations for building energy prediction. A systematic feature selection strategy combining correlation analysis, mutual information, and physically informed reasoning identifies a compact subset of meaningful design variables. Machine learning benchmarking demonstrates that this reduced representation preserves predictive performance comparable to the complete feature set while substantially reducing model complexity. A Generalized Additive Model (GAM) is then employed to reveal nonlinear relationships between building design variables and annual energy demand. Based on the learned GAM response functions, explicit analytical surrogate equations are derived using piecewise quadratic and low-order polynomial approximations, enabling rapid benchmark-supported design screening. The transferability of the proposed workflow is further evaluated using an independent EnergyPlus-based ASHRAE prototype building. The results demonstrate that the same methodology can be successfully applied beyond the original UCI Energy Efficiency benchmark, providing an effective balance between predictive accuracy, engineering interpretability, and practical applicability for early-stage building design. Practical application The proposed analytical surrogate framework provides building services engineers with an efficient decision-support tool during the conceptual design stage. By combining physics-informed feature reduction with explicit analytical equations, the framework enables rapid estimation of annual heating and cooling energy demand, visualization of design trade-offs, and benchmark-supported screening of alternative building configurations without repeatedly performing computationally expensive building energy simulations. Rather than replacing detailed simulation-based analyses, the proposed workflow serves as a computationally efficient preliminary screening approach that helps designers identify promising design regions before undertaking comprehensive building performance assessment.