Automatic evaluation and analysis of indoor visual comfort for sustainable building design using interpretable ensemble learning

Abstract

Sustainable building design increasingly emphasizes daylight access and glare reduction due to their impact on energy efficiency and occupant comfort. However, integrating daylight distribution with dynamic glare risk from an occupant-centered perspective remains a significant challenge. To address this, this paper develops an interpretable Stacking ensemble framework enhanced with SHapley Additive exPlanations (SHAP) method for automated evaluation of indoor visual comfort (IVC). Six ensemble models are optimized through Bayesian optimization and 5-Fold cross-validation. The final Stacking model, which includes ensemble XGBoost, LightGBM, and CatBoost, achieves high predictive accuracy (R² = 0.911) and efficient prediction capability. SHAP analysis identifies six key design variables accounting for 80.6 % of the model's contribution, with building forms (46.6–52.7 %) and fenestration features (22.6–24.9 %) as primary factors. The framework provides rapid feedback in early-stage design, supporting data-driven decisions to optimize IVC and integrate performance analysis into occupant-centered design processes.