This study proposes a machine learning-based framework to predict daylight glare caused by reflected sunlight from neighboring buildings, with a focus on real legal dispute cases in residential areas of Korea. This study was conducted in two main steps: STEP 1 - data acquisition and STEP 2 - machine learning model training. STEP 1 was carried out in four detailed sub-steps. First, a dataset was generated by varying the spatial configurations and optical properties of reflecting buildings (RB) and affected buildings (AB). Second, the critical region where glare is most likely to occur was identified by calculating the reflected sunlight vectors. Third, features such as incident angle and solid angle of glare facade were derived through feature engineering. Fourth, Radiance simulations were conducted to calculate the Daylight Glare Probability (DGP). Based on the data obtained in STEP 1, STEP 2 was carried out in two sub-steps. First, continuous DGP values were predicted by an XGBoost regression model, and glare/non-glare were predicted by an XGBoost classification model. Second, SHAP analysis was conducted to identify the key influencing variables based on the prediction results of the XGBoost regression model. The proposed model achieved high accuracy with significantly reduced computation time compared to conventional simulations, making it practical for early-stage design assessments. The prediction model was generalized by testing cases with different RB sizes. This study proposed a methodology to assess glare risk due to reflected light in residential areas, and the developed prediction model is expected to contribute as a model that can be predicted quickly, has high prediction accuracy, and is interpretable.
