Interpretable Building Energy Performance Prediction Using Xgboost Quantile Regression

Abstract

This study focuses on predicting the Building Energy Performance Ratio (BEPR) of 3,594 residential buildings in Istanbul using machine learning algorithms. The main objective is to identify the factors affecting BEPR, examine their influence, and analyze how these factors differ across buildings with low and high energy efficiency. To achieve this, seven machine learning models were evaluated: Multiple Linear Regression (MLR), Decision Tree (DT), K-Nearest Neighbors (KNN), Support Vector Regression (SVR), Artificial Neural Networks (ANN), Random Forest (RF), and Extreme Gradient Boosting (XGBoost). The results show that XGBoost yields the highest accuracy among all models. To improve the interpretability of XGBoost, the Shapley Additive Explanations (SHAP) method was employed, enabling the assessment of the impact of various features (such as wall U-value, window U-value, and building age) on BEPR. The analysis revealed that building thermal properties and age are critical factors in determining BEPR. Additionally, by applying the XGBoost Quantile Regression (XGBoost-QR) algorithm, the distribution of BEPR across different quantiles (low, medium, and high) was analyzed more effectively. This approach demonstrated that the features influencing BEPR vary between buildings with low and high energy efficiency. Specifically, in the lower quantiles, structural features such as wall and window insulation have a greater impact on BEPR, whereas in the higher quantiles, building age and roof insulation become more influential. This research contributes to a better understanding of the determinants of residential energy performance, introduces the integration of XGBoost-QR into energy performance analysis, and offers valuable insights for enhancing energy efficiency strategies.