Evaluating Machine Learning Models for Sustainable Building Design: Energy, Emissions, and Comfort Metrics

Abstract

This study assesses six machine learning regression models—Random Forest (RF), K-Nearest Neighbors (KNN), Support Vector Regression (SVR), Artificial Neural Network (ANN), Extreme Gradient Boosting (XGBoost), and Linear Regression (LR)—for predicting building performance in a 79.35 m² residential unit in Sari, Iran (ASHRAE Zone 3A). The EnergyPlus model, calibrated with three years of utility data (2021–2023) per ASHRAE Guideline 14, used a synthetic dataset of 1826 configurations with 25 input variables. Five metrics were evaluated: Primary Energy Consumption (kWh), CO₂-equivalent Emissions (kg), Indoor Air Quality (IAQ, ppm), Predicted Percentage of Dissatisfied (PPD, %), and Visual Discomfort Hours (VDH, hr). The dataset was split into 60 % training and 40 % testing sets, with performance measured by RMSE, R², and MAPE. RF and XGBoost excelled, achieving test R² values of 0.9188–0.9578, reducing RMSE by up to 31 % compared to LR (R²: 0.35–0.50). Hyperparameter tuning via Grid Search and Bayesian Optimization improved accuracy, with XGBoost reaching an R² of 0.9578 for IAQ. Sensitivity and SHAP analyses highlighted ventilation and HVAC as key drivers. Scenario analysis with 1000 bootstrap iterations showed trade-offs: increased ventilation increased energy use by 110.7 % and emissions by 76.2 % but improved IAQ by 20.3 %. Optimization reduced energy consumption by 34.3 % and emissions by 38.1 %, enhancing comfort. RF and XGBoost are robust for sustainable building design optimization. © 2025