Predictive optimization of educational buildings' environmental performance under future climate scenarios using Catboost and SHAP

Abstract

<p>Climate change is increasingly impacting building energy consumption and environmental quality, making rapid early-stage performance prediction crucial for optimizing design decisions and improving energy efficiency. This study developed a representative model of educational buildings across three major climate zones in China, simulating performance under four SSP scenarios for 2036–2065 and 2066–2095 using CMIP6 data. A total of 7,200 samples were generated across eight scenarios within each zone. CatBoost was used to efficiently predict and optimize building eco-performance, and results were compared with XGBoost, LightGBM, and Random Forest. Findings show CatBoost outperforms the others on discrete datasets, with an average R² of 0.96, MAPE of 3.68%, and computational speed approximately 600 times faster than traditional simulation-based methods. SHAP analysis indicates that performance is most influenced by the exterior window-to-wall ratio (WWR), glazing type, and orientation. For multi-objective optimization, the NSGA-III algorithm was applied, yielding maximum improvement rates of 35.78% (UDI), 18.29% (UDI-up), 21.12% (ASE), 31.34% (EUI), and 17.49% (PPD) across all scenarios. Optimized results show WWR tends to be below 0.5 in low-emission scenarios and above 0.5 in high-emission ones. Optimal orientation varies by climate zone—favoring south-facing in colder regions and east- or west-facing in warmer ones. The proposed machine learning model enhances building design and energy efficiency by accurately predicting and optimizing educational building performance, supporting climate adaptation.</p>