Research on Climate Response Strategies for Traditional Dwellings Based on Shapley Additive Explanations and Machine Learning

Abstract

High energy consumption in construction often stems from reliance on mechanical systems for indoor comfort. In contrast, traditional Chinese dwellings use climate-responsive strategies, autonomously regulating comfort and providing insights into energy-efficient design. This study introduces a framework combining machine learning, Bayesian optimization, and Shapley Additive exPlanations (SHAP) to investigate the nonlinear relationship between climate strategies and the adaptability of traditional housing. A case study of a traditional residence in northern Guilin, Guangxi, China, is used to simulate photothermal performance and generate a dataset. With Useful Daylight Illumination (UDI) and Predicted Mean Vote (PMV) as main outputs, an extreme gradient boosting (XGBoost) model optimized via Bayesian Optimization-Tree-structured Parzen Estimator (BO-TPE) achieves a cross-validation coefficient of 0.9968. Comparison among three hyperparameter tuning methods—Grid Search, BO-TPE, and Bayesian Optimization-Gaussian Process Regression—shows that BO-TPE is the most effective. SHAP analysis further highlights patio size, orientation, and buffer space as influential parameters. This study expands research on climate adaptability, exploring energy-saving potential in traditional dwellings, improving design feedback, and enhancing model transparency and interpretability.

Keywords: Traditional Chinese dwellings, Climate adaptation, Machine Learning, SHAP, Bayesian optimization, performance prediction