Quantifying the Impact of Spatial Form on Operational Carbon Emissions in High-Rise Residential Districts: A Bayesian Optimization and SHAP Framework

To improve the prediction accuracy of operational carbon emissions and clarify the influence mechanisms of spatial form in high-rise residential buildings in hot-summer and cold-winter regions, this study proposes an integrated framework combining microclimate-coupled simulation, machine learning, and SHapley Additive exPlanations (SHAP) analysis. An automated parametric workflow was developed in Grasshopper to batch-model 116 real high-rise residential districts in Changsha, and district-specific microclimate conditions were generated for each district using Urban Weather Generator to replace the typical meteorological year data in the energy simulation. Key spatial form factors were identified through correlation and multicollinearity analyses, and a Bayesian-optimization-based prediction model for operational carbon emissions was then developed. The BO-SVR model achieved a test set R2 of 0.8479 and an RMSE of 0.8816 kgCO2/(m2·a), outperforming the other five models in the test set comparison. SHAP analysis revealed a three-level “dominant–regulatory–weak response” influence structure, in which the average shape factor was the dominant indicator, followed by a set of regulatory indicators led by the sky view factor, while the remaining indicators showed only weak responses. Feature dependence analysis further identified nonlinear transition points at ASF values of 0.22 and SVF values of 0.37, providing quantitative references for the early-stage low-carbon design of high-rise residential districts in Changsha.