Machine learning prediction and analysis of area-averaged external pressure coefficients for low-rise dual-slope building roofs in a uniform turbulent flow field
Abstract
Accurate prediction of roof area-averaged external pressure coefficients ( $${\overline{C}}_{p,A}$$ ) is crucial for low-rise building design. To isolate geometric and turbulence effects from atmospheric wind shear, wind tunnel experiments were conducted under uniform turbulent flow conditions covering roof pitch angles (0°–30°), wind directions, and turbulence intensities. To overcome the smoothing limitations of conventional regression methods, an ensemble learning framework (Random Forest, XGBoost, and LightGBM) was developed and evaluated using a rigorous Leave-One-Slope-Out (LOSO) cross-validation strategy. Results show that around the roof pitch angle of α = 18.4°, a pronounced variation in aerodynamic pressure response occurs under oblique wind conditions. XGBoost maintains robust predictive performance, outperforming conventional response surface modeling under nonlinear transition conditions. Furthermore, extrapolation failure at α = 30° delineates the practical limits of data-driven models in predicting unseen aerodynamic regimes. Finally, SHAP analysis demonstrates that the model-derived feature attribution patterns are generally consistent with previously reported aerodynamic observations, providing a physically interpretable framework for wind load prediction under complex flow transitions.