A transfer-learning framework to alleviate data scarcity in cross-slope wind pressure modeling

• Four high-performance ensemble learning models are employed to predict the wind pressure distribution on low‑rise building roofs. • Model uncertainty is quantified through the coefficient of variation (CV) and bootstrap method. • The proposed transfer learning framework, built upon ensemble methods, accurately forecasts the wind pressure distribution for target structures with limited data. • Advanced SHAP analysis is integrated with the developed models to provide comprehensive interpretability of their predictions. In the context of rapid urbanization and growing incidence of extreme wind events, wind pressure prediction on low-rise building roofs has gained considerable research attention. However, there remains a notable lack of systematic and scalable prediction studies that account for different roof slope conditions. To address this gap, there is a pressing need for modeling approaches that achieve an optimal balance between predictive accuracy and engineering applicability. This research proposes and validates a transfer learning-based modeling framework for roof wind pressure prediction and evaluates its scalability and practical utility across varying roof slopes. The study begins by pre-training four ensemble learning models—Random Forest, XGBoost, LightGBM, and CatBoost—using source domain data from a 3:12 roof slope. Model performance and robustness were rigorously evaluated using five performance metrics, the coefficient of variation (COV), and uncertainty quantification techniques. Results demonstrate that CatBoost delivered optimal performance in both predictive accuracy and uncertainty control, achieving R² > 0.99 and a mean COV below 5%. Subsequently, the CatBoost model was fine-tuned using a limited target domain dataset (1:12 roof slope), and its transfer efficacy was validated on four independent validation cases. Detailed error analysis confirmed that relative errors at most measurement points remained below 10%. Building on this high-performing model, a novel two-stage CatBoost framework was developed to accurately predict peak wind pressure coefficients. To enhance model interpretability, SHapley Additive exPlanations (SHAP) were employed for both global and instance-level analysis, revealing complex feature interactions and underlying physical mechanisms. Collectively, this research demonstrates the efficacy of cross-slope transfer learning combined with fine tuning strategies in enhancing wind pressure prediction under data-scarce conditions. The proposed framework offers an actionable, engineering-oriented methodology to support roof system reliability assessment and real-time decision-making in digital twin applications.