Prediction and Interpretability Study of Peak Floor Acceleration of Steel Structures Based on XGBoost

Peak floor acceleration is a crucial engineering parameter for evaluating the seismic damage and post-earthquake functional state of non-structural components in buildings. To address the limitations of traditional dynamic time-history analysis, such as complex physical modeling and cumbersome computations, a data-driven rapid prediction model based on Extreme Gradient Boosting (XGBoost) is proposed, utilizing real steel structure array monitoring records. By taking ground motion intensity and spatial structural characteristics as inputs, a Bayesian optimization algorithm based on the Tree-structured Parzen Estimator (TPE) is introduced to achieve intelligent hyperparameter tuning. The results demonstrate that the optimized machine learning model exhibits excellent prediction accuracy and generalization performance. Furthermore, attribution analysis combining feature importance and the SHAP interpretability framework reveals that peak ground acceleration and normalized height are the core features dominating the amplification of floor responses. Validation through a finite element case study confirms that the model can effectively reproduce the variation of peak floor acceleration along the building height using only macroscopic geometric parameters of the structure. This provides a reliable analytical tool with physical interpretability for the rapid seismic performance assessment of non-structural components.