XGBoost-based prediction of on-site acceleration response spectra with multi-feature inputs from P-wave arrivals

On -site earthquake early warning systems use the information from the initial P -wave to quickly estimate the degree of local damage. Existing on -site earthquake early warning models focus on ground motion information and generally ignore the degree of structural response, with many machine learning models characterized by poor interpretability. Therefore, an on -site prediction model based on XGBoost with a continuous time window and multi -feature inputs is proposed. The model uses the response spectrum value, which is closely related to the structural response, as the predicted intensity measure and uses strong motion data from Japanese K -NET stations for training. After using permutation importance to eliminate irrelevant input feature parameters, 11 parameters (Pa, Pd, Pav, Pad, IAV, IV2, Ia, TP, Fmax, TW, T) are determined as the final model input parameters. The results based on a test dataset show that 3 s after P -wave arrival, the mean square error of the model prediction is 15.10*10-4 g. As the input time window is extended, the mean square error of the model prediction gradually decreases, and the overestimation of small values is mitigated. Additionally, 10 s after P -wave arrival, the mean square error decreases by 7.66*10-4 g compared with that at 1 s. After using the PDP and SHAP methods to explain the model, T, Ia, and Fmax were determined as the feature parameters with the greatest impact on the model results. Generalization tests base on large seismic events not included in the training and test datasets indicated that the model has good generalization capabilities.