An XGBoost-Based Onsite PGV Prediction Model

Peak Ground Velocity (PGV) is one of the parameter commonly used to measure the damage potential of ground shaking to building structures, and real-time prediction of PGV is a key technology in emergency response to major engineering earthquakes. To further improve the accuracy of PGV prediction, in this paper it proposes an onsite PGV prediction model based on Extreme Gradient Boosting (XGBoost). The model takes five characteristic parameters, including peak acceleration (P_a), peak velocity (P_v), peak displacement (P_d), cumulative absolute velocity (CAV), and predominant period (T_pd) in the first 3 seconds of the P-wave observed at the station as inputs, and the PGV observed at the station as the prediction target. 6 918 sets of acceleration records from 102 earthquakes recorded by the K-NET station network in Japan were used for model training, and 3 430 sets of acceleration records from 89 earthquakes were used to test the generalization ability of the model. The results show that, within the same dataset, the PGV prediction model based on XGBoost has a predictive value that is closer to a 1:1 ratio with the actual measured values compared to the PGV prediction models based on P_d and support vector machines. Additionally, the standard deviation of the prediction errors is smaller, the mean of the prediction residuals is closer to zero, and the model performs well on actual earthquake cases in China. The PGV prediction model based on XGBoost can be used for the prediction of peak ground motion in local earthquake early warning systems.