A cumulative absolute velocity prediction method based on surrounding strong motion records and deep learning

Cumulative absolute velocity (CAV) is a critical parameter for assessing seismic destructiveness. Existing post-earthquake CAV prediction methods, such as interpolation techniques and ground motion prediction equations (GMPEs), face challenges in simultaneously leveraging historical strong-motion data and real-time observations from strong-motion stations. To address this limitation, this study proposes a novel CAV prediction method based on surrounding strong-motion records and deep learning. The method introduces a station group construction approach, where each group consists of a target station (an unmonitored location) and four surrounding stations with observed data. Using the Japanese strong-motion database, a dataset of 10,463 station groups was established to train the network model. A graph-based feature representation method, designed specifically for station groups, was implemented as the network input. Based on this, a graph neural network (GNN) model, GraphStation, was developed to predict CAV at unmonitored target locations. The performance of the proposed method was compared with interpolation methods and GMPEs, yielding the following key findings: (1) the proposed model achieves a coefficient of determination (R²) of 0.91 for CAV prediction, outperforming existing methods. (2) By training on the station group database and utilizing real-time observations from surrounding stations, the method effectively integrates historical strong-motion data and real-time monitoring data, providing a robust and accurate approach for CAV prediction in regions lacking post-earthquake monitoring data.