5:15 PM - 7:15 PM
[SCG63-P03] Development of a real-time seismic intensity prediction model using a graph neural network
Japan is prone to frequent seismic activity, and from the standpoint of disaster prevention and mitigation, there is a growing need for advanced technology to predict seismic intensity in real time when earthquakes occur. In recent years, methods utilizing deep learning, such as Long Short-Term Memory (LSTM) networks, have been employed for rapid estimation of seismic intensity and spatial interpolation. However, in regions where the density of stations is low, improving accuracy remains a challenge.
In this study, we attempt real-time seismic intensity prediction that simultaneously captures the spatial correlations among observation points and their temporal dependencies by combining a Graph Convolutional Network (GCN) specialized for learning on graph structures, a Gated Recurrent Unit (GRU) for time-series analysis, and an attention mechanism. Specifically, we use K-NET observation data from the National Research Institute for Earth Science and Disaster Resilience (NIED) to estimate the real-time seismic intensity several seconds into the future at multiple stations (such as KNG002) around Yokohama, based on past time-series data of real-time seismic intensity. When building the model, we construct an adjacency matrix using a threshold-based Gaussian kernel derived from the distances between stations and use the real-time seismic intensity at one-second intervals as input data.
We evaluated the prediction performance of the A3T-GCN model under two conditions: (1) a small-scale network of five stations, including the target station, and (2) a larger network of about 40 stations located within 40 km of the target station. The results showed that, while there were cases where the errors were slightly larger compared to previous studies using LSTM for the small-scale network, increasing the number of stations and using a sparser adjacency matrix improved the reproducibility of the maximum seismic intensity and the arrival time. We believe this is because different methods of constructing the adjacency matrix change how spatial features are captured, allowing the spatiotemporal learning capability of A3T-GCN to function more effectively.
Moving forward, we aim to enhance the model’s practicality by tuning hyperparameters and expanding the dataset for application to regions with even sparser stations and data containing diverse earthquake characteristics. Additionally, we plan to explore combining this method with existing techniques as well as utilizing simulated data for areas with few stations, in order to realize more accurate real-time seismic intensity prediction over wider regions.
In this study, we attempt real-time seismic intensity prediction that simultaneously captures the spatial correlations among observation points and their temporal dependencies by combining a Graph Convolutional Network (GCN) specialized for learning on graph structures, a Gated Recurrent Unit (GRU) for time-series analysis, and an attention mechanism. Specifically, we use K-NET observation data from the National Research Institute for Earth Science and Disaster Resilience (NIED) to estimate the real-time seismic intensity several seconds into the future at multiple stations (such as KNG002) around Yokohama, based on past time-series data of real-time seismic intensity. When building the model, we construct an adjacency matrix using a threshold-based Gaussian kernel derived from the distances between stations and use the real-time seismic intensity at one-second intervals as input data.
We evaluated the prediction performance of the A3T-GCN model under two conditions: (1) a small-scale network of five stations, including the target station, and (2) a larger network of about 40 stations located within 40 km of the target station. The results showed that, while there were cases where the errors were slightly larger compared to previous studies using LSTM for the small-scale network, increasing the number of stations and using a sparser adjacency matrix improved the reproducibility of the maximum seismic intensity and the arrival time. We believe this is because different methods of constructing the adjacency matrix change how spatial features are captured, allowing the spatiotemporal learning capability of A3T-GCN to function more effectively.
Moving forward, we aim to enhance the model’s practicality by tuning hyperparameters and expanding the dataset for application to regions with even sparser stations and data containing diverse earthquake characteristics. Additionally, we plan to explore combining this method with existing techniques as well as utilizing simulated data for areas with few stations, in order to realize more accurate real-time seismic intensity prediction over wider regions.