Real-time peak ground acceleration prediction via a hybrid deep learning network

The rapid and accurate prediction of peak ground acceleration (PGA) few seconds after earthquake start is crucial for assessing the potential damage in target areas in impact-based earthquake early warning systems. However, it is difficult to substantially improve the performance of PGA prediction methods based on empirically defined ground motion prediction equations. In this study, we proposed a hybrid deep learning network (HDL-Net) model for PGA prediction based on Japanese and Chinese data sets. The HDL-Net model is capable of extracting useful spatial and temporal features from the input physical feature parameters and three-component waveforms. The test results showed that HDL-Net outperformed the traditional empirical approaches in terms of timeliness and accuracy. To further validate the robustness of the HDL-Net model for PGA prediction, we conducted a potential damage analysis for five earthquakes in Japan. The results showed that the successful alarm rate reached 95.22 per cent, the successful no alarm rate was 100 per cent, and there was no false alarm. The HDL-Net model provides a potential method for earthquake early warning and seismological PGA prediction.