Predictive model for peak ground velocity using long short-term memory networks

Abstract

Peak ground velocity (PGV) is a crucial ground motion parameter correlating with earthquake damage. How to quickly predict PGV at a target site is a core issue of earthquake early warning (EEW) system. By using the embedded characteristics in ground motion sequence, a Long Short-Term Memory (LSTM) networks-based onsite PGV prediction model (LSTM-PGV) is proposed in this paper. The LSTM-PGV model consists of three layer of LSTM and one fully connected layer, and the inputs are sequence features of energy-related, amplitude-related, period-related and distance-related P-wave parameters. The performance of the LSTM model on training, validation and test datasets indicates that the model has good generalization capability, and the predicted PGV and observed PGV can meet the 1:1 relationship in general. Compared with Pd-PGV model, a logarithmic linear regression model where Pd is the peak vertical displacement of the first 3 s P-waves, and LSTM-Pd-PGV model, a LSTM-based model with Pd as the sole input sequency feature where Pd is the maximum vertical displacement continuously changing over time, the proposed model predicts PGV more accurately and stably. Furthermore, the issue of underestimation of PGV for larger earthquakes is alleviated in LSTM-PGV model by using longer length of sequence input. The LSTM model is tested with one off-shore earthquake and one inland earthquake in Japan. The results show that the standard deviation of prediction residual goes from 0.417 at sequence length of 3 s to 0.309 at sequence length of 10 s for the off-shore event, and for the inland event the standard deviation decreases from 0.357 to 0.267 at corresponding sequence length. The prediction timeliness measured by lead time, defined as the time interval between the moment when the observed PGV reaches 17.3 cm/s and the moment when the predicted PGV reaches the same threshold, is also discussed for different magnitudes and hypocentral distances. We believe the proposed LSTM model has promising potential in onsite EEW system for providing accurate and timely PGV prediction.