DEVELOPMENT OF REAL-TIME PREDICTION METHOD FOR TIME SERIES TSUNAMI WATER LEVEL IN WAKAYAMA PREFECTURE COASTAL AREA BY BI-LSTM

Abstract

 This study develops a real-time tsunami prediction method using Bi-LSTM, a type of deep learning, which is more accurate than LSTM because Bi-LSTM uses time series data in both directions. The network was constructed by 900 tsunami simulation results with different epicenters and slip distributions, using 10-minute observations at 56 offshore tsunami stations as input data and 11 cities and towns along the coast of Wakayama Prefecture as output data. The hyperparameters were set using Optuna, an automatic optimization framework. The results showed that the correlation coefficients of the maximum tsunami heights and the arrival times of the tsunamis were better than 0.8 and 10 minutes, respectively, at all locations. Furthermore, the applicability of the proposed method to the Showa Tonankai earthquake and the Nankai Trough giant earthquake model was confirmed.