Accurate tsunami wave prediction using long short-term memory based neural networks

Abstract

Novel tsunami prediction models based on deep learning technique are used for tsunami wave prediction. The Long Short-Term Memory (LSTM) model and its variants are compared with numerical models and observation data. Three different types of tsunami prediction, namely single tsunami prediction at the same location, single tsunami prediction at different locations, and different tsunami prediction at different locations are considered. Different input–output ratios are examined. The error indicators are utilized to evaluate the reliability and accuracy. It is found that the input–output ratio 2:1 is most suitable for tsunami prediction. However, more inputs may not bring better prediction results under the same input–output ratio. Results show that our proposed Bidirectional Long Short-Term Memory (Bi-LSTM) and Convolutional Neural Networks–Long Short-Term Memory(CNN–LSTM) models are superior to the National Oceanic and Atmospheric Administration (NOAA) numerical model in prediction. This may be because the observed tsunami data contains rich information, and the deep learning model can better grasp various information hidden in the data, thus giving excellent prediction results.