Rapid Prediction of Tsunami Waveform with Bayesian Scenario Superposition
Please login to view abstract download link
This study proposes a rapid tsunami prediction method by introducing a scenario superposition concept based on the Bayesian theorem. The target site is the Shikoku region (Japan), which is threatened by tsunamis triggered by the Nankai Trough subduction zone. By a numerical simulation, we generate both training and test data consisting of synthetic tsunami wave heights. The test data is then input as an unknown tsunami event into the proposed method, which sequentially estimates/updates each training data-specific weight by solving a linear regression problem with a Bayesian approach. Based on the evaluated weights, the scenario superposition is achieved by a linear combination of each hypothetical scenario in the precomputed database. The proposed method can predict waveforms that have a much larger/smaller amplitude than all the training scenarios by not only interpolation but also extrapolation. Additionally, a probability distribution for the weight parameters is simultaneously obtained, which enables us to quantify and visualize the prediction reliability. We use the proposed and previous methods to forecast waveforms of test data and compare their results. Through the numerical example, it was found that the proposed method could accurately predict waveforms, which could not be predicted by the previous method, within a 10-minute observation duration.
