Data-Driven Modelling For Tsunami Forecasting Using Computational Intelligence

Abstract

Numerous tsunami disasters have happened in Indonesia as located across Ring of Fire, and it brings casualties to both the economy and welfare of the people in the event of their occurrence. Frequent tectonic earthquakes in the oceanic areas may lead to tsunami disasters that can cause significant damages to the infrastructures and people. Therefore, the development and implementation of a significantly improved early warning system's performance is essential. This paper presents the research on finding an appropriate machine learning algorithm for provisioning fast and accurate tsunami forecasts using spatiotemporal data of tsunami event in Aceh occurred on December 26th, 2004. A mixture of two modelling paradigms: physically based and data-driven modelling was explored and developed by utilizing 3D numerical models with essential measurement data. The outputs of numerical computations are in the form of time series datasets with various time windows and forecast horizons. Three machine learning algorithms namely fully connected neural network (FCNN), convolutional neural networks (CNN), and recurrent neural network (CNN) with long short-term memory (LSTM) were employed and compared to achieve accurate tsunami wave forecasts, evaluated according to specific set of evaluation metrics. The model forecast comparison with window size of 15 minutes and forecast horizon of 1 minute indicate that FCNN model outperform the CNN and RNN with LSTM models, with RMSE of 0.299. This modelling results show that the proposed modelling framework has been able to support in provisioning towards fast and accurate tsunami early warning system.