Robust Storm Surge Forecasts for Early Warning System: A Machine Learning Approach Using Adaptive Monte Carlo Bayesian Model Selection Algorithm

Abstract

Abstract Machine-learning based methods are increasingly employed for the prediction of storm surges and development of early warning systems for coastal flooding. The evaluation of the quality of such methods need to explicitly consider the uncertainty of the prediction, which may stem from the inaccuracy in the forecasted inputs to the model as well as from the uncertainty inherent to the model itself. Defining the range of validity of the prediction is essential for the correct application of such models. Here, a methodology is proposed for building a robust model for forecasting storm surges accounting for the relevant sources of uncertainty. The model uses as inputs the mean sea level pressure and wind velocity components at 10 m above sea level. A set of Artificial Neural Networks are used in conjunction with an adaptive Bayesian model selection process to make robust storm surge forecast predictions with associated confidence intervals. The input uncertainty, characterised by comparing hindcast data and one day forecasted data, is propagated through the model via a Monte Carlo based approach. The application of the proposed methodology is illustrated by considering 24 hour target forecast predictions of storm surges for Millport, in the Firth of Clyde, Scotland, UK. It is shown that the proposed approach improves significantly the predictive performance of existing Artificial Neural Network based models and provides a meaningful confidence interval that characterises both model and input uncertainty.