Computing extreme storm surges in Europe using neural networks

This report investigates the performance of neural networks at predicting extreme storm surges at 9 tide-gauge stations in Europe when trained with a cost-sensitive learning approach based on the density of the observed storm surges. Because of the computational costs of computing storm surges with hydrodynamic models, projections of changes in extreme storm surges are often based on small ensembles of climate model simulations. This may be resolved by using data-driven storm-surge models instead, which are computationally much cheaper to apply than hydrodynamic models.

So far, data-driven storm-surge models have primarily been used to predict short time series of local water levels or peak heights during specific events, using the characteristics of tropical cyclones traveling over the region as predictors. While neural networks still predominantly underestimate the highest extreme storm surges, the authors conclude that addressing the imbalance in the training data through density-based weighting helps to improve the performance of neural networks at predicting the extremes and forms a step forward towards their use for climate projections.