Item

Abstract

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. However, the potential performance of data-driven models at predicting extreme storm surges, which are underrepresented in observations, is unclear because previous studies did not train their models to specifically predict the extremes. Here, we investigate 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. We find that density-based weighting improves both the error and timing of predictions of exceedances of the 99th percentile made with Long-Short-Term-Memory (LSTM) models, with the optimal degree of weighting depending on the location. At most locations, the performance of the neural networks also improves by exploiting spatiotemporal patterns in the input data with a convolutional LSTM (ConvLSTM) layer. The neural networks generally outperform an existing multi-linear regression model, and at the majority of locations, the performance of especially the ConvLSTM models approximates that of the hydrodynamic Global Tide and Surge Model. While the neural networks still predominantly underestimate the highest extreme storm surges, we 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.