A Hybrid Approach Using Hydrodynamic Modeling and Artificial Neural Networks for Extreme Storm Surge Prediction

On coastlines with shallow shelf areas (e.g. North Sea), a combination of high tides, storm surges, wind waves and mutual interactions generally represent the major sources of coastal flood risks: The contribution of the mutual interactions between the various components still remains the least understood, despite advancements in the current operational storm-tide models. The greatest difficulties arise from the fact that most of these nonlinear interactions are physically unknown and that the entire storm-tide system is highly complex and of stochastic nature. A pragmatic data-driven approach, which can use artificial neural networks (ANNs), is required to assess the contributions of these nonlinear interactions to the resulting extreme storm-tide. Such a pragmatic approach is proposed, which is based on two types of extreme water level ANNs models called nonlinear autoregressive eXogenous inputs (NARX): (i) NARX neural network model to predict the extreme storm-tide (Type-A), (ii) NARX neural network model to nonlinearly correct the numerical storm-tide results from TELEMAC2D and TOMAWAC (Type-B). Ensembles methods are then used to reduce variance and minimize error especially in extreme storm-tide events. Several ensemble fitting neural (EFN) network models are developed and tested. The approach was applied for two pilot sites in the North Sea (Cuxhaven and Sylt). The results show that the ensemble models are able to extract the contribution of the nonlinear interaction between the different extreme storm-tide components at both sites by subtracting the results of the hydrodynamic models (linear superposition of storm-tide components) from the ensemble results. In most extreme storm-tide events considered in this study, the contribution of the nonlinear interaction resulted in the reduction of the extreme water levels when compared with the linear superposition of extreme storm-tide components. However, under certain conditions, the nonlinear interactions might result in higher storm-tides than the linear superposition (e.g. storm of January 2000 at Cuxhaven and Sylt).