Development of a neural network model to estimate the maximum elevation of storm surge in coastal Virginia

Abstract:

Storm surge by hurricanes is one of the most devastating threats to coastal communities in Virginia. To build resilient coastal communities, a robust probabilistic storm surge hazard assessment is necessary, which requires maximum elevation estimations based on a large number of scenarios. Physics-based numerical models are usually used to estimate the maximum elevation of storm surge. The problem, however, is that these physics-based numerical models are computationally expensive, limiting the number of simulated scenarios available for hazard assessment. Herein, we will present a neural network model that can rapidly estimate the maximum elevation of storm surge in coastal Virginia from hurricane parameters. In this study, we used the pre-computed high-resolution dataset of the U.S. Army Corps of Engineers’ North Atlantic Comprehensive Coastal Study (NACCS). We will demonstrate (1) a pre-processing of imbalanced data with mostly low maximum elevation, (2) a hyperparameter-tuning process to optimize the neural network structure, and (3) dimensional-reduction techniques to reduce the training time.