Machine learning-based assessment of storm surge in the New York metropolitan area

Storm surge generated from low-probability high-consequence tropical cyclones is a major flood hazard to the New York metropolitan area and its assessment requires a large number of storm scenarios. High-fidelity hydrodynamic numerical simulations can predict surge levels from storm scenarios. However, an accurate prediction requires a relatively fine computational grid, which is computationally expensive, especially when including wave effects. Towards alleviating the computational burden, Machine Learning models are developed to determine long-term average recurrence of flood levels induced by tropical cyclones in the New York metropolitan area. The models are trained and verified using a data set generated from physics-based hydrodynamic simulations to predict peak storm surge height, defined as the maximum induced water level due to wind stresses on the water surface and wave setup, at four coastal sites. In the generated data set, the number of low probability high-level storm surges was much smaller than the number of high probability low-level storm surges. This resulted in an imbalanced data set, a challenge that is addressed and resolved in this study. The results show that return period curves generated based on storm surge predictions from machine learning models are in good agreement with curves generated from high-fidelity hydrodynamic simulations, with the advantage that the machine learning model results are obtained in a fraction of the computational time required to run the simulations.