Rapid prediction of peak storm surge from tropical cyclone track time series using machine learning

Rapid and accurate prediction of peak storm surges across an extensive coastal region is necessary to inform assessments used to design the systems that protect coastal communities’ life and property. Significant advances in high-fidelity, physics-based numerical models have been made in recent years, but use of these models for probabilistic forecasting and probabilistic hazard assessment is computationally intensive. Several surrogate modeling approaches based on existing databases of high-fidelity synthetic storm surge simulations have been recently suggested to reduce computational burden without substantial loss of accuracy. In these previous studies, however, the surrogate modeling approaches relied on a tropical cyclone condition at one moment (usually at or near landfall), which is not always most correlated with the peak storm surge. In this study, a new one-dimensional convolutional neural network model combined with principal component analysis and a k-means clustering (C1PKNet) is presented that can rapidly predict peak storm surge across an extensive coastal region from time-series of tropical cyclone conditions, namely the storm track. The C1PKNet model was trained and cross-validated for the Chesapeake Bay area of the United States using existing database of 1031 high-fidelity storm surge simulations, including both landfalling and bypassing storms. Moreover, the performance of the C1PKNet model was evaluated based on observations from three historical hurricanes (Hurricane Isabel in 2003, Hurricane Irene in 2011, and Hurricane Sandy in 2012). The results indicate that the C1PKNet model is computationally efficient and can predict peak storm surges from realistic tropical cyclone track time-series. We believe that this new surrogate model can enhance coastal resilience by providing rapid storm surge predictions.