Advancing spatio-temporal storm surge prediction with hierarchical deep neural networks

Coastal regions in North America face significant threats from storm surges caused by hurricanes and nor’easters. While traditional numerical models offer high-fidelity simulations, their computational costs limit their use for real-time predictions and risk assessments. Recently, deep learning has been developed for efficient storm surge prediction using storm parameters as inputs. However, resolving small scales of storm surge in both time and space over long durations and large areas often requires large neural networks prone to accumulating prediction errors over time. This study introduces the hierarchical deep neural network (HDNN) technique integrated with a convolutional autoencoder to accurately and efficiently predict storm surge time series. The autoencoder reduces the dimensionality of the storm surge data, streamlining the learning process. The HDNNs then map storm parameters to the low-dimensional representation of storm surge, enabling sequential predictions across different time scales. Specifically, the current-level neural network predicts future states with larger time steps, which are passed as inputs to the next-level neural network for smaller time-step predictions. This process continues sequentially for all time steps. The simulation results from different-level neural networks across various time steps are then stacked to acquire the entire time series of storm surge. The simulated low-dimensional representations are finally decoded back into storm surge time series. The model was trained and evaluated using synthetic data from the North Atlantic Comprehensive Coastal Study (covering critical coastal regions within New York and New Jersey), achieving excellent performance on the test scenarios (root mean square error = 0.055 m, mean absolute error = 0.027 m, and coefficient of determination = 0.966), respectively. The obtained results demonstrate its ability to effectively handle high-dimensional surge data while mitigating error accumulation, making it a promising tool for advancing spatio-temporal storm surge prediction.