Machine Learning Approximation for Rapid Prediction of High-Dimensional Storm Surge and Wave Responses

Storm surge and waves are responsible for a substantial portion of the tropical and extratropical cyclones-induced damage in coastal areas of the USA and Canada. High-fidelity, numerical models can provide accurate simulation results of the water elevation, where a hydrodynamic model (e.g., ADCIRC) is coupled with a wave model (e.g., SWAN). However, they are computationally expensive, hence cannot be employed as part of an early warning system for urban flooding hazards or implemented in probabilistic tropical and extratropical cyclones' risk assessment. In this study, an alternative and efficient approach is proposed based on hybrid machine learning approaches. First, a dimensionality reduction technique based on deep autoencoder is developed to encode the spatial information in a reduced state space. Then, a machine learning-based model is developed in the latent space to predict the maximum surge and significant wave height. The latent space is then decompressed back to the original high-dimensional space using the decoder. The high-fidelity data are retrieved from the North Atlantic Comprehensive Coastal Study (NACCS), released by the US Army Corps of Engineers. Due to its high efficiency and accuracy, the proposed methodology can be employed to analyze the impact of input uncertainties on the simulation results. Four machine learning algorithms are used to predict the maximum surge and significant wave height including artificial neural network (ANN), support vector regression (SVR), gradient boosting regression (GBR), and random forest regression (RFR). The coupled autoencoder-ANN model for the prediction of the storm surge (significant wave height) outperformed all other algorithms with a coefficient of determination R-2 of 0.953 (0.921) for the testing set. In addition, the comparison between deep autoencoder and the widely used principal component analysis (PCA) technique indicated the superior performance of the former since it is able to accurately capture the inherent nonlinearities within the data.