Using Neural Networks to Predict Hurricane Storm Surge and to Assess the Sensitivity of Surge to Storm Characteristics

Hurricane storm surge represents a significant threat to coastal communities around the world. Here, we use artificial neural network (ANN) models to predict storm surge levels using hurricane characteristics along the US Gulf and East Coasts. The ANN models are trained with storm surge levels from a hydrodynamic model and physical characteristics of synthetic hurricanes which are downscaled from National Centers for Environmental Prediction (NCEP) reanalysis using a statistical-deterministic hurricane model. The ANN models are able to accurately predict storm surge levels with root-mean-square errors (RMSE) below 0.2 m and correlation coefficients > 0.85. The ANN models trained with the NCEP data also show satisfactory accuracy (RMSE below 0.7 m; correlation > 0.7) in predicting storm surge levels for hurricanes downscaled from future climate projections. Once trained, we use the ANN models to assess the sensitivity of storm surge levels to variations in hurricane characteristics and local geophysical features. Progressively stronger maximum wind speeds and larger outer radius sizes independently increase storm surge levels at all locations along the US East and Gulf Coasts. The response of storm surge levels to changes in hurricane translation speed, however, is found to be sensitive to coastal configuration, with increases in hurricane translation speed amplifying (reducing) storm surge levels in open ocean (semi-enclosed) regions.