Use of a WRF-ADCIRC Ensemble to Explore Storm Surge Predictability Associated with Hurricane Irma (2017)

On 10 September 2017 Hurricane Irma made its final landfall near Marco Island in southwest Florida. The storm’s immense wind field produced a substantial storm surge along the northeast Florida coast and as far north as South Carolina. However, Irma’s eastward jog before landfall substantially reduced storm surge inundation closer to the landfall location. While a storm surge of 7.8 feet was recorded at Fernandina Beach in far northeast Florida, storm surges recorded in southwest Florida were lower than expected. Indeed, both Naples (5.1 feet) and Fort Myers (3.9 feet) had substantially less surge (Cangialosi et al. 2018). Clearly, Irma provides a challenging case for predicting the distribution of storm surge.

Predictability of the storm surge distribution associated with Irma is explored here using an ensemble of 51 Weather Research and Forecasting (WRF) simulations of the storm to force an ensemble of ADvanced CIRCulation (ADCIRC) storm surge simulations. Ensemble forecasts from the European Centre for Medium-Range Weather Forecasts (ECMWF) provide initial and boundary conditions for the 51 convection-permitting WRF simulations that are initialized several days before Irma’s Florida landfalls. The Stochastic Kinetic Energy Backscatter Scheme is employed to ensure that there is sufficient spread in the ensemble track and intensity evolutions of Irma. Output from each WRF simulation is used as input data for a corresponding ADCIRC storm surge simulation to model the water levels and coastal inundation associated with each evolution of Irma.

The spread of track and intensity within the WRF ensemble is compared to the ECMWF ensemble from which it was derived; intensity of Irma among the WRF simulations is also compared to Irma’s actual evolution. Results from the 51 ADCIRC simulations are analyzed to investigate the predictability of Irma’s surge at different locations and the sensitivity of coastal inundation to variations in the atmospheric ensemble. Finally, regression mixture models are used to partition the tracks of Irma among the 51 WRF simulations into clusters, and ADCIRC simulation results from each cluster are compared to evaluate the inter-cluster and intra-cluster variations in surge.