The Use of Machine Learning in Correcting 3-day WRF Simulated Tropical Cyclone Track and Landfall Positions

While Tropical Cyclone (TC) track forecasting has improved over time, especially in the short term (1-2 day) time period, large track errors are still persistent with the dynamical models at the medium range (3-5 days). Some efforts to improve medium range track forecasting have focused heavily in improving dynamic model initial conditions. However, few studies have explored the use of machine learning in correcting medium-range track forecasts provided by dynamical model simulations of TCs. As such, this study will explore the utility of machine learning techniques in improving 3-day landfall location forecasts of United States landfalling TCs. A database of the GFS forecasts three days before landfall for all US landfalling TCs spanning from 2008-2023 was obtained from the HURDAT2 database to calculate the model's track errors in the landfall location. These GFS fields were used to simulate each TC through WRF-ARW (with hurricane model physics) to model the TC track and determine the errors. From there, machine learning (ML) methods (support vector machines and neural networks) were trained using TC track predictors from the WRF TC simulations. The resulting ML provides the updated landfall location by correcting model-derived track forecasts in 3-day TC landfall location. The results of this project demonstrate the potential of utilizing machine learning to improve the 3-day forecast when implemented alongside the dynamic model simulations of Atlantic TCs.