Exploring the Use of Machine Learning in Correcting GFS Tropical Cyclone Track and Landfall Errors

While tropical cyclone (TC) track forecasting has improved over time, especially in the short term (1-2 day) time period, somewhat large track errors still persist with some of our dynamic models at the medium range (3-5 days). Many advances in medium range track forecasting, including the implementation of Kalman filtering, a new barotropic model, etc., have been offered, which provide improved initial conditions to dynamic model simulations. Few studies have explored the utility of machine learning in correcting medium-range track forecasts from dynamic model simulations of TCs. The purpose of this study is to explore the utility of machine learning in correcting dynamic model simulations of TC landfall location for Atlantic basin TCs. A database of GFS forecasts for all US landfalling TCs spanning 2018-2020 was obtained, with days 3-5 retained to quantify model track errors in landfall locations (as represented by HURDAT2) for these TCs. Once track errors were quantified, an artificial neural network (ANN) was trained using predictors relevant to TC track over a large database of GFS TC simulations. The resulting ANN provides updated landfall locations by correcting GFS track forecasts for both TC advancement speed and direction. Improvements in the TC track were quantified relative to the baseline GFS performance. The results of this project demonstrate the potential utility of improving medium-range TC track forecasting when implemented alongside dynamic model simulations of Atlantic TCs.