An Updated Machine Learning Ensemble for Atlantic Tropical Cyclone Rapid Intensification Forecasting

Prior work has produced a multi-member machine learning ensemble for prediction of rapid intensification (RI) within Atlantic Basin tropical cyclones (TCs), defined as an increase in 1-minute maximum sustained winds of at least 30 kt in 24 hours. However, the ensemble’s performance during the 2017 Atlantic hurricane season revealed important limitations in RI classification. The updated ensemble incorporates predictors from a combination of Statistical Hurricane Intensity Prediction Scheme Rapid Intensification Index (SHIPS-RII) predictors, track data from the National Hurricane Center’s best track (HURDAT2) database, and meteorological fields from the Global Forecast System Final Analysis (FNL) fields for all Atlantic TC timesteps from 1999 – 2016. Feature selection parsed through these 35000+ predictors to identify an optimal combination from which an ensemble of support vector machines (SVMs), random forests (RFs), and artificial neural networks (NNETs) was identified. Ensemble members were identified via cross-validation (a leave-one-year out approach) designed to emulate an operational hurricane season’s RI forecasts (since performance in a given season is a common evaluation metric in Atlantic RI research). The optimal ensemble configuration included 8 SVMs, 6 RFs, and 10 NNETs. The AI forecasts were then combined into a single probabilistic forecast using an SVM, where the individual member predictions were predictors for the SVM. The results yielded considerable improvement over the previous version of the ensemble used to forecast the 2017 season (via increases in Brier and Heidke skill scores), with numeric results more in line with current operational predictive skill found by the National Hurricane Center. and results were in line with current operational models employed by the National Hurricane Center.