A Statistical-Dynamical Model for Probabilistic Hurricane Intensity Prediction

Abstract

Hurricane intensity prediction at various lead times is essential to provide warnings of approaching hurricanes to coastal communities. A novel probabilistic statistical-dynamical model based on gradient boosted trees and quantile regression is described. A linear regression baseline, based on the elastic net regression, is used to mimic the performance of the linear Statistical Hurricane Intensity Prediction Scheme (SHIPS) model, and evaluate the performance of the probabilistic model. The models are trained on each hurricane season from 2001-2016, using climatology, persistence, and synoptic predictors along the best track of hurricanes occurring from 1989. Results show that the gradient boosted tree model improves on the linear regression model by 7-9% over the 12h to 48h forecasts and 4-5% over the 3-5 day forecasts, across the seasons 2001-2016. The prediction intervals of the probabilistic model show high accuracy. The application of an artificial neural network as the basis of a statistical model is also discussed.