Application of Statistical Methods to Improving Model Predictions of Rapid Intensification in Tropical Cyclones

The goal of this research is to improve predictions of rapid intensification (RI) in tropical cyclones (TCs) using a statistical algorithm. Forecasting RI in TCs is notoriously difficult because even in generally favorable conditions (e.g., high column water vapor, low vertical wind shear and high sea surface temperature), the occurrence and timing of RI is uncertain. In this talk, we discuss the use of logistic regression as a post-processing tool for numerical weather prediction (NWP) TC models. The baseline NWP model used in this study is the Hurricane Weather Research and Forecasting (HWRF) model based on the 2018 operational configuration. We have extensive reforecasts from 2015-2017 and real-time data from 2018 for both the Atlantic and Eastern Pacific basins. From the HWRF forecasts, we derive predictors describing aspects of a TC’s environment and inner-core, which are tested for inclusion in the HWRF LOGistic regression (HLOG) model. Top HLOG predictors were determined for each basin using objective feature selection and leave-one-year-out cross-validation. Independent testing over the years 2015-2017 gives Brier Skill Scores (BSSs) of 0.30 and 0.36 for Atlantic and East Pacific, respectively. Testing the model on 2018 real-time HWRF forecasts, the BSS values were 0.24 and 0.13 for the Atlantic and East Pacific models. This talk will also compare these results with other probabilistic RI methods. Examining individual cases in the 2018 season, HLOG generally produced higher probabilities for storms that experienced RI compared to storms that did not experience RI. Overall, the HLOG appears to be a promising, simple, and computationally cheap NWP model post-processing technique.