Adaptive Machine Learning Approaches to Seasonal Prediction of Tropical Cyclones

AbstractTropical cyclones (TCs) are devastating phenomena that cause loss of life and catastrophic damage, owing to destructive winds, flooding rains and coastal inundation from storm surges. Accurate seasonal predictions of TC frequency and intensity are required, with a lead-time appropriate for preemptive action. Current schemes rely on linear statistics to generate forecasts of the TC activity for an upcoming season. Such techniques employ a suite of intercorrelated predictors; however, the relationships between predictors and TCs violate assumptions of standard prediction techniques.We extend tradition linear approaches, implementing support vector regression (SVR) models. Multiple linear regression (MLR) is used to create a baseline to assess SVR performance. Nine predictors for each calendar month (108 total) were inputs to MLR. MLR equations were unstable, owing to collinearity, requiring variable selection. Stepwise multiple regression was used to select a subset of three attributes adaptive to specific climatological variability. The R2 for the MLR testing data was 0.182. The SVR model used the same predictors with a radial basis function kernel to extend the traditional linear approach. Results of that model had an R2 of 0.255 (∼ 40% improvement over linear model). Refinement of the SVR to include the Quasi-Biennial Oscillation (QBO) improved the SVR predictions dramatically with an R2 of 0.564 (∼ 121% improvement over SVR without QBO)