Warm-season Thermodynamically-driven Rainfall Prediction with Support Vector Machines

Dynamic numerical weather prediction models have been designed to deal with large-scale, highly predictable midlatitude atmospheric patterns. However, the capability of these models to simulate thermodynamically driven warm-season rainfall events, such as afternoon airmass thunderstorm formation in subtropical summers, is highly limited. Current methods of addressing this issue have included ensemble numerical weather prediction simulations, where an ensemble mean of multiple simulations with varied model physics is used as an improved prediction over any individual ensemble member. These approaches still yield only modest skill primarily due to inherent biases in each ensemble member. As such, the current research will utilize machine learning to combine logically ensemble members into a single prediction of warm-season rainfall. In particular, a support vector machine classification scheme that employs members of a 30 member ensemble as predictors and observed rainfall patterns as a predictand will be formulated on multiple warm-season rainfall days in an effort to develop an improved prognosis of warm-season rainfall that can be implemented in operational meteorology forecasts. The primary goal of the work is to obtain a statistically significant improvement of predictive skill over currently utilized ensemble member approaches.