Using support vector machines to predict the type and relative severity of severe weather outbreaks

Forecasting severe weather outbreaks has improved considerably in recent years. However, the accurate prediction of the type and severity of these events remains challenging, despite the development and implementation of powerful, high-resolution numerical prediction models. Because of the substantial societal impacts of these severe weather events, methods to quantify a mesoscale model's ability to diagnose the characteristics and meteorological significance of past outbreaks are needed.

Using multivariate indices developed to determine the type and meteorological significance of the top 30 outbreak days from the period 1960 to 2006, the Weather Research and Forecasting (WRF) model is used to forecast 100 training cases and 100 testing cases. A principal component analysis is performed on the forecasts from each set of cases for a number of meteorological parameters associated with severe weather. The principal component scores of the training set are the input matrices for support vector regression, which is used to determine if the WRF can predict the multivariate indices. After a statistical model is developed from the training data, predictions of the multivariate indices for each of the cases in the testing set are compared to the actual values. Results indicate that the WRF is able to predict the multivariate indices accurately for a large number of cases, suggesting the WRF is capable of consistently diagnosing the type of outbreak, as well as its meteorological significance, one day in advance of the outbreak. Sensitivities exist with this capability, depending on the meteorological parameters analyzed and spatial/temporal errors in the simulations.