On the use of support vector machines to evaluate the capability of the WRF to discriminate tornadic and nontornadic outbreaks

One of the primary goals in forecasting severe weather outbreaks is to determine whether the event will feature the development of numerous long-lived and significant tornadoes. Much progress has been made in recent years in the forecast of these outbreaks, but the need for more accurate predictions of outbreak type remains. This study will present an objective methodology for determining a mesoscale model's ability to discriminate outbreak type one, two, and three days in advance of the outbreak.

Fifty tornadic and fifty primarily nontornadic outbreaks are selected, and simulations for each of the three forecast periods are performed using the Weather Research and Forecasting model (WRF). Input data for model initializations are of the synoptic scale to investigate the degree to which synoptic-scale processes influence the occurrence or absence of tornadoes in these outbreaks. Because the model is not able to resolve tornadoes explicitly, numerous meteorological parameters associated with severe weather and tornadoes (known as covariates) are analyzed to determine if the model differentiates the two types of outbreaks. Principal component analyses are performed on the model output for the parameters (individually and in combination), a large subset of the cases is selected to train a support vector machine, and the remaining cases are used to test the proposed function to be implemented to distinguish the two types of outbreaks. The process is repeated using a jackknife technique, so that each case is tested an acceptable number of times with different combinations of trained cases.

Results of these analyses indicate that synoptic-scale processes play a substantial role in the occurrence or absence of tornado outbreaks, and that the WRF is capable of consistently predicting the outbreak type up to three days in advance. “Synoptic” parameters in the low levels of the atmosphere (such as geopotential height and mean sea level pressure), as well as low-level shear parameters (e.g., 0-1 km storm-relative environmental helicity), are most helpful in distinguishing the two types of outbreaks. There is a strong signal of seasonal dependence in the results, which necessitates careful selections of cases. The implications of these results for severe weather forecasting and suggestions for utilization of this technique for other forecasting problems will be discussed.