92nd American Meteorological Society Annual Meeting (January 22-26, 2012)

Sunday, 22 January 2012
Validation of Statistical Classification Methods for Warm-season Convective Wind Forecasting for Cape Canaveral Air Force Station and Kennedy Space Center
Hall E (New Orleans Convention Center )
Bonnie E. Anderson, Plymouth State University, Plymouth, NH

A16-year (1995 to 2010) climatology of 1500 UTC warm-season (May through September) rawinsonde observation (RAOB) data from the Cape Canaveral Air Force Station (CCAFS) Skid Strip (KXMR) and 5-minute wind data from 36 wind towers on CCAFS and Kennedy Space Center (KSC) was used to evaluate statistical forecast methods from previous research and improve on their performance. One specific statistical method examined had previously showed excellent results, but lacked sufficient independent data. This method was re-examined with additional independent data.

Using metrics from previous research, “better” statistical methods were explored, including the previous “best” performing model with different combinations of iterations, algorithm types, and different size classification and regression trees (CART). The models were built using data from the 1995 to 2009 warm-season RAOB data and were validated with an independent data set from the 2010 warm-season RAOB data. The ensemble CART algorithm that this research focused on was the boosting algorithm.

Probability forecasts were also examined, using the ensemble CART models and predictive results. Probability forecasts were of an interest to the 45 WS in order to better the forecasts and forecasting methods on which days were more likely to produce wind gusts associated with convection. The potential for reaching warning thresholds for convective wind gusts was based upon the warning criteria set out by the 45th Weather Squadron (45 WS). These probability forecasts were created with the output from the ensemble CART models and the prediction technique done on each one. Verification was performed on each of the models based on the probability output and the “best” model was chosen from there.

Exploration of different types of the ensemble CART method showed promising results in the way of good forecast verification scores. The ensemble CART model with the boosting algorithm showed the most promise in predictive performance, however only in the binary response of predicting whether a wind gust associated with convection will occur.

Supplementary URL: