Application of statistical models to Boulder windstorm prediction

Downslope windstorms affect Boulder, Colorado each year, causing extensive property damage with gusts of over 35 m/s. Minimal operational techniques exist to deal with these clear-air storms. The scope of this study is to implement statistical models on a training set of Boulder windstorm data in an attempt to predict peak winds from operational forecast soundings.

A set of three statistical models, a stepwise linear regression (LR) model, a support vector regression (SVR) model, and a neural network (NN), were trained using a National Bureau of Standards peak wind dataset valid for Boulder, Colorado, which included peak wind gusts on days where windstorms occurred. Observed soundings from five sites: Grand Junction, Colorado, North Platte, Nebraska, Denver, Colorado, Lander, Wyoming, and Salt Lake City, Utah, were used to compute the 18 parameters for the training set. An independent testing dataset, which was derived from forecast soundings created from 80 km grid spacing output from the North American Mesoscale model (NAM), was applied to the training models.

Overall, SVR had the best root mean square error (RMSE) results, with RMSE over the 10 cases of 8.51 m/s. For NN, RMSE was higher (9 m/s), and LR had a mean RMSE of 11.35 m/s. LR tended to overforecast many events, while NN handled strong wind events well and weak events poorly. SVR tended to handle the weaker events (which are more common) with more consistency, but severely underforecast the strong events. Reasons for these problems include the lack of strong windstorm cases, the use of numerical model output for testing, and the implicit coarsening of 12 km native output NAM to 80 km for our dataset.