Statistical modeling of downslope windstorms in Boulder, Colorado

Downslope windstorms are of major concern to those living near the Boulder, Colorado area. These storms often strike with little warning, bringing clear air wind gusts of 35-50 m/s or higher, producing widespread damage across the city. No accurate modeling technique is currently used for forecasting these dangerous events. This study documents an attempt to apply different linear and non-linear statistical modeling techniques to a 10-year mountain-windstorm dataset.

A set of eighteen predictors, based on a decade of data, was calculated for use in this study, including temperature advection, 700 mb geostrophic wind direction, 700 mb geostrophic wind speed, the cross-mountain component of the wind in Denver and nearby sites, mb – 500 mb geostrophic shear direction, ratio of the 700 mb wind speed by the 700 mb geostrophic wind speed at Denver, difference between 700 mb wind direction and the 700 mb geostrophic wind direction, mountain top relative humidity, cross mountain height difference, static-stability ratio, Froude height, integrated Scorer parameter, characteristic impedance ratio, lowest tropopause level at nearby locations, local tropopause height at several nearby locations, postfrontal parameter and Sangster parameter.

Linear regression, neural networks and support vector models were used to relate the predictors to windstorm events. In the linear model, stepwise linear regression was used. It is difficult to determine which predictors are the most important, although significance testing indicated 700 mb flow is highly significant. Both support vector regression and the feedforward neural network did not filter out any predictors but, instead, fit a non-linear function to all predictors, so that no important data were discarded. The study did not conclusively discover any set of these predictors that did a significantly better job of forecasting peak winds than any of the other predictors but, rather, many different methods and predictors can be used to forecast peak winds.

Through comparison of both RMSE and median residuals, it is shown that support vector regression model performed best. Stepwise linear regression yielded results that were accurate to within 8 m/s, whereas a neural network reduced errors of 6 to 7 m/s and support vector regression had errors of only 4 to 6 m/s. 85% of these forecasts based on nonlinear techniques predicted maximum wind gusts with an RMSE of less than 6 m/s, and all of our forecasts predicted wind gusts with an RMSE of below 12 m/s. For comparison, a linear model forecast wind gusts better than 6 m/s 60% of the time, and better than 12 m/s 95% of the time. These results suggest that meaningful improvements can be seen by application of newer non-linear techniques, such as neural networks and support vector regression, to mountain wind forecasts. It is time to apply such techniques operationally.