Perfect prog statistical approach to prediction of Boulder downslope windstorms

Downslope windstorms are a common phenomenon for those living in the lee of the Rocky Mountains. Boulder, Colorado is particularly vulnerable to these dangerous events, which often can bring damaging wind gusts with little or no advance warning. Currently, no numerical modeling technique is used to directly predict these storms. This study is intended to provide the basis for a statistical “Perfect Prog” approach to forecasting, in which operational model forecast output in the form of BUFR soundings at selected National Weather Service (NWS) operational raob sites could be used to predict downslope wind gusts at Boulder.

In this study, eighteen numerical predictors are computed for several downslope windstorm events (events with a wind gust of over 18 m s-1) at Boulder CO from five nearby NWS operational sounding sites. All sounding data are vertically interpolated using cardinal splines to obtain 100m vertical resolution for all five sounding sites. Three different statistical modeling techniques are applied to these predictor datasets, including stepwise linear regression, neural networks, and support vector regression. All techniques are tested on different storm types (prefrontal or postfrontal). Errors of 4 to 6 m s-1 are noted in support vector regression, while errors of 7-8 ms-1 are seen in neural networks and errors of 8 m s-1 noted with stepwise linear regression. Based on results from ten years of training data, support vector regression is found to be the best statistical forecasting technique for Boulder downslope windstorms.