Statistical Modeling of Downslope Windstorms in Boulder, Colorado

Abstract

Downslope windstorms are of major concern to those living in and around Boulder, Colorado, often striking with little warning, occasionally bringing clear-air wind gusts of 35–50 m s−1 or higher, and producing widespread damage. Historically, numerical models used for forecasting these events had lower than desired accuracy. This observation provides the motivation to study the potential for improving windstorm forecasting through the use of linear and nonlinear statistical modeling techniques with a perfect prog approach. A 10-yr mountain-windstorm dataset and a set of 18 predictors are used to train and test the models. For the linear model, a stepwise regression is applied. It is difficult to determine which predictor is the most important, although significance testing suggests that 700-hPa flow is selected often. The nonlinear techniques employed, feedforward neural networks (NN) and support vector regression (SVR), do not filter out predictors as the former uses a hidden layer to account for the nonlinearities in the data, whereas the latter fits a kernel function to the data to optimize prediction. The models are evaluated using root-mean-square error (RMSE) and median residuals. The SVR model has the lowest forecast errors, consistently, and is not prone to creating outlier forecasts. Stepwise linear regression (LR) yielded results that were accurate to within an RMSE of 8 m s−1; whereas an NN had errors of 7–9 m s−1 and SVR had errors of 4–6 m s−1. For SVR, 85% of the forecasts predicted maximum wind gusts with an RMSE of less than 6 m s−1 and all forecasts predicted wind gusts with an RMSE of below 12 m s−1. The LR method performed slightly better in most evaluations than NNs; however, SVR was the optimal technique.