McCandless, T., S. E. Haupt, and G. S. Young, 2011

Statistical guidance methods for predicting snowfall accumulation in the northeast United States

National Weather Digest, 35, 149-162

Abstract

The accuracy of snowfall accumulation forecasts has widespread economic and safety consequences. Due to the complex structure and dynamics of winter weather systems, snowfall accumulation forecasts tend to have a large degree of uncertainty associated with them. Numerical weather prediction (NWP) ensemble prediction systems were developed specifically to address the uncertainty in weather forecasts. An accurate deterministic forecast as well as an estimate of the uncertainty is of utmost importance to the public; therefore, the ensemble mean is often used as the deterministic forecast and the ensemble spread as the basis for a forecast uncertainty estimate. This approach works for weather parameters directly forecast by the model; however, snowfall accumulation is not directly forecast by the Global Ensemble Forecast System (GEFS). Therefore this study examines nine artificial intelligence methods for producing a 24-h snowfall accumulation prediction from the parameters directly output from the GEFS. These methods are then examined by their deterministic forecast skill using the mean absolute error of the ensemble mean forecast as well as the degree to which the ensemble spread corresponds to forecast uncertainty, which is examined by spread-skill relationships and quantile-quantile plots. Out of the nine methods – an artificial neural network, linear regression, least median squares regression, support vector regression, radial basis function network, conjunctive rule, k-nearest neighbor, regression tree, and an average of these methods–the k-nearest neighbor method produces significantly more accurate forecasts as well as the best calibrated ensemble spread. This postprocessing method would be appropriate for operational forecasting.