Forecasting the Skill of a Regional Numerical Weather Prediction Model

Abstract

It is demonstrated that the skill of short-term regional numerical forecasts can be predicted on a day-to-day basis. This was achieved by using a statistical regression scheme with the model forecast errors (MFE) as the predictands and the initial analysis, together with the model forecast, at proximate points, as the predictors. In a first attempt to assess the utility of the method, the technique was applied in a long-term quasi-operational trial to 24 h forecasts of mean sea level pressure in two seasonal periods (one summer and one winter period) on the Australian region forecast domain. Correlation coefficients were computed between the predicted and observed root-mean-square (rms) MFE and were found to be 0.54 and 0.51, respectively, averaged over the full region, for the 90-day summer and winter periods. Using standard Student's t-tests these correlations were shown to be highly significant. In addition, the regional forecasts were divided into four categories of rms MFE, and were verified against the observed rms MFE. Using a contingency table skill score (relative to chance), it was demonstrated that the category forecasts exhibited a very high level of skill. The procedure also was applied to subdomains of the Australian region grid and it was found that the predictions of model forecast skill were improved further for these local forecasts.