Predicting Regional Forecast Skill Using Single and Ensemble Forecast Techniques

View More View Less
  • 1 Bureau of Meteorology Research Centre, Melbourne, Australia
© Get Permissions
Full access

Abstract

The potential for predicting the skill of 36-h forecasts from the Australian region limited area model is investigated using three predictors of model forecast error (MFE) for mean sea level pressure. Two of the predictors utilize single forecasts: one is based on statistical regression of the MFE against the initial analysis and the forecast; and the other uses a measure of the degree of persistence in the forecast. The third predictor utilizes the divergence, or spread, of an ensemble of forecasts from other NWP centers.

Based on a 5-month period of daily 36-h forecasts, correlations were found between the above predictors and the MFE of 0.58, 0.18, and 0.40, respectively. Combining the three predictors in an optimal linear manner increased the correlation to 0.71. Further testing of the combined predictors on a 2-month independent dataset produced a correlation of 0.67. Thus, application of the technique to both dependent and independent datasets explained approximately 50% of the variance in the MFE. This demonstrates that the technique has operational utility for differentiating overall poor and good model forecasts. Using case studies concentrating on southeastern Australia, it is further demonstrated that the predictors can provide excellent differentiation of forecast skill across the forecast domain.

Abstract

The potential for predicting the skill of 36-h forecasts from the Australian region limited area model is investigated using three predictors of model forecast error (MFE) for mean sea level pressure. Two of the predictors utilize single forecasts: one is based on statistical regression of the MFE against the initial analysis and the forecast; and the other uses a measure of the degree of persistence in the forecast. The third predictor utilizes the divergence, or spread, of an ensemble of forecasts from other NWP centers.

Based on a 5-month period of daily 36-h forecasts, correlations were found between the above predictors and the MFE of 0.58, 0.18, and 0.40, respectively. Combining the three predictors in an optimal linear manner increased the correlation to 0.71. Further testing of the combined predictors on a 2-month independent dataset produced a correlation of 0.67. Thus, application of the technique to both dependent and independent datasets explained approximately 50% of the variance in the MFE. This demonstrates that the technique has operational utility for differentiating overall poor and good model forecasts. Using case studies concentrating on southeastern Australia, it is further demonstrated that the predictors can provide excellent differentiation of forecast skill across the forecast domain.

Save