Tuesday, 13 January 2004
Intercomparison of global research and operational forecasts
Limits of atmospheric predictability are investigated through comparison of correlation and error statistics of operational and research global models for two winter seasons. In 1993, bias-corrected models produce anomaly correlations of 0.6 after 6.5 to 7 days, with relatively little forecast skill beyond that point. In 2003, the forecast skill of a more developed, higher resolution operational model has been extended 36 hours, while the skill of the unchanged, low-resolution research model has been extended 6 hours. The relative importance of improved model resolution/physics and improved initial state to the lengthening of forecast skill is diagnosed through evaluation of rms evolution of analyzed and forecast differences of 500-mb height and meridional wind. Results indicate that forecast sensitivity to initial data is less important than sensitivity to model used. However, sensitivity to model used is less important than the rms forecast error of either model, indicating model forecasts are more similar to each other than to reality. In 1993, anomaly correlations of models forecasts to each other reach 0.6 by roughly 8 days, i.e., the models predict each otherís behavior 1.5 days longer than they predict the atmosphere. In 2003, the low-resolution research model has nearly as much skill in predicting the higher resolution operational model forecasts as the latter have in predicting the atmosphere, implying similarity in model errors despite differing model complexities. Correlations of model errors to each other quantify this similarity, with correlations exceeding the asymptotic value of 0.5 through the 14-day forecasts. Investigations of initial state error evolution by wavenumber show long waves (0-15) account for 50% more of the total error growth in 14-day research model integrations than do short waves (16-42). Results indicate the current limit of predictability may be impacted by model sophistication, but error pattern similarities suggest a common deficiency of models, perhaps initial state uncertainty.