VERIFICATION OF ETA-RSM SHORT-RANGE ENSEMBLE FORECASTS

Motivated by the success of ensemble forecasting at the medium range, the performance of a prototype short-range ensemble forecast system is examined. The ensemble dataset consists of 15 case days from Septembe r 1995 through January 1996. There are 15 members of the ensemble, 10 from an 80-km version of the era model and five from the regional spec tral model. Initial conditions include various in-house analyses avail able at the National Centers for Environmental Prediction as well as b red initial conditions interpolated from the medium-range forecast ens emble. Forecasts from the 29-km mesoeta model were archived as well fo r comparison. The performance of the ensemble is first evaluated by th e criterion of ''uniformity of verification rank.'' Assuming a perfect forecast model, equally plausible initial conditions, and the verific ation is a plausible member of the ensemble, these imply the verificat ion when pooled with the 15 ensemble forecasts and sorted is equally l ikely to occur in each of the 16 ranks. Hence, over many independent s amples, a histogram of the rank distribution should be nearly uniform. Using data from the ensemble forecasts, rank distributions were popul ated and found to be nonuniform. This was determined to be largely a r esult of model and initial condition deficiencies and not problems wit h the verification data. The uniformity of rank distributions varied w ith atmospheric baroclinicity for midtropospheric forecast variables b ut not for precipitation forecasts. Examination of the error character istics of individual ensemble members showed that the assumption of id entical errors for each member is not met with this particular ensembl e configuration, primarily because of the use of both bred and nonbred initial conditions in this test, further, there were both differences in the accuracy of eta and regional spectral model bred member foreca sts. The performance of various summary forecasts from the ensemble su ch as its mean showed that the ensemble can generate forecasts that ha ve similar or lower error than forecasts from the 29-km mesoeta, which was approximately equivalent in computational expense. Also, by combi ning the ensemble forecasts with rank information from other cases, re liable ensemble precipitation forecasts could be created, indicating. the potential for useful probabilistic forecasts of quantitative preci pitation from the ensemble.