Extended‐range predictions with ECMWF models: Time‐lagged ensemble forecasting

Abstract A set of 16 ensembles of time‐lagged extended‐range forecasts have been run at different times of year using the T63 version of the European Centre for Medium Range Weather Forecasts (ECMWF) operational model. Each ensemble was composed of 9 integrations from consecutive 6‐hourly analyses. Theoretical properties of ensemble‐mean skill and ensemble spread are studied using a simple model of error growth with a parametrization of the ECMWF model error. the impact of systematic error on the potential improvement in skill of the ensemble‐mean forecast is discussed. the presence of model errors considerably reduces the gain from ensemble averaging. In practice, about a third of the ensemble‐mean forecasts, at forecast days 11‐20, were more skilful than both persistence and climate, and, in addition, were more skilful than the latest member of the ensemble. At days 21‐30. only one of the ensemble‐mean forecasts was similarly skilful. Whilst there is an overall hemispheric‐scale correlation between ensemble spread and skill, a substantial part of this reflects the impact of the annual cycle on both quantities. In the winter period, however, no clear spread/skill correlation was found. Within the winter period, there was considerable case‐to‐case variability in forecast skill. of all the ensembles that of January 1986 was poorest, whereas that of February 1986 was one of the best. the different character of these two ensembles was shown by considering phase‐space trajectories of the ensemble forecasts in the plane spanned by the two principal forecast EOFs of 500 mb height. During the first 15 days, the trajectories of the January ensemble forecasts were consistent with each other, but not with the observed atmospheric trajectory (which was associated with the onset of blocking over Europe). During the last 15 days, as the January ensemble forecasts migrated from positive to negative PNA index, the trajectories dispersed quite strongly, becoming disordered. By contrast, the trajectories of the February forecasts remained mutually consistent and in agreement with the real atmosphere's trajectory throughout most of the forecast period. In order to investigate possible reasons for the failure of the January 1986 ensemble to develop the European block during the first half of the forecast period, two further integrations were made. In the first, the control integration was re‐run with a more recent version of the ECMWF model. Development of the block continued to be missed, though the trajectory of this forecast in phase space in the medium range was quite different from any of the members of the original ensemble. This suggests that a methodology for Monte Carlo forecasting might include perturbations of the model formulation as well as perturbations of the initial conditions. Secondly, motivated by known systematic errors in the model's simulation of tropical divergence, and the diagnostic study of Hoskins and Sardeshmukh (1987), an integration was run in which the model's tropical fields were relaxed towards the verifying analysis. In this integration, substantial ridging over the Euro/Atlantic area occurred, and the extratropical skill scores were noticably improved. Phase‐space trajectories confirm this partial success on a hemispheric scale. However, the intensity of the block was not well captured. It would appear therefore that failure to capture the block is associated partially with the problem of predictability, and partially with the problem of the model's systematic errors, particularly in the tropics. The EOF decomposition of the ensemble forecasts was also used as an objective criterion to test for clustering within the ensemble. According to this analysis, one of the January forecasts (but not the latest) showed quite distinct behaviour, and its extended‐range skill was well above that for the other members of the ensemble. The cluster analysis was performed on all winter forecasts. When the forecasts were categorized into three clusters (which vary from ensemble to ensemble), it was found that the skill at days 11‐20 of one of the categories was always better than the ensemble‐mean forecast. However, this more skilful group of forecasts did not always correspond to the more densely populated ensemble. We suggest that this may be associated with a problem of sampling. Given the results from the present sample of forecasts we believe that prediction beyond the medium range is not currently viable. However, when forecast systematic error, particularly in the tropics, is reduced, and when techniques that can identify the most rapidly growing perturbations have been fully developed, then the cluster analysis suggests that probablistic forecasting of extratropical time‐mean weather may be feasible in the time range of up to two to three weeks into the future.