Ensembles using multiple models and analyses

The performance of the European Centre for Medium-Range Weather Forecasts (ECMWF) Ensemble Prediction System (EC EPS) is compared with that of a number of alternative configurations which incorporate information from additional analyses or an additional model or both. Each configuration is approximately equivalent in size, resolution and computational cost and could therefore in principle provide an alternative operational EPS. A multi-centre ensemble system (MC EPS) is constructed by replacing half of the EC EPS with integrations of the Met Office (UKMO) model perturbed about the UKMO analysis. Two additional configurations are constructed using the ECMWF model only, but adding information available from four other operational analyses. A 'consensus' ensemble (CONS EPS) is generated by adding the operational EC EPS perturbations to the mean of the available analyses (the consensus analysis); a multi-analysis ensemble (MA EPS) is generated by combining smaller ensembles of ECMWF-model integrations perturbed about each of the available operational analyses. The different systems are compared over a large sample of 60 cases covering the period October 1998 to July 1999. All the alternative configurations are found to improve over the EC EPS in terms of ensemble-mean skill, ensemble spread, and probabilistic evaluation. For 500 hPa height, the MC EPS is consistently the most skilful configuration, but for 850 hPa temperature the MC and MA systems are equivalent. The CONS EPS is generally less skilful than the other two alternatives. The benefit of the MC EPS over the MA EPS for 500 hPa height can be attributed to the combination of different models within the MC EPS. However, it is shown that application of a simple bias correction to the ensemble members removes the advantage of the MC EPS. It is also noted that, by construction, the MC and MA systems have larger initial spread than the EC and CONS configurations. It is demonstrated that increasing the amplitude of the initial perturbations of the CONS EPS improves the performance of this configuration to be comparable with the MC and MA systems. It is concluded that the benefit of the MC EPS can be realized using the single-model EC EPS incorporating information from additional operational analyses (provided that the issue of model bias is addressed), thus avoiding the technical difficulties of maintaining more than one model in an operational EPS.