SEARCH

SEARCH BY CITATION

Keywords:

  • Predictability;
  • Medium-range forecasts;
  • Multi-analysis ensembles;
  • Multi-model ensembles

Abstract

There are numerous occasions on which operational forecasts from different meteorological centres diverge in the medium range. These differing evolutions may be due to the different analyses used by the centres or may be a result of the different models used to produce the forecasts; or a combination of both factors may be involved. Ensemble forecasts in routine production at the European Centre for Medium-Range Weather Forecasts (ECMWF) are designed to address the problem of the impact of small analysis differences on forecast evolution. the initial perturbations, generated using the singular-vector technique, are designed to maximise spread within the ensemble while being consistent with possible analysis errors. the aim of the current work is to examine, in the context of ensemble forecasting, the effects of model and analysis differences on forecast evolution. Two detailed case-studies have been performed using the ECMWF model at T63 resolution and the UK Meteorological Office (UKMO) Unified Model run at a comparable resolution. In each case four 33-member ensembles were produced using all possible combinations of ECMWF and UKMO models and analyses. the same initial perturbations, derived using the singular-vector technique with the ECMWF model, were used for all four ensemble sets.

Significant differences between all four ensemble sets are found in each case-study. the relative impact of model and analysis dependencies varies with geographical region and forecast range; however the greater overall impact is, for both region and range, from the model differences. the varying evolutions of the four ensemble sets are shown to have a substantial impact on synoptic features in the predicted 500 hPa height fields, such that a forecaster would have issued different predictions depending on which ensemble data were available. the synoptic differences are consistent with the statistically significant model and analysis dependencies found using analysis of variance applied to anomaly correlations and to the distribution of ensemble members in phase space. It is demonstrated that the calculation of the initial perturbations is relatively insensitive to differences between the initial analyses of the two centres, but that the dispersion of initial conditions resulting from the 16 singular vectors used to generate the perturbations does not cover the dispersion arising from the analysis fields. In both of the case-studies each of the four ensemble sets contains information not available in any of the other ensembles. These results raise the possibility that, to maximise the information available to the forecaster, both models and both analyses should be used in the generation of ensemble predictions.