Evaluation of ‘GLAMEPS’—a proposed multimodel EPS for short range forecasting


Corresponding author. Present address: ECMWF, Shinfield Park, RG2 9AX, United Kingdom. e-mail: trond.iversen@met.no


Grand Limited Area Model Ensemble Prediction System (GLAMEPS) is prepared for pan-European, short-range probabilistic numerical weather prediction of fine synoptic-scale, quasi-hydrostatic atmospheric flows. Four equally sized ensembles are combined: EuroTEPS, a version of the global ECMWF EPS with European target; AladEPS, a downscaling of EuroTEPS using the ALADIN model; HirEPS_K and HirEPS_S, two ensembles using the HIRLAM model nested into EuroTEPS including 3DVar data-assimilation for two control forecasts. A 52-member GLAMEPS thus samples forecast uncertainty by three analysed initial states combined with 12 singular vector-based perturbations, four different models and the stochastic physics tendencies in EuroTEPS. Over a 7-week test period in winter 2008, GLAMEPS produced better results than ECMWF's EPS with 51 ensemble members. Apart from spatial resolution, the improvement is due to the multimodel combination and to a smaller extent the dedicated EuroTEPS. Ensemble resolution and reliability are both improved. Combining uncalibrated ensembles is seen to produce a better combined ensemble than the best single-model ensemble of the same size, except when one of the single-model ensembles is considerably better than the others. Bayesian Model Averaging improves reliability, but needs further elaboration to account for geographical variations. These conclusions need to be confirmed by long-period evaluations.