Towards process‐level representation of model uncertainties: stochastically perturbed parametrizations in the ECMWF ensemble. (4th December 2016)

• Record Type: Journal Article
• Authors: Ollinaho, Pirkka; Lock, Sarah‐Jane; Leutbecher, Martin; Bechtold, Peter; Beljaars, Anton; Bozzo, Alessio; Forbes, Richard M.; Haiden, Thomas; Hogan, Robin J.; Sandu, Irina
• Is Part Of: Quarterly journal of the Royal Meteorological Society. Volume 143:Number 702(2017)
• Page Start: 408
• Page End: 422
• Publication Date: 2016-12-04
• Subjects: numerical weather prediction -- ensemble prediction system -- model uncertainty representation -- model error representation -- stochastic physics -- stochastic parametrizations -- closure parameters -- parameter perturbations; Meteorology -- Periodicals; 551.5
• DOI: 10.1002/qj.2931
• Languages: English
• ISSNs: 0035-9009

Abstract

Abstract : Ensemble forecasts depend on representations of model uncertainties. Here, we introduce a model uncertainty representation where a novel approach is taken to the established methodology of perturbing model parameters. The Stochastically Perturbed Parametrizations (SPP) scheme applies spatially and temporally varying perturbations to 20 parameters and variables in the ECMWF IFS model. The perturbed quantities are chosen from the IFS parametrizations of (a) turbulent diffusion and subgrid orography, (b) convection, (c) clouds and large‐scale precipitation, and (d) radiation. The perturbations are drawn from prescribed distributions. Numerous configurations of SPP are compared in experiments with the ECMWF ensemble forecasts at T L 399 resolution up to 15 day lead times. Halving the standard deviations of the perturbations considerably reduces the ensemble spread. Smaller variations of the standard deviations lead to minor changes to the ensemble spread. Experiments with different space and time correlations for the perturbations suggest optimal correlation scales of 2000 km and 72 h. SPP displays a lower skill for upper‐air variables in the medium range than the current operational model uncertainty scheme Stochastically Perturbed Parametrization Tendencies (SPPT) for a given set of fixed initial‐state perturbations. However, in short ranges the two schemes display similar skill. Moreover, verification against surface observations shows SPP is more skilful than SPPT in 2 m temperature for the first couple of forecast days. We show that the direct perturbation of cloud (and radiation) processes in SPP has a greater impact on radiative fluxes than the indirect perturbation via SPPT. SPP also produces a better model climate for a range of variables when comparing long model integrations with the two schemes, indicating the potential advantage of a physically consistent model uncertainty representation. A comparison of the tendency perturbations introduced by SPP and SPPT suggests that the two schemes represent different aspects of model uncertainty.