This work package aims at optimizing the ensemble characteristics of the operational regional decadal prediction system towards a realistic range of uncertainty between different ensemble members. An optimal subset of the global MPI-ESM ensemble members is defined for dynamical downscaling with CCLM and the ensemble spread of decadal predictions is calibrated towards external multi-model ensembles applying a combination of dynamical and low-cost statistical-dynamical downscaling. The use of an ensemble-subset enables a more efficient regionalization of MPI-ESM decadal hindcasts in terms of computing time for the operational prediction system.
The work package is structured by several tasks:
1) By means of multivariate statistics and Bayesian approaches an optimal subset of the global MPI-ESM ensemble members representing the statistical properties of the whole ensemble is selected for the dynamical downscaling with CCLM. This approach permits a real-time application to operational predictions (C3-2.1).
2) The ensemble spread of the regional predictions in Central Europe is calibrated towards the adequate spread of the regional multi-model ensembles from the MiKlip I, EURO-CORDEX and ENSEMBLES projects for different variables, seasons and regions applying a combination of regression models, probability matching and Bayes statistics (C3-2.2).
3) Statistical-dynamical downscaling methodologies for the regionalization of the three key variables temperature, precipitation and wind are developed and applied to the full MPI-ESM hindcast ensemble. (C3-2.3 and C3-2.4).
4) A strategy is developed to combine the results from the dynamical and statistical-dynamical downscaling in terms of the optimization of the ensemble characteristics which can be applied to all regional predictions of MiKlip II (C3-2.5).
5) The resulting subset of global ensemble simulations and the optimized ensemble spread of the regional predictions define the optimal configuration and post-processing of the operational regional prediction system and the developed methods are integrated in the central evaluation system for comparison with global predictions (C3-2.6).
The deliverables of the work package are given by the milestones of the different tasks: the optimized subset of global ensemble simulations (M1, month 12), the optimized ensemble spread for different variables, seasons and regions (M2, month 30), the methodologies for the statistical-dynamical downscaling (M3, month 27), homogenous time series for temperature, precipitation, and wind as derived from the statistical-dynamical downscaling (M4, month 33), the ensemble optimization applying the combination of dynamical and statistical-dynamical downscaling (M5, month 33) and the integration of the developed methods in the central evaluation system and the operational prediction system (M6, month 36)
The statistical-dynamical downscaling methodology is based on weather type analysis. In MiKlip-II focus will be given to the Hess and Brezowsky Grosswetterlagen. So far, an objective Grosswetterlagen classification has been applied to ERA40/ERA-Interim and the full available ensemble of the MiKlip decadal prediction system (baseline0, baseline1, prototype, historical runs).
The ensemble spread of CCLM baseline1 (50 runs, 10 ensemble members per decade) and Preop (205 runs, 5 ensemble members per year) have been compared with that of different global and regional multi-model ensembles including CORDEX (13 runs from 10 models), CMIP3 (58 ensemble runs from 24 models) and CMIP5 (88 runs from 38 models). CCLM (Preop) temperature ensembles are under-dispersive, but seem to have the best ensemble spread score among all datasets. Recalibration of temperature ensembles has been tested and the results demonstrate improved model skills.
Paeth, H. | J. Li, F. Pollinger, W. A. Müller, H. Pohlmann, H. Feldmann, H.-J. Panitz