Stochastic treatment of cloud related processes in nonhydrostatic weather prediction models
Zusammenfassung der Projektergebnisse
For numerical weather prediction (NWP) ensemble approaches allow to estimate the forecast uncertainty and the ensemble mean usually is superior to a single deterministic forecast. In classical ensembles the initialization / boundary data are varied to map the uncertainties of the underlying observations. The fact that these ensembles are usually underdispersive is a strong hint that also the intrinsic model uncertainties have to be considered. Stochastic physics is a currently discussed strategy to include these uncertainties due to model parameters / processes. The question is how to choose the stochastic processes the deterministic model parameters are to replaced by. Turbulence is a driving process for small scale fluctuations and has been successfully treated by direct numerical simulation techniques. Here, this has been exploited to construct physically based autoregressive processes for several model parameters. The advantage of such physically based ”stochastic parameters” is that especially their autocorrelation times no longer have to be choosen ad hoc, but can be determined from the turbulent kinetic energy and the eddy dissipation rate both provided by the NWP model itself. This also allows for stochasic parameters to automatically adjust to the current meteorological situation. To implement such a stochastic physics approach (SPA) and to demonstrate its behavior, the stochastic parameter set here has been limited to the most energetic atmospheric processes, i.e. the cloud related processes (initial ice concentration, the collection kernel for cloud particle accretion, subgrid scale cloud cover, convective trigger, convective entrainment/detrainment and the convective closure). Numerical simulations have been performed with the COSMO model of Deutscher Wetterdienst (DWD) including the convection scheme HYMACS which had especially been developed for grid sizes of a few kilometers where convection is partially resolved on the model grid. Drawing different realizations of pseudo-random number sets needed to drive the stochastic processes, ensembles of 24 members have been generated for real cases comprising different meteorological situations. Observational data (hourly precipitation sums) for validation has been provided by the spatially dense SYNOP station network of DWD. Both the ensemble spread and the ensemble skill grow with simulation time, but reach a saturation after several hours because in a spatially limited model domain the deterministic boundary values advected into the model domain can develop stochastically only over a limited time period. The single stochastic parameters show a strong interaction, except for the start of the simulation, since this interaction needs a few hours to build up. Air mass convection, which is locally driven by small scale disturbances, is found to be more sensitve to stochastic physics than frontal precipitation events strongly prescribed by synoptic processes. Both ensemble spread and skill break down for fair weather situations, since stochastic physics here was limited to clud processes. As classical ensembles, the ensembles of the SPA are still underdispersive, because both the classical approach and stochastic physics are to produce only parts of the total spread. An important question is, which spread contribution (produced by a specific stochastic parameter) increases skill (”good spread”) and which one is only unstructured noise with a detrimental effect on the skill (”bad spread”). The SPA system significantly produces ensemble skill classifying the ensemble spread as good spread. The only contribution of bad spread has been found to be generated by the stochastic subgrid scale cloud cover parameterization. The reason for this was lacking sufficiently pronounced spatial patterns in subgrid scale cloud cover, so that this parameterization scheme has been left in its deterministic version. As a conclusion, stochastic physics with underlying stochastic processes based on turbulent dynamics, as introduced in the present project, can improve ensemble forecasts and complete classical ensemble approaches.
Projektbezogene Publikationen (Auswahl)
- 2013: Stochastische Parametrisierung von Wolkenprozessen in nichthydrostatischen NWV-Modellen. DACH-Meteorologentagung, Innsbruck, Austria, September 2-6, 2013
Kuell, V. and A. Bott
- 2014: Stochastic parameterization of cloud processes. Atmos. Res., 143, 176–197
Kuell, V. and A. Bott
(Siehe online unter https://doi.org/10.1016/j.atmosres.2014.01.027)
