A6 - Representing the evolution of forecast uncertainty
In Phase 1 a stochastic parameterization scheme for convective-scale numerical weather prediction (NWP) was developed that improves the representation of the diurnal cycle of precipitation and shows a beneficial impact in skill scores. This work will be followed up by examining the role of such schemes in the context of the larger predictability problem. The aim is to understand how predictability of convective precipitation decays with forecast lead time, and how it should be represented in a prediction system at different stages of the evolution. The project will include data assimilation to represent the initial uncertainty, nowcasting for the initial highly-predictable stage, stochastic physics to represent model error, and ensembles of boundary conditions to represent the uncertainty of the synoptic flow. The development of forecast error and spread will be followed over approximately 24 hours, from the first minutes, where individual convective cells are represented, until predictability decays to a background "climatological" distribution determined by the synoptic-scale weather pattern.
Current research is limited by computational constraints that restrict ensemble sizes in most cases to 100 members or less, which are not sufficient to accurately represent the nonlinear evolution of the probability density. The first part of the project will develop and use an idealized, computationally efficient modeling framework based on the shallow water equations modified to include convective instability. Ensemble sizes of up to 10000 members are possible, allowing an accurate representation of probability distributions.
In parallel with this work, the second part of the project will perform analogous experiments using a modern NWP system. This effort will cooperate with the SINFONY project of Deutscher Wetterdienst (DWD), which implements a complete forecasting chain for central Europe, including the stochastic parameterization developed in Phase 1. Idealized and real weather events will be simulated with up to 1000 members. For both the idealized model and the realistic simulations, a range of scale-dependent and statistical measures will be used to quantify the properties of the forecast distribution, looking particularly for non-Gaussian features resulting from nonlinearity and persistent weather systems such as organized convection.