The assessment of the uncertainty of NWP is currently based upon the so called "ensemble forecasting". This technique implies that the largest forecasting errors are essentially due to the uncertainty in the initial conditions (which are estimated on the basis of scattered observations) while the NWP model is implicitly assumed as perfect. Nonetheless, when data assimilation are envisaged to be used in order to improve model performances and reducing model divergence, it is essential to evaluate the "model uncertainty" (which derives from the discretisation of the equations and from the used parameterisations), because the largest improvements are obtained if the relative uncertainty between model and assimilated data can be specified.
Several techniques have been proposed for NWP model uncertainty assessment, such as the ones based upon model adjoint or based upon Maximum Likelihood and Simplified Kalman Filters (ML/SKF) approach.
Partner 2, in collaboration with Partners 4 and 7, will set up a Kalman Filter Approach based upon the of optimality conditions in terms of independence in time of the innovation process (IIP) and will compare the results with the ML/SKF approach.
Partner 4 will investigate the possibilities for online estimation of forecast error standard deviations (by ML/SKF or KF techniques) to be used within the framework of variational data assimilation.