Professor Heikki Järvinen
Finnish Meteorological Institute
Firstly, I will present a small result, or a “dilemma” concerning optimal prediction model closure parameters. We have found that in a simplified prediction system (data assimilation + prediction model), the optimal prediction model parameter values depend on the approximations made in the data assimilation component. In our case, the optimized Lorenz’95 model becomes sub-optimal when the data assimilation component is changed from EKF to a more approximate EAKF. This result tends to suggest that model tuning is not only needed when a prediction model undergoes major changes, but also when data assimilation system is upgraded.
Secondly, I will talk about an online algorithm for adaptive model tuning using ensemble prediction system infra-structure. The background for this work is the fact that integration of physical parameterization schemes with NWP model dynamics is increasingly challenging. One reason is that the parameterization schemes of sub-grid scale physical processes contain tunable parameters, and it is very hard to manually specify their optimal values. We developed recently a method called “Ensemble prediction and parameter estimation system” (EPPES) which utilizes ensemble prediction infra-structure for parameter estimation in a very cost-effective way via closure parameter variations. Here, the method is first illustrated by using a modified Lorenz’95 model in ensemble prediction context. The method correctly detects the unknown and wrongly specified parameters values, and leads to an optimal forecast skill. Second, results with an ensemble prediction system emulator, based on the ECHAM5 atmospheric GCM (T42L31) show that the model tuning capability of EPPES scales up to realistic models and ensemble prediction systems. Finally, encouraging results of the EPPES algorithm in the context of the ECMWF forecasting system are presented.
Thursday, 1 November 2012, 3:30 PM
Refreshments 3:15 PM
3450 Mitchell Lane
Bldg 2, Large Auditorium Room 1022