On the initial-state assimilation for limited-area air-quality forecasts

An ensemble of 9 regional Air Quality models is being run operationally within CAMS-50 project providing the 3D fields of air-pollutant distribution over Europe. The models are initialized from their previous-day's forecasts for 00Z and run for 4 days forward. The same models are used for near-real-time reanalysis of the previous day involving the air-quality observations to adjust the modelled  fields via data assimilation methods, such as 3D-var or optimal-interpolation procedures.  In this set-up the observed near-real-time data do not affect the forecasts.  Development of a method to improve the forecast quality by using the assimilated fields from the previous-day analysis is one of the goals for the CAMS-61 project.

As a prototype evaluation for this study, we made several tests with SILAM model (http://silam.fmi.fi) initializing the simulations from the assimilated or non-assimilated states and evaluated the evolution of the model skill scores along the forecast lead time. The tests were made for summer and winter seasons and for initialization time of 00Z vs 12Z.  In order to generalize the results, and make them independent on particular implementation of 3D-VAR in SILAM, the tests were made also with initialization from the analyses made with other CAMS-50 models.  That experiment utilized the list of species and vertical available in the CAMS-50 product catalog. 

The results of the simulation corroborated with our earlier studies that showed a quite quick relaxation of the scores for runs initialized from analyses to the free-run state: with certain variability between the species, the runs converged to the free-run trajectory generally within several hours.  We also investigated the issues connected with initialization from the incomplete set of species and sparse vertical, which might make the scores of the forecast initialized from the incomplete assimilated model state being worse than the ones from the free-run model.