Improving the handling of model bias in data assimilation
Laloyaux, Patrick ; Bonavita, Massimo
Errors in numerical weather prediction arise from two main sources: incorrect initial conditions and deficiencies in the dynamics and the physical parametrizations of the forecast model. To correct initial errors, four-dimensional variational data assimilation (4D-Var) adjusts the initial state of the atmosphere to find the model trajectory that best fits the most recent meteorological observations. The new initial state from which forecasts start is called the analysis. In the standard 4D-Var formulation, known as strong-constraint 4D-Var (SC-4DVar), the model is assumed to be perfect and any systematic model errors (biases) which gradually accumulate in the short-range forecasts used in the data assimilation system (the first guess) are not taken into account. A version of 4D-Var which relaxes the assumption that the model is perfect, known as weak-constraint 4D-Var (WC-4DVar), has been run at ECMWF for some years, but without significant positive impact on analysis accuracy and forecast scores (Tre?molet & Fisher, 2010). This has motivated the work reported in this article, which aimed to revise the operational WC-4DVar configuration to better mitigate the effect of model biases during the assimilation step.<br><br>Tests of the new WC-4DVar configuration have shown that analysis and first-guess temperature biases in the stratosphere are reduced by up to 50%. In view of these results, the revised WC-4DVar will be implemented in the next upgrade of the Integrated Forecasting System (IFS Cycle 47r1) planned for later this year. The initial implementation is restricted to the stratosphere, but work is under way to extend WC-4DVar to the troposphere using deep learning methods.</p>
Accès à la notice sur le site du portail documentaire de Météo-France