Estimation and Correction of the GFS systematic errors

Systematic model errors contribute significantly to the total forecast error growth in weather prediction models like Global Forecast System (GFS). We estimate the GFS systematic errors at the analysis time so they can be corrected empirically, within the model, and at the same time provide guidance for the development of improved parameterizations of subgrid-scale physical processes. We follow the approach of Danforth and Kalnay (2008), who showed that the online model bias correction not only corrected the systematic components of the errors as well as the statistical corrections performed a posteriori on the forecasts, but also significantly reduced the random errors.

Here we estimate the 6hr model error bias as the time average of the 6hr analysis increment (AI), which represents the best estimate of the model bias before the systematic errors grow non-linearly. The fields of mean AI for the GFS indicate that the estimated model bias is robust despite operational changes in the model and data assimilation schemes, and that it takes place on larger scales, well resolved by T62. The analysis of errors in diurnal cycle shows that four leading AI eigenmodes represent the diurnal cycle errors exceedingly well, implying the possibility of correcting the GFS diurnal errors with just a few terms in the time derivative of each variable.

We plan to test the impact of correcting the model online using these estimations of model bias and diurnal cycle errors, and verify whether the encouraging reduction of systematic and random errors obtained by Danforth and Kalnay are still present in a much more realistic NWP system.