What's wrong with the way ECMWF generates ensemble forecasts?
Thomas HamillNOAA-CIRES Climate Diagnostics Center
Abstract |
Do ECMWF's initial-time TE SVs sample the analysis-error probability distribution? Recent development of ensemble-based data assimilation methods based on the Kalman filter permit us to examine this question more quantitatively. These ensemble methods generate a random set of analyses. With a large ensemble of forecasts from these analyses, we can determine what forecast structures have the largest errors. Under assumptions of linearity of error growth, we can also determine what initial conditions evolve into these structures with the largest errors. We can then determine whether the structure of these "analysis-error covariance singular vectors" (AEC SVs) resembles the structures of ECMWF's total-energy singular vectors. We generate a set of analysis-error covariance singular vectors using a T31 L15 dry general circulation model under perfect-model assumptions. A 400-member ensemble square-root filter data assimilation system is used to assimilate a sparse network of simulated rawinsonde observations every 12 h. 48h forecasts are generated from the the ensemble of analyses, and the forecast and initial-time analysis-error covariance singular vectors are estimated. Results suggest that the structure of initial-time AEC SVs are radically different than TE SVs. AEC SVs have their largest amplitude near the tropopause, TE SVs in the middle-troposphere. AEC SVs grow less rapidly than TE SVs, are larger in scale, are more nearly balanced, and have most of their initial energy in the wind component rather than the temperature component. These results suggest that ECMWF could improve their ensemble forecasts if they changed the manner of generating their initial ensemble. We discuss the implications of these results for ensemble forecasting and for adaptive observation strategies. |
2 PM/ DSRC 1D 403
(Coffee at 1:50 PM)