Initial perturbations based on the ensemble transform (ET) technique in the NCEP global operational forecast system

Since modern data assimilation (DA) involves the repetitive use of dynamical forecasts, errors in analyses share characteristics of those in short-range forecasts. Initial conditions for an ensemble prediction/forecast system (EPS or EFS) are expected to sample uncertainty in the analysis field. Ensemble forecasts with such initial conditions can therefore (a) be fed back to DA to reduce analysis uncertainty, as well as (b) sample forecast uncertainty related to initial conditions. Optimum performance of both DA and EFS requires a careful choice of initial ensemble perturbations. DA can be improved with an EFS that represents the dynamically conditioned part of forecast error covariance as accurately as possible, while an EFS can be improved by initial perturbations reflecting analysis error variance. Initial perturbation generation schemes that dynamically cycle ensemble perturbations reminiscent to how forecast errors are cycled in DA schemes may offer consistency between DA and EFS, and good performance for both. In this paper, we introduce an EFS based on the initial perturbations that are generated by the Ensemble Transform (ET) and ET with rescaling (ETR) methods to achieve this goal. Both ET and ETR are generalizations of the breeding method (BM). The results from ensemble systems based on BM, ET, ETR and the Ensemble Transform Kalman Filter (ETKF) method are experimentally compared in the context of ensemble forecast performance. Initial perturbations are centred around a 3D-VAR analysis, with a variance equal to that of estimated analysis errors. Of the four methods, the ETR method performed best in most probabilistic scores and in terms of the forecast error explained by the perturbations. All methods display very high time consistency between the analysis and forecast perturbations. It is expected that DA performance can be improved by the use of forecast error covariance from a dynamically cycled ensemble either with a variational DA approach (coupled with an ETR generation scheme), or with an ETKF-type DA scheme.