Representing model error in ensemble data assimilation
- 1European Centre for Medium-Range Weather Forecasts, Reading, UK
- 2University of Ljubljana and Center of Excellence SPACE-SI, Ljubljana, Slovenia
Abstract. The paper investigates a method to represent model error in the ensemble data assimilation (EDA) system. The ECMWF operational EDA simulates the effect of both observations and model uncertainties. Observation errors are represented by perturbations with statistics characterized by the observation error covariance matrix whilst the model uncertainties are represented by stochastic perturbations added to the physical tendencies to simulate the effect of random errors in the physical parameterizations (ST-method). In this work an alternative method (XB-method) is proposed to simulate model uncertainties by adding perturbations to the model background field. In this way the error represented is not just restricted to model error in the usual sense but potentially extends to any form of background error. The perturbations have the same correlation as the background error covariance matrix and their magnitude is computed from comparing the high-resolution operational innovation variances with the ensemble variances when the ensemble is obtained by perturbing only the observations (OBS-method). The XB-method has been designed to represent the short range model error relevant for the data assimilation window. Spread diagnostic shows that the XB-method generates a larger spread than the ST-method that is operationally used at ECMWF, in particular in the extratropics. Three-dimensional normal-mode diagnostics indicate that XB-EDA spread projects more than the spread from the other EDAs onto the easterly inertia-gravity modes associated with equatorial Kelvin waves, tropical dynamics and, in general, model error sources.
The background error statistics from the above described EDAs have been employed in the assimilation system. The assimilation system performance showed that the XB-method background error statistics increase the observation influence in the analysis process. The other EDA background error statistics, when inflated by a global factor, generate analyses with 30–50% smaller degree of freedom of signal. XB-EDA background error variances have not been inflated.
The presented EDAs have been used to generate the initial perturbations of the ECMWF ensemble prediction system (EPS) of which the XB-EDA induces the largest EPS spread, also in the medium range, leading to a more reliable ensemble. Compared to ST-EDA, XB-EDA leads to a small improvement of the EPS ignorance skill score at day 3 and 7.