Sensitivity of ensemble-based variances to initial background perturbations

Abstract

[1] In this study, the sensitivity of ensemble-based background error variances to initial background perturbations is investigated. This is done in a quasi-operational NWP framework, by comparing two experimental ensembles with a reference one. The first experimental ensemble is a cold-start configuration, with initial background perturbations equal to zero. The second experimental ensemble uses initial background perturbations given by random draws of a specified background error covariance matrix B. The reference ensemble is based on a warm-start configuration, launched 6 days before the start of the experimental period. A formal linear analysis of the perturbation evolution suggests that error variance estimates are likely to converge after a few analysis/forecast steps, even if they are expected to be underestimated in the first steps of the cold-start approach. This is confirmed by experimental results, which indicate that the convergence is approximately reached after 3–4 days of cycling. Using random draws from a background error covariance matrix provides variance structures which are more consistent with the reference ensemble than in the cold-start approach. Detailed results are also found to be somewhat sensitive to the amplitude of the initial random draws. Moreover, as expected, the impact of ensemble initiation techniques on forecast scores is nearly neutral once variance estimates have converged toward similar values, while the sensitivity is larger at the beginning of the period.