Simulating model uncertainty of subgrid-scale processes by sampling model errors at convective scales

Ideally, perturbation schemes in ensemble forecasts should be based on the statistical properties of the model errors. Often, however, the statistical properties of these model errors are unknown. In practice, the perturbations are pragmatically modelled and tuned to maximize the skill of the ensemble forecast. In this paper a general methodology is developed to diagnose the model error, linked to a specific physical process, based on a comparison between a target and a reference model. Here, the reference model is a configuration of the ALADIN (Aire Limitee Adaptation Dynamique Developpement International) model with a parameterization of deep convection. This configuration is also run with the deep-convection parameterization scheme switched off, degrading the forecast skill. The model error is then defined as the difference of the energy and mass fluxes between the reference model with scale-aware deep-convection parameterization and the target model without deep-convection parameterization. In the second part of the paper, the diagnosed model-error characteristics are used to stochastically perturb the fluxes of the target model by sampling the model errors from a training period in such a way that the distribution and the vertical and multivariate correlation within a grid column are preserved. By perturbing the fluxes it is guaranteed that the total mass, heat and momentum are conserved. The tests, performed over the period 11-20 April 2009, show that the ensemble system with the stochastic flux perturbations combined with the initial condition perturbations not only outperforms the target ensemble, where deep convection is not parameterized, but for many variables it even performs better than the reference ensemble (with scale-aware deep-convection scheme). The introduction of the stochastic flux perturbations reduces the small-scale erroneous spread while increasing the overall spread, leading to a more skillful ensemble. The impact is largest in the upper troposphere with substantial improvements compared to other state-of-the-art stochastic perturbation schemes. At lower levels the improvements are smaller or neutral, except for temperature where the forecast skill is degraded.