Ensemble Forecasting in a System with Model Error

Abstract

Error in weather forecasting is due to inaccuracy both in the models used and in the estimate of the current atmospheric state at which the model is initiated. Because weather models are thought to be chaotic and, therefore, sensitive to initial conditions, the technique of ensemble forecasting has been developed in part to address the latter effect. An ensemble of forecasts is made with perturbed initial conditions, with the aim being to produce an estimate of the probability distribution function for the future state of the weather. Some ensemble schemes also include changes to the model. While the ensemble approach is quite widely adopted, however, its verification is complicated, and the effect of model error on ensemble performance is not clear. In this paper, the effect of model error on ensemble behavior for a version of the Lorenz ?96 system is investigated. It is shown that estimates of the model?s ability to shadow the observations, obtained using the model drift, are robust to observational error and smoothing schemes, such as four-dimensional variational data assimilation (4DVAR), and help reveal the effect of model error on ensemble performance. Comparisons are made with full weather models. The aim is to provide a study of ensemble error in the context of the Lorenz ?96 system, which may be useful in formulating questions and experiments for weather models.