| The primary purpose of ensemble (or, more generally, probabilistic) prediction is to define an a priori estimate of the uncertainty on the future state of the flow. One conceptually simple way to describe that uncertainty is by the conditional probability distribution for the state of the flow, given all the available relevant information. In that perspective, ensemble forecasting must aim at defining a sample of the conditional probability distribution. The latter depends on both the initial uncertainty at the beginning of the forecast, and on the uncertainty in the dynamics that governs the evolution of the flow, as described by the numerical prediction model (or models, in the case of a multi-model system). A first series of questions has precisely to do with the identification and quantification of the uncertainties in the initial state and the forecasting model. The uncertainty on the initial conditions is determined by a complex combination of three causes: the recent observations, which have decreased the uncertainty in the observed components, the dynamics of the flow, which has increased the uncertainty in the unstable components of the flow, and decreased it in the stable components, and the model error, which has globally increased the uncertainty. All three effects are represented, at least to some degree of accuracy, in present assimilation algorithms, and it is argued that the best way to define the initial conditions of ensemble prediction is through an appropriate ensemble assimilation system. Another series of questions has to do with the cost efficiency of ensemble prediction. For instance, given the choice, is it better to increase the quality of the prediction model (e. g., through an increase of spatial resolution), or to increase the size of the ensembles? It is argued that no gain can in practice be obtained by increasing ensemble size beyond a few tens of units, even in the case of a multi-model system. This is fundamentally due to the unavoidably limited size of the verifying sample. The limitations of probabilistic prediction are discussed. Now, some results suggest that ensemble prediction can have other uses than estimation of uncertainty. In the case of prediction of precipitation, there is often a high correlation between the estimated probability of precipitation and the observed amount of precipitation. Ensemble prediction can therefore be used as a substitute for deterministic prediction. This aspect is discussed. |
|