A strategy for high-resolution ensemble prediction. I: Definition of representative members and global-model experiments

In the last few years, tens of alternative weather forecasts have been made available to forecasters by operational ensemble prediction systems. In ma ny forecasting applications, it is useful to identify (possibly in an objec tive way) a few representative ensemble members, deemed to represent the mo st interesting weather scenarios. In this paper, a strategy to select repre sentative members (RMs hereafter) from an ensemble prediction is developed, and applied to four cases of medium-range ensemble forecasts performed wit h the European Centre for Medium-Range Weather Forecasts (ECMWF) Ensemble P rediction System (EPS). The four case-studies correspond to events of very intense rainfall (leading to localized floods) in the Alpine region, select ed as benchmarks for numerical simulations in the Mesoscale Alpine Programm e. The RM selection procedure uses a cluster analysis of the ensemble forec asts as its first step. For each cluster, an RM is defined to be the member with the smallest ratio between its average distance from the members of i ts own cluster and its average distance from the members of the other clust ers. Distances are computed either using an L2-norm, applied to 700 hPa geo potential height fields or an L1-norm to precipitation fields. RMs are compared with cluster centroids in the four case-studies of extreme rainfall. By definition, RMs are characterized by a synoptic-scale atmosph eric flow similar to the flow of the corresponding cluster centroid, but th ey contain more small-scale features, especially in the prediction of weath er parameters such as precipitation. RM initial conditions can be used to initiate higher-resolution global fore casts; alternatively, RMs may be used to define initial and boundary condit ions for nested high-resolution forecasts with limited-area models. Integra tions of RMs with the ECMWF global model at T-1 319 horizontal resolution ( compared with the T-1 159 resolution used in the EPS) were performed. Resul ts indicate that each higher-resolution forecast, started from RM initial c onditions, remains closer to the low-resolution RM than to other ensemble m embers, but provides a more detailed forecast of weather parameters, especi ally in regions of complex topography. Experiments with a nested limited-ar ea model, started from the same set of RMs, are described in a companion pa per.