Systematic error analysis of heavy-precipitation-event prediction using a 30-year hindcast dataset

Abstract

Abstract. The western Mediterranean region is prone to devastating flash floods induced by heavy-precipitation events (HPEs), which are responsible for considerable human and material losses. Quantitative precipitation forecasts have improved dramatically in recent years to produce realistic accumulated rainfall estimations. Nevertheless, there are still challenging issues which must be resolved to reduce uncertainties in the initial condition assimilation and the modelling of physical processes. In this study, we analyse the HPE forecasting ability of the multi-physics-based ensemble model Prévision d’Ensemble ARPEGE (PEARP) operational at Météo-France. The analysis is based on 30-year (1981–2010) ensemble hindcasts which implement the same 10 physical parameterizations, one per member, run every 4 d. Over the same period a 24 h precipitation dataset is used as the reference for the verification procedure. Furthermore, regional classification is performed in order to investigate the local variation in spatial properties and intensities of rainfall fields, with a particular focus on HPEs. As grid-point verification tends to be perturbed by the double penalty issue, we focus on rainfall spatial pattern verification thanks to the feature-based quality measure of structure, amplitude, and location (SAL) that is performed on the model forecast and reference rainfall fields. The length of the dataset allows us to subsample scores for very intense rainfall at a regional scale and still obtain a significant analysis, demonstrating that such a procedure is consistent to study model behaviour in HPE forecasting. In the case of PEARP, we show that the amplitude and structure of the rainfall patterns are basically driven by the deep-convection parametrization. Between the two main deep-convection schemes used in PEARP, we qualify that the Prognostic Condensates Microphysics and Transport (PCMT) parametrization scheme performs better than the B85 scheme. A further analysis of spatial features of the rainfall objects to which the SAL metric pertains shows the predominance of large objects in the verification measure. It is for the most extreme events that the model has the best representation of the distribution of object-integrated rain.