Reforecasting two heavy-precipitation events with three convection-permitting ensembles

Abstract

We investigate the potential added value of running three limited-area ensemble systems (with the WRF, Meso-NH and MOLOCH models and a grid spacing of approximately 2.5 km) for two heavy-precipitation events in Italy. Such high-resolution ensembles include an explicit treatment of convective processes and dynamically downscale the ECMWF global data, which have a grid spacing of approximately 18 km. The predictions are verified against rain-gauge data and their accuracy is evaluated over that of the driving coarser-resolution ensemble system. Furthermore, we compare the simulation speed (defined as the ratio of simulation length to wall-clock time) of the three limited-area models to estimate the computational effort for operational convection-permitting ensemble forecasting. We also study how the simulation wall-clock time scales with increasing numbers of computing elements (from 36 to 1152 cores). Objective verification methods generally show that convection-permitting forecasts outperform global forecasts for both events, although precipitation peaks remain largely underestimated for one of the two events. Comparing simulation speeds, the MOLOCH model is the fastest and the Meso-NH is the slowest one. The WRF model attains efficient scalability, whereas it is limited for the Meso-NH and MOLOCH models when using more than 288 cores. We finally demonstrate how the model simulation speed has the largest impact on a joint evaluation with the model performance because the accuracy of the three limited-area ensembles, amplifying the forecasting capability of the global predictions, does not differ substantially.Significance statement. We reforecasted two heavyprecipitation events that struck Italy in autumn 2011 by using recent versions of three limited-area weather models, which use a more accurate description of convective processes than the driving global model. By reforecasting past events, the aim of this work is to assess the information content carried by current numerical forecasting systems in the event of similar high-impact events happening again. In view of the potential use for operational forecasting, this study evaluates how the time needed to deliver the forecasts decreases as the computer speed increases. It is shown that such a scaling capability depends not only on the computing elements but also on the geometry of the model domain.