Predictions of convective cells from nowcasting, numerical forecasting and a combination of both

Summer thunderstorms can cause strong socio-economic impacts over Germany. The project SINFONY (Seamless integrated forecasting system) at German Weather Service has the goal to improve short-range predictions of these storms. Nowcasting (NWC) currently is superior to numerical weather prediction (NWP) on the very-short range up to about two hours in predicting convective cells while NWP performs better afterwards. Within SINFONY products are developed that integrate both approaches for a seamless prediction. High-resolution reflectivities from the German radar network are used as observational data base and for NWC initialization. The respective reflectivities from NWP models are derived by employing the radar forward operator EMVORADO. Convective cells are identified from these reflectivities using the KONRAD3D cell detection tool.

We present forecasts of convective cells from standalone NWC and NWP predictions and from a product that combines both systems. The recently developed NWC ensemble system KONRAD3D-EPS comprises 20 members with stochastic differences in the positions and life cycles of NWC objects. NWP objects come from ICON-D2-EPS (20+1 members) simulations employing a two-moment microphysics scheme with a forecast horizon of 8 hours. To combine these 41 members the NWP objects are clustered spatially, compared with each observation and the cluster closest to an observation is selected. Several properties of objects within a selected cluster are compared with the matching observed object using the Total Interest. Model objects that are similar enough to the observed object are selected and spatially shifted to make their centroid position equal to the observed cell. The trajectories of shifted NWP objects are then used, together with the NWC objects, as forecasts of convective cells. NWP objects that develop later in the forecast are considered as well but without the assignment to an observed object. Thus, with increasing lead time and the decease of cells that existed during the initialization of the combination, the product smoothly transitions into a purely model-based forecast.

Having an ensemble of predicted objects necessitates some kind of information reduction. Here the pseudomember method is employed that selects the locally most representative objects from the ensemble. For the object-based verification we use the Median of Maximum Interest. It reveals for JJA 2022 that the methods described above deliver a seamless object-based forecast product for convective cells which unifies the strengths of NWC and NWP. For lead times between 30 and 120 minutes the combined product performs even better than each prediction type standalone.