Convection-Allowing Versus Convection-Parameterizing Ensembles for Medium-Range Precipitation Forecasts

Benefits of convection-allowing ensembles (CAEs) relative to coarser convection-parameterizing ensembles are well-established for short-term precipitation forecasts, but it is less clear whether these benefits extend to medium-range (3–8-day) forecast periods. To further understand these potential benefits, 3–8-day precipitation forecasts from two CAEs with ~3-km horizontal cell spacing over the conterminous United States (CONUS) were examined and compared to precipitation forecasts from coarser ensembles. One CAE had 10 members and used GFDL’s “C-SHiELD”, a prediction model based on the Finite-Volume Cubed-Sphere Dynamical Core (FV3), while the second CAE had 5 members and used NSF NCAR’s Model for Prediction Across Scales (MPAS). The two CAEs were initialized daily at 0000 UTC across both May 2023 and 2024 and compared to precipitation forecasts from a 15-km MPAS-based ensemble and NCEP’s ~25-km Global Ensemble Forecast System (GEFS).

Both ~3-km CAEs more accurately represented the observed diurnal cycle of precipitation than the GEFS and 15-km MPAS ensemble throughout medium-range periods. However, regarding forecast skill as objectively measured by percentile-based fractions skill scores that assessed whether it rained at the right place and time, results differed across the years. Specifically, for May 2023, which featured weak synoptic-scale forcing over the CONUS, precipitation forecasts from both CAEs were more skillful than corresponding forecasts from the GEFS and 15-km MPAS ensemble through at least 5 days. Conversely, for May 2024, which featured strong synoptic-scale forcing over the CONUS, the CAEs did not produce more skillful precipitation forecasts than the GEFS and 15-km MPAS ensemble after ~3–4 days. These findings suggest that benefits of CAEs relative to coarser ensembles for medium-range precipitation forecast skill may depend on the meteorological regime.