Forecast skill and spread in small convection-resolving and large non-convection-resolving WRF model ensembles

Regional scale ensembles composed of members using mixed-physics parameterizations and perturbed initial conditions (ICs) and lateral boundary conditions (LBCs) have been shown to provide valuable information for short-range forecasts (~0-48 hours). Typically, these regional scale ensembles use grid-spacing that is too coarse to resolve convection [e.g., The National Centers for Environmental Prediction's (NCEP) short-range ensemble forecasting (SREF) system], thus, the members use various convective parameterization schemes. Although mixing the convective schemes is an efficient means to generate spread in an ensemble, recent work has shown that using convective schemes can introduce large systematic errors in forecasts of rainfall, and that simulations using convective schemes are not able to resolve fine-scale features in rainfall systems. Simulations using convection-resolving grid-spacing have been shown to have an improved diurnal precipitation cycle as well as the ability to depict fine-scale features that coarser simulations cannot replicate. In addition, convection-resolving simulations should have greater forecast spread due to more rapid growth of errors than coarser grid simulations. Thus, it is believed that an ensemble composed of a limited number of convection-resolving members may exhibit more predictive skill than an ensemble composed of many more members run at relatively coarse grid-spacing.

To test this hypothesis, forecast spread and skill will be analyzed from a convection- resolving WRF model ensemble using mixed physics and dynamic cores with a limited number of members (4-8), along with a WRF model ensemble that also uses mixed physics and dynamic cores, but is run at a coarser grid-spacing, uses convective parameterization schemes, and has many more members (16-20) than the convection-resolving ensemble. To account for initial condition uncertainty and to minimize the effects of reduced spread caused by boundary condition infiltration, each ensemble member from both ensembles will be initialized with a unique bred-perturbation from the Global Forecast System (GFS) Ensemble. Forecast skill will be compared using traditional measures of skill applied to ensembles, such as area under the ROC curve, Brier skill scores, and reliability diagrams. Statistical consistency and forecast spread will be evaluated by comparing the ensemble variance to the mean-square-error of the ensemble mean, and rank histograms will also be used to analyze ensemble spread. In addition, effects of bias correction on rainfall forecasts from both ensembles will be examined.