Evaluation of CAPS Convection-Allowing FV3-LAM Ensembles during the 2022 HWT Spring Forecasting Experiment to Inform the Design of the Rapid Refresh Forecast System (RRFS)

To inform the optimization of the future operational Rapid Refresh Forecast System (RRFS), the Center for Analysis and Prediction of Storms (CAPS) performed three sets of CONUS-domain 3-km finite-volume cubed-sphere limited-area model (FV3-LAM) ensemble forecasts with various configurations during the 2022 NOAA Hazardous Weather Testbed Spring Forecasting Experiment in real time. The first set used different physics parameterizations, the second set included additional initial and lateral boundary condition perturbations, and the third introduced additional stochastic physics perturbations. This study evaluates precipitation, temperature, dewpoint, and wind forecasts and compares them to those of the operational High-Resolution Ensemble Forecast (HREF) and Global Ensemble Forecast System (GEFS). Precipitation forecasts are verified against NCEP stage-IV precipitation analyses, while temperature, dewpoint, and wind forecasts are verified against Unrestricted Mesoscale Analysis (URMA) surface analyses and radiosonde observations. Overall, the ensemble configurations tested are generally suitable for predicting spring season convective rainfall. The CAPS forecasts generally outperform GEFS in terms of ETS, frequency bias, and area under the receiver operating characteristic (ROC) curve, approaching (but not exceeding) the performance of HREF in some metrics. Including stochastic physics perturbations resulted in forecasts objectively very similar to those without such perturbations, except for a small but consistent positive impact on the ensemble spread of surface variables throughout the 84-h forecast period. The CAPS forecasts have a near-neutral to slightly negative bias in total precipitation coverage. Forecasts using the NSSL microphysics scheme have more total rainfall than forecasts using the Thompson scheme and stage-IV analyses, while forecasts using the Noah-MP land surface model generally have lower total precipitation than other forecasts.