Evaluating the NCAR Ensemble’s Initialization Approach

Between 7 April 2015 and 30 December 2017, the National Center for Atmospheric Research (NCAR) produced experimental, real-time, 10-member, 48-h, convection-permitting ensemble (CPE) forecasts with 3-km horizontal grid spacing over the conterminous United States. These forecasts were unique in that continuously-cycling ensemble Kalman filter (EnKF) analyses provided flow-dependent initial condition (IC) perturbations.

This study investigates whether the NCAR ensemble’s EnKF-based initialization is preferable to initializing CPEs from more traditional ad hoc methods by examining output from several 10-member CPEs configured identically to the NCAR ensemble except for their IC perturbations. All forecasts were integrated for 48 h over a 31-day period and objectively verified with a focus on probabilistic precipitation.

CPEs initialized by adding either Short Range Ensemble Forecast (SREF) perturbations or random, correlated noise to Global Forecast System (GFS) analyses produced better precipitation forecasts than the EnKF-initialized NCAR ensemble. Similarly, CPE forecasts initialized by re-centering EnKF perturbations about GFS analyses were also better than the NCAR ensemble. Finally, adding random, correlated noise to EnKF mean analyses degraded precipitation forecasts compared to the NCAR ensemble.

Overall, these results suggest flow-dependent IC perturbations and high-quality ensemble mean analyses are critical for good CPE forecasts and that some characteristics of NCAR ensemble mean analyses were suboptimal. Nonetheless, as EnKFs possess nice theoretical properties, these findings should motivate efforts to improve EnKFs for CPE forecast initialization and offer some insights toward realizing these improvements.