High-Resolution Rapid Refresh (HRRR) Development for Deterministic and Ensemble Prediction

The 3-km convection-allowing High-Resolution Rapid Refresh (HRRR) is an hourly updating weather forecast model that uses a specially configured version of the Advanced Research WRF (ARW) model and assimilates many novel and most conventional observation types on an hourly basis using Gridpoint Statistical Interpolation (GSI). The HRRR is run hourly out to eighteen forecast hours over a domain covering the entire conterminous United States.  The HRRR transitioned, for the first time, from a real-time experimental research model into the operational NCEP production suite in September 2014 (version one) with a major operational upgrade in mid-2016 (version two). Changes to the real-time experimental HRRR were also deployed in mid-2016 as an initial step towards a third version of the model and included an extension of forecasts to 36 hours once every three cycles.  In this presentation, we will highlight the HRRR data assimilation and model physics changes associated with both the second and third versions that include attribution to specific model bias reductions and associated increases in forecast accuracy, especially as they relate to convective weather.

In addition to the development of the deterministic HRRR, two efforts to extend the hourly updating forecast capability into ensemble prediction are underway.  The first of these efforts leverages the HRRR forecasts in a cost-effective time-lagged ensemble (HRRR-TLE) to estimate hourly-updating likelihood probabilities of various weather hazards associated with severe thunderstorms and heavy precipitation over the CONUS out to 24 hours.  We will highlight the post-processing techniques used to produce these HRRR-TLE products including quantile-mapping and  bias correction and temporal-spatial filtering to produce statistically reliable forecast probabilities.

The second effort involves the development of a more expensive HRRR 3-km 40-member data assimilation and forecast ensemble (HRRRE) for a limited sub-CONUS domain.  Ensemble spread is produced through initial condition perturbations and hourly-cycling of assimilated conventional observations using a GSI-based Ensemble Kalman Filter system.  The hourly data assimilation uses 40 3-km HRRR members with ensemble forecasts of 3-18 members.  We will show comparisons of deterministic HRRR, HRRR-TLE and HRRRE forecasts with both case-studies from real-time forecasts and longer term statistics to highlight both the strengths of each system and avenues for future improvements including consolidation of all three efforts.