RRFS

From the RUC to the RRFS

Hourly updated forecasting has a long history in the United States. The Rapid Update Cycle (RUC), developed at NOAA's laboratory in Boulder, went into operation at the National Centers for Environmental Prediction (NCEP) in 1994 with a 60 km grid and a new analysis every 3 hours; it moved to hourly updates at 40 km in 1998, to 20 km in 2002 and to 13 km in 2005, when it became NCEP's first model to assimilate cloud observations. On 1 May 2012 the Rapid Refresh (RAP) replaced it over North America, and on 30 September 2014 the High-Resolution Rapid Refresh (HRRR) brought hourly 3 km forecasts over the contiguous United States.

Alongside them, NCEP ran a patchwork of other regional systems built on different codes: the North American Mesoscale model (NAM), a 12 km model with 3 km nests over the contiguous US, Alaska, Hawaii and Puerto Rico, the High-Resolution Window (HiresW) runs, and the SREF and HREF ensembles. In July 2016 NOAA chose FV3, the finite-volume dynamical core of its Geophysical Fluid Dynamics Laboratory in Princeton, for its next global model, and built around it the Unified Forecast System (UFS), one modelling system meant to serve from global to regional scales. The RRFS is the UFS's regional, convection-allowing application, developed by NCEP's Environmental Modeling Center and the Global Systems Laboratory with many partners.

The road was not smooth. In the warm season of 2023 the RRFS produced thunderstorms that were too intense and too wet, and a deep convection scheme was added that August to tame them; then, during NOAA's 2024 Hazardous Weather Testbed, it missed or weakened several severe-weather outbreaks, and in July 2024 the team announced that version 1 would be delayed. A study requested by NOAA's leadership traced part of the problem to the way FV3 arranges its variables on the grid and recommended that version 2 move to MPAS, the dynamical core developed at NCAR. Version 1 was nevertheless finished on FV3.

In May 2026 the National Weather Service announced the implementation of RRFS v1.0 for 31 August 2026; the date then slipped to 6 October, to 14 October and, in the notice updated on 15 October 2026, to 3 November 2026. A real-time parallel feed has been public since 12 August 2026. On the day RRFS goes live, NCEP retires the NAM, the HiresW runs (except over Guam), the HREF and SREF ensembles and the NAM-based statistics. The RAP and HRRR are not part of that retirement: the HRRR even supplies two members to the new RRFS ensemble products.

One 3 km model for a continent

FV3, for Finite-Volume Cubed-Sphere, was designed at NOAA's Geophysical Fluid Dynamics Laboratory: it solves the non-hydrostatic equations of the atmosphere with finite volumes, which keep track of what flows in and out of each cell. In the RRFS it runs as a limited-area model over a single computational domain at 3 km that covers the whole of North America, instead of the separate regional models of the past. At that spacing the model is called convection-allowing: large thunderstorms and squall lines begin to be drawn explicitly, though a single cell or gust front is only a few grid points wide.

The atmosphere is cut into 65 levels, the lowest about 8 m above the surface and the top at 2 hPa, against 50 levels and a 20 hPa top for the HRRR. FV3 stores its winds on the edges of the cells in a so-called D-grid arrangement, which the study mentioned above blamed for cold pools and storms that grow too broad; that is the main reason why the next version is planned on MPAS.

The physics comes largely from the RAP and HRRR lineage. Clouds and precipitation follow the aerosol-aware Thompson-Eidhammer microphysics; the boundary layer, where the wind felt at sea is shaped, is handled by the MYNN eddy-diffusivity mass-flux scheme, which also treats shallow cumulus, above an MYNN surface layer; deep convection by the Grell-Freitas scheme; radiation by RRTMG; land by the RUC land surface model. The Great Lakes are represented by the FVCOM lake model and smaller lakes by the CLM lake model, and smoke and dust from fires are forecast from satellite fire emissions.

How a run is computed

The RRFS renews its analysis every hour, combining a 1-hour forecast with the latest observations: surface stations, mesonets, buoys, coastal stations and ships, radiosondes, aircraft, the NEXRAD weather radars (reflectivity, radial wind and the wind profiles derived from them), lightning seen from the GOES satellites, cloud and wind products from GOES, and the radiances of the polar-orbiting sounders AMSU-A, MHS, ATMS, CrIS, IASI and SSMIS. As of 2025, satellite radiances are used in clear sky only.

The method is a hybrid 3D-EnVar run with NOAA's GSI software. A variational analysis looks for the atmospheric state that best fits both the first guess and the observations; "hybrid" and "En" mean that the expected errors of the first guess come partly from fixed statistics and partly from a 30-member ensemble at 3 km, itself updated every hour by an ensemble Kalman filter (EnKF), so that they reflect the weather of the hour. Radar reflectivity is assimilated directly in this scheme, which lets the model start with the showers and storms that are already on the radar instead of having to grow them.

A regional model cycled hour after hour on itself can drift away from the large-scale flow. Twice a day, at 03 and 15 UTC, the RRFS therefore restarts its atmosphere from NOAA's global analysis and cycles it hourly for several hours until it rejoins the main cycle, while the land state keeps cycling without interruption; information from the global ensemble is also blended into the regional ensemble twice a day. At its edges the domain receives its boundary conditions from the global model GFS.

A forecast starts every hour: it runs to 18 hours, except at 00, 06, 12 and 18 UTC when it runs to 84 hours. At those four times, five ensemble members with different initial states, boundaries and physics also run to 60 hours, and the REFS combines them and the deterministic run with those of the cycle 6 hours earlier, plus two HRRR members over the contiguous US and Alaska, into means, spreads and probabilities. The output is cut into regional grids: 3 km over the contiguous US (Lambert conformal projection) and Alaska, 2.5 km over Hawaii and Puerto Rico (Mercator), and 13 km over North America, with 15-minute fields for some variables.

The files are published in GRIB2 on NOAA's NOMADS server and in part broadcast on NOAAPORT. They are free to use: on Amazon's open data registry, which carried the development prototype, NOAA states that there are no restrictions on the use of RRFS data and asks only for attribution and that no endorsement by NOAA be implied. Until the switch to operations, the parallel feed is not an operational product and can be late or interrupted.

At sea

Its 3 km grid, its hourly cycle and its radar data make the RRFS one of the sharpest views available of coastal North American waters. It can draw sea breezes that build in the afternoon along the Florida and Gulf coasts or on Chesapeake Bay, the wind funnelled through straits and bays, lake breezes and the cold-water effects of the Great Lakes, which it represents with a lake model, and the strong northerlies that hug the California coast in summer. Each new run starts from the storms and fronts as they actually are an hour earlier.

Some of its strengths are also its limits. Convection has been its weak point during development, with storms too intense in 2023 and storms missed in 2024, so a squall drawn by the RRFS, or its absence, deserves some caution; the 10 m wind it gives is a mean wind, and gusts under a storm can be much stronger. Where the Gulf Stream's north wall separates warm and cold water, the wind and the sea can change sharply over a few miles; a 3 km model renders that contrast only as well as the sea temperature it is given.

A regional model also depends on what enters through its edges: a low arriving from the Pacific or the Atlantic is only as good in the RRFS as it was in the GFS that drives it. Beyond a day or two, it is therefore wise to compare the RRFS with independent global models such as ECMWF's IFS or DWD's ICON. For a tropical storm or a hurricane, the RRFS is not a hurricane model: the advisories of the National Hurricane Center come first.

In Saily WR

Saily WR uses the RRFS as three separate models, one per output grid: the contiguous United States, Puerto Rico and Hawaii. Each computes its own route whenever its area covers both the start and the end, sailed in the model's 10 m wind interpolated in space and in time, and shown next to the routes of the other models so that they can be compared.

The app takes the four long runs of the day, at 00, 06, 12 and 18 UTC, hourly from the start of the run to 84 hours. It places the 3 km Lambert grid of the contiguous US onto a regular 0.03° grid, and the 2.5 km Mercator grids of Puerto Rico and Hawaii onto 0.025° grids. Until the RRFS becomes operational, planned for 3 November 2026, the app reads NOAA's parallel feed; a server fetches each new run as soon as it is published, and the Data page shows live which run each of the three is on, how far it reaches and the area it covers.