REFS stands for RRFS Ensemble Forecast System. It is not a stand-alone weather model in the way GFS (Global Forecast System), HRRR (High-Resolution Rapid Refresh), or RRFS (Rapid Refresh Forecast System) are. Instead, REFS is the built-in ensemble component of RRFS itself, a way of running RRFS many times with small variations and combining the results into a probability-based forecast rather than a single “this is what will happen” prediction (NOAA Global Systems Laboratory).
Most weather models most people are familiar with, such as GFS or HRRR, are “deterministic.” Each model run starts from one best estimate of the current state of the atmosphere and produces one forecast: a single, specific outcome. But the atmosphere is chaotic, and small uncertainties in that starting estimate, or in how the model handles processes it can’t fully resolve, can grow into large differences in the forecast a few hours or days out.
An ensemble forecast addresses this by running a model many times with slightly different starting conditions or slightly different internal settings, then looking at the whole spread of results together, rather than trusting any single run. The National Weather Service’s own glossary defines an ensemble as “a collection of numerical model results that show slightly different possible outcomes,” produced from “multiple predictions from an ensemble of slightly different initial conditions and/or various versions of models” (National Weather Service Glossary). If most of the individual runs agree on, say, an inch of rain in a given area, forecasters can have more confidence in that outcome. If the runs disagree widely, that disagreement itself is useful information: it tells forecasters the situation is genuinely uncertain, rather than giving them false confidence in one number.
REFS applies this idea to RRFS. According to NOAA’s Global Systems Laboratory (GSL), the office responsible for developing RRFS, REFS is built from a control member (a single baseline run) plus six additional members, one of which is HRRR itself, along with multiple time-lagged members (versions of the model started at an earlier time and allowed to “catch up,” which adds useful spread to the ensemble without the computing cost of running an entirely separate extra model). GSL describes the ensemble’s spread as coming from a combination of methods: “initial condition perturbations, stochastic physics, physics diversity” and time-lagging (NOAA Global Systems Laboratory). In plain terms, some members start from slightly different snapshots of the atmosphere, some use slightly different internal physics settings, and some are simply the same forecast run at a different time, all so that the collection of forecasts captures a realistic range of “what might actually happen.”
Separately, in NOAA’s official notice announcing RRFS and REFS’s operational rollout, the mechanism is described somewhat more precisely: REFS is described as a “time-lagged ensemble product generation system” that combines forecast output from the current and six-hour-old cycles of the RRFS deterministic and ensemble runs, with HRRR contributing two further members (also drawn from its current and six-hour-old runs) for the continental United States and Alaska (National Weather Service, 2026b). The exact bookkeeping of “how many members” differs slightly between these two NOAA descriptions, but the underlying idea is the same either way: REFS is a set of RRFS-based and HRRR-based forecasts, some run from slightly different starting points and some simply time-shifted copies, blended together into a single probability product.
REFS is built directly on top of RRFS and runs at the same roughly 3-kilometer grid spacing as RRFS’s primary forecast domain (NOAA Global Systems Laboratory). Where RRFS itself is used to produce a single, deterministic forecast, REFS takes that same underlying model system and reruns it multiple times (plus folds in HRRR) to produce probability information instead. Because every REFS member, RRFS’s own reruns as well as the two HRRR-based members, runs at this same fine grid spacing, REFS carries forward RRFS’s ability to resolve terrain features like ridges, valleys, and coastlines directly rather than smoothing them out, the same design rationale NOAA has documented for HRRR (Dowell et al., 2022). What the ensemble adds on top of that shared terrain detail is a sense of uncertainty in exactly how terrain shapes a given forecast, for example how confident to be about where a thunderstorm first forms along a ridge or how a valley channels wind, rather than trusting a single run’s placement of that effect. REFS’s members, though, don’t all handle terrain in quite the same way under the hood. RRFS’s members use FV3’s own terrain-following vertical coordinate, while the two HRRR-based members use a separately-built terrain-following coordinate developed for WRF, the modeling system HRRR runs on. Both approaches track the real ground surface in their lower layers rather than using flat, unchanging altitude levels, but they were built independently rather than sharing code (NOAA Geophysical Fluid Dynamics Laboratory).
HRRR (High-Resolution Rapid Refresh) plays a notable double role here. It is a separate, well-established NOAA model in its own right: an hourly-updated, roughly 3-kilometer-resolution model that assimilates fresh radar data every 15 minutes and, like RRFS’s deterministic run, produces a single, specific forecast rather than a range of probabilities (NOAA Rapid Refresh Program). What makes REFS distinctive is that HRRR is not just a separate model running alongside RRFS. It is literally nested inside REFS as one of REFS’s own ensemble members. NOAA built REFS this way on purpose. HRRR is already a proven, independently run model with its own separate dynamical core, so folding its output in as one of REFS’s members adds a genuinely different perspective to the mix rather than just another variation of RRFS. NOAA’s own materials describe HRRR output as directly contributing member forecasts to the REFS ensemble (NOAA Global Systems Laboratory; National Weather Service, 2026b).
Because REFS is built from multiple forecast runs rather than one, its output takes the form of probabilities rather than a single predicted outcome. GSL describes this as giving forecasters “a new level of confidence” by allowing them to see, for example, the likelihood of a storm-scale event occurring at a given location, rather than a single yes-or-no forecast (NOAA Global Systems Laboratory). In practice, this kind of ensemble product is typically presented as things like the percentage of ensemble members that forecast measurable rainfall in an area, or the probability that a rainfall or wind threshold will be exceeded at a given location, sometimes called a “probability of exceedance.” That kind of output is meant to be read alongside deterministic forecasts (from RRFS’s own control run, HRRR, or GFS) as a measure of confidence and spread, not as a replacement for them.
REFS’s forecasts extend out to 60 hours for the four main forecast cycles per day (00, 06, 12, and 18 UTC), matching the timing of RRFS’s own extended-range deterministic runs, and are intended to support forecasting of aviation hazards, severe convective weather, heavy precipitation, winter weather, and similar short-range, storm-scale forecasting needs (National Weather Service, 2026b). REFS is intended to succeed NOAA’s older High Resolution Ensemble Forecast system (HREF), providing wider geographic coverage and a longer forecast window than HREF did (National Weather Service, 2026b).
Sources
National Oceanic and Atmospheric Administration, Global Systems Laboratory. “Rapid Refresh Forecast System (RRFS).” https://gsl.noaa.gov/rrfs/.
National Weather Service. “Glossary: Ensemble.” https://forecast.weather.gov/glossary.php?word=ENSEMBLE.
National Weather Service. Service Change Notice 26-48 Updated, “RRFS and REFS Implementation.” https://www.weather.gov/media/notification/pdf_2026/scn26-048_Updated_RRFS_and_REFS_Implementation_aad.pdf. 2026.
Dowell, D.C., et al. “The High-Resolution Rapid Refresh (HRRR): An Hourly Updating Convection-Allowing Forecast Model. Part I: Motivation and System Description.” Weather and Forecasting, 37(8), 1371-1395, 2022.
National Oceanic and Atmospheric Administration, Geophysical Fluid Dynamics Laboratory. “FV3: Key Components.” https://www.gfdl.noaa.gov/fv3/fv3-key-components/.
National Oceanic and Atmospheric Administration. “Rapid Refresh / High-Resolution Rapid Refresh (HRRR).” https://rapidrefresh.noaa.gov/hrrr/.