The HRRR model: what it is and how to read it

The HRRR model (High-Resolution Rapid Refresh) is NOAA’s hourly updating, convection-allowing weather model, run at 3 km grid spacing over the contiguous United States and Alaska. It was developed at NOAA’s Global Systems Laboratory and runs operationally at the National Centers for Environmental Prediction (NCEP), where the first version went live on 30 September 2014.

The outlooks and forecast models guide makes the central point: simulated reflectivity describes the character of a day, not the position of a storm at a given minute. This page goes underneath that — how the model is built, where radar enters it, what its own developers measured about its errors, and how to read several runs together.

The model at a glance

The specifications below come from the two-part system description published by the HRRR’s developers (Dowell et al. and James et al., 2022, Weather and Forecasting) and from NCEP’s Environmental Modeling Center.

Earlier versions went operational on 23 August 2016 (v2) and 12 July 2018 (v3). The 48-hour runs every six hours arrived with v4; v3 had extended those runs to 36 hours. Dowell et al. describe v4 as the final WRF-based version of the system.

Convection-allowing versus parameterized

A thunderstorm updraft is a few kilometers wide. A model with a 13 km grid, such as the RAP, cannot draw one, so it uses a convective parameterization: a scheme that estimates the net effect of storms inside each grid box — rainfall, heating, moisture transport — without simulating the storms themselves.

At 3 km the HRRR switches that scheme off. Storms have to emerge from the model’s own equations on the grid. The term for that is convection-allowing: the grid is fine enough to permit explicit storms, not fine enough to resolve their interiors. The storms it produces carry recognizable structure: lines, bows, discrete cells, stratiform shields. By simple arithmetic, a mesocyclone 2 to 10 km across spans only one to three 3 km grid boxes, and nothing at tornado scale exists in the model at all.

How radar gets into the model

Every hourly HRRR run begins one hour early. The model integrates a “pre-forecast” from one hour before the run time up to the run time, and during that hour it is nudged with latent heating derived from radar reflectivity: the three-dimensional Multi-Radar Multi-Sensor (MRMS) mosaic, augmented by lightning data, at four times spaced 15 minutes apart. Where radar sees precipitation, the model is heated as if condensation were happening there, which spins up updrafts where storms actually exist. After the pre-forecast, conventional observations — surface stations, balloons, aircraft — are assimilated, and the free forecast begins.

HRRRv4 added a 36-member, 3 km ensemble called the HRRRDAS, cycled every hour, which also assimilates MRMS reflectivity and supplies the starting state and error statistics for the main run. James et al. report that a pre-operational HRRR without direct radar assimilation showed a sharp spurious jump in reflectivity coverage in the first few hours, and the radar pre-forecast introduced with the first operational version reduced it.

James et al. also show the benefit of assimilation in the relatively high skill of the first few hours of each run. The further a forecast hour sits from that radar-anchored start, the more its storms are the model’s own.

Simulated reflectivity is a scenario

The HRRR’s best-known output is simulated composite reflectivity, rendered in radar colors (the radar reading guide covers what those colors mean in real data). It looks like a radar image from the future. It is a physically consistent picture of one way the afternoon could unfold.

The HRRR’s own verification reflects this. James et al. scored reflectivity forecasts after averaging both forecast and MRMS observations onto a 20 km grid, precisely so that a correctly forecast storm a few grid points off would not be scored as both a miss and a false alarm. If the developers do not grade the model at 3 km placement, a reader should not trust it at that precision either.

A documented bias

James et al. (2022) reported that across the eastern United States in the 2020 warm season (15 April to 15 October), HRRRv4 predicted too much convection at the 35 dBZ composite reflectivity threshold. The error had a daily rhythm: the over-forecast bias bottomed out near 1900 UTC and peaked near 0200 UTC, which the authors read as the model growing storms upscale too aggressively after initiation. Forecasts valid between 1500 and 0300 UTC also showed a dip in skill tied to the difficulty of predicting convective initiation.

In plain terms: in that verification, evening HRRR output tended to show more and larger organized convection than MRMS observed. That is one season, one region and one threshold, and it should be read as a tendency to check rather than a correction to apply.

Reading successive runs

Because a new HRRR arrives every hour, several runs are always valid for the same future time: the 18 UTC run’s 6-hour forecast, the 19 UTC run’s 5-hour forecast, and so on. Stacking them is called a time-lagged ensemble. GSL built an experimental version, the HRRR-TLE, that post-processes recent runs into hourly updating probabilities of hazards.

Two cautions from its own evaluation in the 2018 Hazardous Weather Testbed Spring Forecasting Experiment: time-lagged members share the same model and much of the same starting information, so their spread “may be limited”, “often resulting in an overconfident solution”. The same study found it still gave useful probabilistic severe guidance on many days, and on some outperformed formal convection-allowing ensembles.

What that means when you flip through runs yourself:

• Agreement is weak evidence, not proof. Five runs showing the same line can be one model repeating one error.
• Trends carry information. A line that initiates earlier in each new run, or shifts steadily east, is telling you something the single latest run does not.
• Newer runs have fresher radar. For the next few hours, the latest cycle has seen the current storms; older cycles have not.
• Check against observation. If the run from two hours ago has storms that radar does not show now, discount the rest of that run.

What comes after the HRRR

NOAA’s successor system is the Rapid Refresh Forecast System (RRFS), a 3 km, hourly updating model built on the Unified Forecast System. As of an NWS Service Change Notice dated 9 September 2026, NCEP plans to implement RRFS and its ensemble, REFS, on 14 October 2026, subject to delay if that day is declared a Critical Weather Day. That implementation retires the NAM, SREF, HREF and HiresW. The notices do not list the HRRR or RAP among the retired systems. Dowell et al. had projected HRRR replacement for 2024; that date passed, so treat any successor timeline as provisional until NCEP announces it.

WeatherStory Radar extends its radar loop with HRRR forecast reflectivity hourly to +12 hours, with a bright seam marking where observation ends and the model begins.