The HRRR, NOAA's Three-Kilometer Rapid-Refresh Model That Redraws the Thunderstorm Forecast Every Single Hour

The HRRR - NOAA’s High-Resolution Rapid Refresh - is a numerical weather model that simulates thunderstorms on a 3-kilometer grid across the continental United States and produces a fresh forecast every hour, 24 times a day. Unlike older, coarser models that only estimate where storms are likely, the HRRR is fine enough to grow individual storm cells inside the simulation and show you where they’ll form and which way they’ll track. For pilots deciding whether to launch ahead of building convection, it’s effectively a simulated radar picture of an afternoon that hasn’t happened yet.

What Is a Weather Model, and Why Does Resolution Matter?

A numerical weather model is a large physics simulation of the atmosphere. It takes the current state of the air - temperature, pressure, moisture, and wind at millions of points - feeds it into the equations that govern how fluids and heat move, and steps that simulation forward in time. Run it far enough and you get a picture of the sky six hours, twelve hours, or two days from now.

Every forecast you’ve ever seen traces back to a model like this. Two properties separate the merely useful ones from the exceptional ones: resolution and refresh rate.

Resolution is the size of the grid. Picture a checkerboard laid over the country, with the weather calculated in each square. If those squares are 13 kilometers across - roughly 7 nautical miles - then anything smaller than one square is invisible to the model. A thunderstorm updraft is only a couple of miles wide, so a coarse model literally cannot see it. It can only sense that conditions are ripe somewhere in a general area and fill in the rest with a statistical shortcut called convective parameterization - essentially the model saying, “there should be about this much storm activity somewhere in this box, but don’t ask me where.”

How Is the HRRR Different From Older Forecast Models?

The HRRR throws that shortcut out. It runs on a 3-kilometer grid - under 2 nautical miles per square - coast to coast across the continental U.S. At that scale, the model crosses a threshold meteorologists call convection-allowing: the grid is fine enough to actually grow an individual thunderstorm inside the simulation. It doesn’t guess that storms will form; it builds them cell by cell and shows where they’ll pop and how they’ll move.

That’s a fundamentally different product, not just a sharper version of the old area forecast. It’s an actual simulated radar image of the future.

Why Does the HRRR Update Every Hour?

The HRRR refreshes every hour - a complete run, top to bottom, 24 times a day. Most runs forecast out 18 hours, while the longer runs at the four synoptic times extend to 48 hours.

Hourly matters because convection is fast and chaotic. A forecast built twelve hours ago is working with twelve-hour-old information about a situation that reorganizes itself every fifteen minutes. The HRRR constantly folds in the newest data and starts over. So when you refresh your app at noon and again at one o’clock and the storm timing shifts, that isn’t the app glitching - it’s a brand-new supercomputer run that just ingested another hour of reality.

How Does Radar Make the HRRR So Accurate Short-Term?

Radar is the ingredient that really sets the HRRR apart. Most models are fed mainly by balloon soundings, surface stations, satellites, and aircraft reports. The HRRR uses all of that too, but it also swallows the national NEXRAD radar mosaic. In the moments before each hourly run kicks off, it assimilates radar reflectivity in roughly 15-minute chunks over the preceding hour.

The practical effect is significant. The model isn’t just told the atmosphere is unstable - it’s shown the storms that already exist right now on radar and uses them to nudge its own simulated storms into the right position before forecasting forward. This technique, called radar reflectivity assimilation, is a big reason the HRRR performs so well in the zero-to-six-hour window - exactly the window that matters when you’re deciding whether to launch.

Who Builds and Runs the HRRR?

The HRRR is a product of the National Oceanic and Atmospheric Administration (NOAA). It was developed largely at NOAA’s laboratory in Boulder, Colorado - the group now known as the Global Systems Laboratory - and runs operationally at the National Centers for Environmental Prediction on the national weather supercomputers.

Under the hood, the version most pilots have been using is built on a research model called the WRF (Weather Research and Forecasting model) in its advanced research configuration. The system grew out of the research community, went live nationally in 2014, and has been upgraded roughly every year or two since - each version bringing better handling of low clouds, sharper icing signals, and improved near-surface wind fields.

What Are the HRRR’s Limitations?

A weather model is a tool with sharp edges. Four limits are worth understanding before you lean on it.

It shows skill, not pinpoint detail. The HRRR can tell you a line of storms will cut across your route between three and five in the afternoon. But if it draws that line ten or fifteen miles east of where it sets up and shifts the timing by an hour, that’s completely normal. Fly the pattern the pixels are showing you, not the exact pixels.

It’s deterministic. For most of its life the HRRR has produced one run, one answer, with no built-in statement of confidence - and a confident-looking wrong answer is the most dangerous kind. The workaround is to compare consecutive runs. If noon, one o’clock, and two o’clock all paint the squall line in about the same place, your confidence should climb. If the storms jump around every run, the atmosphere hasn’t decided yet. An experimental HRRR ensemble runs multiple members to address uncertainty directly, but the everyday product in your app is usually the single deterministic run.

It’s a short-range tool. Eighteen hours, mostly. This is the model for this afternoon and this evening - not for planning next weekend’s trip.

It’s computationally expensive. Running a 3-kilometer simulation over the entire country 24 times a day, on a deadline, is a brutal problem - and a forecast that arrives late is worthless. That cost is precisely why the grid is 3 kilometers and not 1, and why most runs stop at 18 hours rather than 48. Every one of those numbers is a tradeoff between forecast quality and delivery speed.

Where Do Pilots Actually See HRRR Output?

More places than you’d expect. The short-range depictions in the Graphical Forecasts for Aviation (GFA) on aviationweather.gov lean heavily on high-resolution guidance like this. When your flight-planning app shows a simulated radar or convective forecast layer - sometimes just labeled “forecast radar” - you’re very often looking at HRRR output. And when your electronic flight bag animates a future radar loop, that future came from a convection-allowing model, which in the U.S. usually means the HRRR or its close relatives.

How Should I Use the HRRR to Time Thunderstorms?

Treat it as a fast, detailed second opinion you cross-examine - not an oracle. Say it’s a July Saturday with cumulus building to the west. Don’t just glance at current radar, which only describes the storm that already exists. Pull the forecast radar loop and watch where the model grows new cells over the next three or four hours. Then, critically, refresh it an hour later and see whether the new run agrees with the old one.

Two runs in agreement is a plan. Two runs that disagree is a warning that the atmosphere is unpredictable today - which is itself one of the most useful things a model can tell you.

What Is Replacing the HRRR?

The HRRR is on its way out, and that’s a good thing. The WRF model it’s built on is aging research code, and NOAA is transitioning to a new unified system called the Rapid Refresh Forecast System (RRFS). The RRFS is built on a more modern engine - the FV3 dynamical core, the same family that powers the American global model. The goal is to retire the current zoo of separate models (the global model, the HRRR, the older Rapid Refresh), each with its own quirks, and consolidate onto one modern framework.

Crucially, the RRFS is designed to be an ensemble from the ground up. Multiple members are baked in, so the uncertainty pilots currently reverse-engineer by comparing runs becomes a number the system simply hands you. A forecast that knows how much it doesn’t know is worth more than a sharper forecast that’s silently overconfident.

Operational transitions like this never happen overnight. The new system runs in parallel with the old one for a long stretch while forecasters build trust and work out bugs, and target dates have moved before. But the machine quietly timing your thunderstorms is being replaced by something built to be smarter about its own limits.

The Bottom Line for Pilots

The HRRR represents a shift from forecasts that describe the atmosphere to forecasts that simulate it - from a model that says storms are likely in your area to one that grows the storms in front of you and lets you watch them move. It’s a remarkable piece of engineering: 3 kilometers, every hour, radar-fed, coast to coast. But it’s a simulation, not a prophecy. The pilots who get the most from it understand what it can and can’t see, watch it change run to run, and let a disagreeing forecast make them more humble, not less.

Key Takeaways

• The HRRR runs on a 3-kilometer, convection-allowing grid that simulates individual thunderstorms rather than just estimating storm probability in a general area.
• It refreshes every hour and forecasts out 18 hours for most runs (up to 48 hours at the four synoptic times), making it ideal for the same-day, zero-to-six-hour decision window.
• Its radar reflectivity assimilation of the NEXRAD mosaic is a key reason it excels at short-range storm timing.
• The model is deterministic, so compare consecutive runs to gauge confidence - agreement builds trust, disagreement is a warning.
• Developed by NOAA’s Global Systems Laboratory and operational since 2014, the HRRR is being replaced by the ensemble-based RRFS built on the modern FV3 core.