The High Resolution Rapid Refresh, NOAA's Three-Kilometer Model Running Every Hour, and the Short-Range Forecast Revolution Changing How Pilots Plan Around Convection

The High Resolution Rapid Refresh (HRRR) is a numerical weather prediction model operated by NOAA that runs at three-kilometer horizontal grid spacing and produces a fresh forecast cycle every hour. It covers the continental United States, with regional domains for Alaska, Hawaii, and Puerto Rico. For general aviation pilots planning around convection, it is the most skillful short-range guidance tool available - and it is almost certainly already embedded in the weather products on your Electronic Flight Bag.

What Is the HRRR and Who Built It

The HRRR was developed through NOAA’s Environmental Modeling Center and the Earth System Research Laboratories in Boulder, Colorado, led by Stan Benjamin and his team over roughly a decade of experimental work beginning in the mid-2000s. It became fully operational at NOAA’s National Centers for Environmental Prediction in September 2014.

The model sits on top of a parent model called the RAP (Rapid Refresh), which runs at thirteen kilometers across the full North American domain every hour. The HRRR nests inside the RAP at three-kilometer resolution, inheriting its rapid-refresh architecture while resolving features the RAP cannot see.

Why Hourly Updates Are the Core Innovation

Every other major model the aviation community relies on runs far less frequently. The GFS (Global Forecast System) runs four times per day - at 00, 06, 12, and 18 Zulu - at roughly thirteen-kilometer resolution. At that scale, individual storm cells are invisible; the model characterizes the environment that supports convection but cannot place where a storm forms. The NAM (North American Mesoscale model) runs a nested three-kilometer domain over the CONUS but also only four times per day.

The fundamental problem those models share: by the time initialization has completed and post-processing has made data available to users, an hour or more has passed. For synoptic features - fronts, large troughs - that lag barely matters. For convective development on a loaded summer afternoon, it matters enormously.

The HRRR was designed specifically to close that gap.

How the HRRR Ingests Real-Time Observations

Every hour, the HRRR assimilates a continuous stream of real observations rather than starting from a predicted state. The data sources include:

NEXRAD Doppler radar from the Next Generation Radar network
GOES satellite infrared and visible imagery
ASOS (Automated Surface Observing System) surface reports
AMDARs - Aircraft Meteorological Data Relay observations from commercial aircraft
Upper-air soundings
Surface mesonet stations across the country

This process - rapid-refresh data assimilation - is what makes the HRRR different in kind, not just degree, from older high-resolution models. The initial atmospheric state it forecasts from has been continuously corrected against observed reality for the previous hour.

What the HRRR Actually Produces

Most pilots associate HRRR output with simulated radar overlays, but the model produces a full atmospheric suite updated every hour:

Surface temperature and dewpoint
Boundary layer depth
Surface and upper-level winds at multiple altitudes
Precipitation type and intensity
CAPE (Convective Available Potential Energy)
CIN (Convective Inhibition)
Lifted index

The CAPE/CIN combination is particularly actionable for convective planning. CAPE is the energy available to a parcel of air once it accesses the unstable layer above it - the fuel for storm intensity. CIN is the cap that holds convection back. A strong cap over high CAPE is the classic dangerous afternoon setup: the atmosphere looks benign at noon, surface heating erodes the cap through early afternoon, initiation fires, and the storms that develop have enormous energy to work with. The HRRR resolves this thermodynamic structure at three-kilometer spacing, updated each hour - giving pilots not just “is today a convective day” but a running picture of where the cap is weakest and when it will break.

Where the HRRR Proves Most Valuable

Warm-season Great Plains and Midwest flying. The diurnal convective cycle in that corridor - afternoon storms developing along the dryline, organizing into convective systems through the evening, occasionally growing into overnight mesoscale systems pushing east - is where the HRRR’s accuracy is most operationally significant. For pilots flying the central corridor between May and September, HRRR output is the tool departure timing gets built around, not a supplemental check.

Mountain aviation. At three-kilometer resolution, the HRRR resolves significant terrain features that older models at 13–25 km effectively smoothed into broad gradients. The Rockies, Sierra Nevada, and Cascades appear in the model as physically recognizable terrain rather than gentle hills. Orographic lifting that fires convection on windward slopes, wave activity and rotor zones downwind of major ridges - these are features the HRRR can resolve in ways previous publicly available models could not. Forecasters at mountain aviation facilities specifically use HRRR output for convective timing guidance. It is not perfect at the precise scale of rotor location or exact initiation timing above treeline, but the improvement over what existed fifteen years ago is substantial.

Understanding the HRRR’s Limits

The spin-up problem. When the HRRR initializes each hour, it ingests radar data showing where precipitation exists - but not the full three-dimensional thermodynamic structure of every storm cell. For the first one to two hours of any given run, the model is still reconstructing vertical structure from horizontal observations. Spurious convective cells can appear as initialization artifacts; real convection may be slightly misrepresented in intensity or position. Apply more skepticism to HRRR output at forecast hours one and two than at hours four through six.

The eighteen-hour ceiling. The HRRR produces forecasts out to eighteen hours in standard configuration. Beyond that range, convective skill drops sharply - the atmosphere becomes too chaotic at the scales the model resolves. The HRRR is explicitly a day-of-flight tool. For planning two or more days out, the GFS, NAM, and ensemble models remain the appropriate starting point.

Single-solution limitation. The HRRR produces one forecast, one run at a time. That single solution may or may not verify. The correct workflow is to evaluate multiple consecutive hourly runs for consistency. If four or five consecutive runs have placed convective initiation in the same area at roughly the same time, that is a stable signal. If the most recent run has shifted initiation two hours later and a hundred miles east from where it had been sitting, something changed in the initialization - that deserves scrutiny against current observations.

Resolution versus storm-scale tracking. Three kilometers is the finest resolution available in a publicly operational model of this kind. Individual storm cells are often smaller. The HRRR resolves convective environment and broad storm-scale evolution. It does not track individual supercells with the precision of a forecaster doing hand analysis of the mesoscale picture. The Storm Prediction Center’s convective outlooks and real-time radar interpretation remain essential complements.

The Correct Workflow for Using HRRR Data

Strategic phase (morning briefing): Use the GFS and NAM to determine whether the day is flyable in principle. Is the synoptic pattern favorable? Does a significant weather system intersect the route? Is the broad convective picture manageable? Longer-range models answer these questions.

Tactical phase (one to two hours before departure): Pull fresh HRRR output to nail down timing. When does that convective line reach the destination? Is there a window before the afternoon build? Can the flight be on the ground with two hours of margin before the HRRR shows the threat arriving?

One important distinction: the HRRR is a forecast tool, not a nowcast. Simulated reflectivity shows where the model projects storms to be at a given forecast hour - it is not a depiction of current conditions even at hour zero. For current storm position, use actual radar. Confusing the two is a common mistake with real safety implications.

All HRRR output is publicly available through aviationweather.gov and NOAA’s operational model archive. It is also accessible through Windy as a selectable model layer and through HRRR-derived convective overlays in most major EFB applications.

What’s Coming Next: The RRFS

NOAA has been developing the Rapid Refresh Forecast System (RRFS) as the eventual successor to both the RAP and the HRRR. Built on the Unified Forecast System (UFS) - an open-source modeling framework with contributions from NOAA, academia, and the private sector - the RRFS is designed to run deterministic forecasts, ensemble forecasts, and regional high-resolution runs from a single unified codebase.

The key capability the HRRR has never had: ensemble output. A single HRRR run produces one solution with no built-in measure of forecast uncertainty. Operators currently approximate uncertainty by comparing multiple consecutive runs - a workable technique that requires experience and time. An ensemble-capable RRFS would provide a direct measure: how spread are the solutions, and how confident is the model in its convective timing? That information would flow eventually to EFB platforms and directly improve go/no-go decision quality.

The RRFS has been progressing through testing and parallel operations alongside the existing HRRR. Current operational status and implementation announcements are tracked by NOAA’s National Centers for Environmental Prediction at ncep.noaa.gov.

The fundamental architecture - high-resolution, rapidly-refreshing, continuously assimilating real-time observations - is not being replaced. It is being improved.

Key Takeaways

The HRRR runs at 3-kilometer resolution every hour, assimilating live radar, satellite, and surface data to start each forecast from a corrected atmospheric state - a fundamentally different approach from models that run four times daily.
Inside six hours, the HRRR is generally the most skillful convective guidance available for the CONUS; beyond eighteen hours, its skill drops sharply and longer-range models take over.
The spin-up artifact affects the first one to two forecast hours of each run - apply extra skepticism to simulated reflectivity at those lead times.
The correct workflow is synoptic models for strategic planning, HRRR for tactical execution: GFS/NAM in the morning to evaluate flyability, fresh HRRR in the final one to two hours before departure for convective timing.
The HRRR is a forecast tool, not a radar substitute - never use simulated reflectivity as a proxy for current storm position.
The HRRR’s successor, the RRFS, will add ensemble capability, giving pilots and forecasters a direct measure of forecast uncertainty rather than inferring it from run-to-run comparison.