Spatial domain | Continental United States |
Spatial resolution | 3km |
Time domain | Forecasts initialized 2018-07-13 12:00:00 UTC to Present |
Time resolution | Forecasts initialized every 6 hours. |
Forecast domain | Forecast lead time 0-48 hours ahead |
Forecast resolution | Hourly |
⎘
The High-Resolution Rapid Refresh (HRRR) is a NOAA real-time 3-km resolution, hourly updated, cloud-resolving, convection-allowing atmospheric model, initialized by 3km grids with 3km radar assimilation. Radar data is assimilated in the HRRR every 15 min over a 1-h period adding further detail to that provided by the hourly data assimilation from the 13km radar-enhanced Rapid Refresh.
This dataset is an archive of past and present HRRR forecasts. Forecasts
are identified by an initialization time (init_time
) denoting the
start time of the model run.
Each forecast has an hourly forecast step along the lead_time
dimension. This dataset contains only the 00, 06, 12, and 18 hour UTC
initialization times which produce the full length, 48 hour forecast.
This dataset uses the native HRRR Lambert Conformal Conic projection,
with spatial indexing along the x
and y
dimensions.
The example notebook shows how to use the embedded spatial reference to
select geographic areas of interest.
Open notebook in github | Open notebook in colab |
import xarray as xr # xarray>=2025.1.2 and zarr>=3.0.8 for zarr v3 support
ds = xr.open_zarr("https://data.dynamical.org/noaa/hrrr/forecast-48-hour/[email protected]")
ds["temperature_2m"].sel(init_time="2025-01-01T00", x=0, y=0, method="nearest").max().compute()
min | max | units | |
---|---|---|---|
init_time | 2018-07-13T12:00:00 | Present | seconds since 1970-01-01 |
lead_time | 0 days 00:00:00 | 2 days 00:00:00 | seconds |
x | -2700000 | 2700000 | m |
y | -1600000 | 1600000 | m |
units | dimensions | |
---|---|---|
categorical_freezing_rain_surface | 0=no; 1=yes | init_time × lead_time × y × x |
categorical_ice_pellets_surface | 0=no; 1=yes | init_time × lead_time × y × x |
categorical_rain_surface | 0=no; 1=yes | init_time × lead_time × y × x |
categorical_snow_surface | 0=no; 1=yes | init_time × lead_time × y × x |
composite_reflectivity | dBZ | init_time × lead_time × y × x |
downward_long_wave_radiation_flux_surface | W/(m^2) | init_time × lead_time × y × x |
downward_short_wave_radiation_flux_surface | W/(m^2) | init_time × lead_time × y × x |
expected_forecast_length | seconds | init_time |
geopotential_height_cloud_ceiling | gpm | init_time × lead_time × y × x |
ingested_forecast_length | seconds | init_time |
latitude | degrees_north | y × x |
longitude | degrees_east | y × x |
percent_frozen_precipitation_surface | % | init_time × lead_time × y × x |
precipitable_water_atmosphere | kg/(m^2) | init_time × lead_time × y × x |
precipitation_surface | mm/s | init_time × lead_time × y × x |
pressure_reduced_to_mean_sea_level | Pa | init_time × lead_time × y × x |
pressure_surface | Pa | init_time × lead_time × y × x |
relative_humidity_2m | % | init_time × lead_time × y × x |
temperature_2m | C | init_time × lead_time × y × x |
total_cloud_cover_atmosphere | % | init_time × lead_time × y × x |
valid_time | seconds since 1970-01-01 | init_time × lead_time |
wind_u_10m | m/s | init_time × lead_time × y × x |
wind_u_80m | m/s | init_time × lead_time × y × x |
wind_v_10m | m/s | init_time × lead_time × y × x |
wind_v_80m | m/s | init_time × lead_time × y × x |
The source grib files this archive is contructed from are provided by NOAA Open Data Dissemniation (NODD) and accessed from the AWS Open Data Registry.
Forecasts initialized through 2020-12-02T06 UTC include data only for the first 36 hours; steps 37–48 are filled with NaNs. Starting with the 2020-12-02T12 UTC initialization, forecasts cover the full 48 hours.
Storage for this dataset is generously provided by Source Cooperative, a Radiant Earth initiative.
The data values in this dataset have been rounded in their binary floating point representation to improve compression. See Klöwer et al. 2021 for more information on this approach. The exact number of rounded bits can be found in our reformatting code.