Skip to main content

Version 1.0

A regional solar irradiance model that predicts multiple components of solar radiation with coverage of North & South America (80°S to 80°N, -5°W to -150°W).

Helios is powered by the EarthNet foundation model, which uses an advanced transformer architecture to learn spatiotemporal relationships across diverse data sources.

Helios has no reliance on Numerical Weather Prediction (NWP) models. Instead, Helios rapidly integrates satellite observations from GOES-16 Avanced Baseline Imager(ABI) track cloud movement, formation, and opacity in near real-time, while solar geometry and clear-sky modeling factor in seasonal, daily, and hourly variations in sunlight. Training labels were derived from NREL's hourly National Solar Radiation Database (NSRDB).

Variables​

  • GHI = global horizontal irradiance (W/m²)
  • DNI = direct normal irradiance (W/m²)
  • DHI = diffuse horizontal irradiance (W/m²)

Coordinates​

  • lat: Latitude, 0.06 degrees, float64`
  • lon: Longitude, 0.06 degrees, float64
  • time: Initial time of forecast, datetime64
  • prediction_timedelta: Forecast lead time as a datetime.timedelta

Location​

Historical data on AWS S3 - s3://zeus-public/helios/v1/helios-v1-historical-0.06-hourly.zarr

Forecast on AWS S3 - s3://zeusai-data/prod/helios/v1/forecast/{year}/{month}/{day}/helios.v1.forecast.6h.{year}{month}{day}{hour}00.zarr

Format​

Historical analysis

<xarray.Dataset> Size: 4TB
Dimensions:  (time: 43848, lat: 2620, lon: 2644)
Coordinates:
  * time     (time) datetime64[ns] 351kB 2020-01-01 ... 2024-12-31T23:00:00
  * lat      (lat) float64 21kB -78.57 -78.51 -78.45 ... 78.45 78.51 78.57
  * lon      (lon) float64 21kB -154.3 -154.2 -154.2 -154.1 ... 4.17 4.23 4.29
Data variables:
    dhi      (time, lat, lon) float32 1TB dask.array<chunksize=(10, 500, 500), meta=np.ndarray>
    dni      (time, lat, lon) float32 1TB dask.array<chunksize=(10, 500, 500), meta=np.ndarray>
    ghi      (time, lat, lon) float32 1TB dask.array<chunksize=(10, 500, 500), meta=np.ndarray>

Forecast

<xarray.Dataset> Size: 639MB
Dimensions:               (time: 1, prediction_timedelta: 7, lat: 2745,
                           lon: 2770)
Coordinates:
  * time                  (time) datetime64[ns] 8B 2025-04-15T17:00:00
  * prediction_timedelta  (prediction_timedelta) timedelta64[ns] 56B 00:00:00...
  * lat                   (lat) float64 22kB -78.57 -78.52 ... 78.52 78.57
  * lon                   (lon) float64 22kB -154.3 -154.2 -154.2 ... 4.24 4.29
Data variables:
    dhi                   (time, prediction_timedelta, lat, lon) float32 213MB dask.array<chunksize=(1, 1, 1000, 1000), meta=np.ndarray>
    dni                   (time, prediction_timedelta, lat, lon) float32 213MB dask.array<chunksize=(1, 1, 1000, 1000), meta=np.ndarray>
    ghi                   (time, prediction_timedelta, lat, lon) float32 213MB dask.array<chunksize=(1, 1, 1000, 1000), meta=np.ndarray>

Getting started​

Historical data is available for model validation and backtesting. Load a historical Zarr file with one line of code in our Colab notebook