GFT-US
What if your weather forecast wasn’t just “good enough” but was consistently sharp, accurate, and reliable, updating every hour to support your most important decisions?
That’s what our new model, GFT-US, brings to the contiguous continental U.S. It offers an AI-based alternative to NOAA’s widely used HRRR model. GFT-US is highly accurate, adds new variables, and is available twenty minutes earlier than HRRR:
That's not all, GFT-US supports:
- High resolution forecasts at 3km.
- Forecasts every hour on the top of the hour: GFT-US delivers forecasts every hour, with a median delivery time of 1 minute after the hour. That's 23 minutes faster than HRRR.
- Day-ahead forecasts: Each GFT-US forecast goes up to 30h, compared to the typical 18h HRRR forecast.
- Higher accuracy: For most lead times, variables and regions, GFT-US is more accurate and less biased than HRRR.
- Extensive variable set: GFT-US predicts wind at 10m and 80m, temperature, total precipitation, solar radiation as well as 10m wind gusts.
Game changer for energy forecasting
Probabilistic forecasts for wind and solar: On request, GFT-US can provide explicit uncertainty estimates (10th and 90th percentiles, a.k.a “P10” and “P90”) for wind and solar, making these forecasts a sharp decision-making tool for the energy sector.
Why hourly forecasting matters
Frequent updates are key, especially when dealing with rapidly evolving weather conditions. Hourly forecasts enable businesses, community leaders, and emergency services to adapt quickly and respond confidently to changing conditions:
Energy Producers: Hourly wind forecasts with clear uncertainty bands let wind farm operators precisely optimize work schedules around projected high generation periods, thereby reducing the chance of revenue shortfalls.
Transportation & Logistics: Frequent updates enable dynamic rerouting of shipments around evolving severe weather events, minimizing delays and financial losses.
Emergency Response & Public Safety: Frequent and accurate forecasts allow authorities to act with high confidence when issuing warnings and mobilizing personnel to safeguard life and property.
Why resolution matters
The weather isn't uniform. It's hyper-local. GFT-US's 3km high-resolution forecasts capture nuances like heat islands, wind patterns near mountains, sea breezes and intense isolated storms, transforming weather forecasts into well-defined, actionable insights. In contrast, most AI models currently operate at the resolution of the HRES model from ECMWF (10km)
Comparison between a 10m wind speed forecast of the ECMWF HRES model and GFT-US forecast. Most AI models to date operate at an even lower resolution than HRES.
How does GFT-US stack up?
To demonstrate the advantage of GFT-US, we have collected temperature and wind speed forecast error statistics at over two thousand weather stations distributed across the continental United States. Here is how GFT-US compares to NOAA’s HRRR model at those same weather stations.
In addition to arriving before HRRR, GFT-US is the clear winner for overall temperature forecast accuracy up to hour-24 and overall wind speed forecast accuracy up to hour-30. But what does this mean for a location near you? The following interactive chart allows comparison of forecast and actual temperature and wind speed at over 2000 weather stations across the contiguous United States.
Relative Comparison at Weather Stations
Try the GFT-US model
Create an Earth API account, install the latest silurian SDK and use the snippet below to get more accurate short-term forecasts for any location within the contiguous United States!
from silurian import Earth # version >= 0.0.8
def main():
latitude = 32.77
longitude = -96.79
timezone = "utc"
units = "metric"
client = Earth(api_key=API_KEY)
response = client.weather.experimental.regional.usa(
latitude=latitude,
longitude=longitude,
timezone=timezone,
units=units,
)
print(response)
if __name__ == "__main__":
main()When evaluating forecast performance at individual weather stations, keep in mind that data quality is variable and regional biases exist.
