Google DeepMind and Google Research have introduced WeatherNext 3, an AI-based global weather forecasting model that incorporates real-time satellite data and generates hourly forecasts at resolutions as fine as 5 kilometers.

According to a company statement, the model represents a significant upgrade over its predecessor, WeatherNext 2, which produced forecasts on a 25-kilometer grid in six-hour increments. WeatherNext 3 delivers what Google describes as a global weather picture roughly five times sharper than the previous model.

The new model ingests live geostationary satellite data to update forecasts every hour, bypassing the six-hour data lag associated with traditional numerical weather prediction systems. It also trains directly on sparse weather station observation data to account for local topography and regional variations.

In precipitation forecasting, the company reports improvements of up to 60% in Continuous Ranked Probability Score against NASA's IMERG satellite dataset, 30% against MRMS, and 10% against rain gauge measurements for early lead times.

WeatherNext 3 also introduces variables relevant to renewable energy, including 100-meter wind speed forecasts suited for wind turbine planning and high-resolution solar radiation estimates for solar farm output projections.

Google states that WeatherNext 3 is now integrated into Google Search, the Gemini app, Google Maps, the Google Maps Platform Weather API, and Google Earth Engine. The company says users planning one or more days ahead will see up to 50% more accurate precipitation forecasts compared to the previous model, with the largest gains in historically underserved regions.

Forecast data is also available for researchers and developers through BigQuery, Earth Engine, and Google Cloud Storage. Independent live evaluations cited by Google were conducted by Brightband.