AI weather models have struggled for years with two stubborn issues: forecasts that are too coarse to capture local terrain features like valleys, coastlines, and mountains, and a six‑hour delay in initialization because they rely on outdated numerical weather prediction analyses. These gaps limit the usefulness of AI‑driven forecasts for applications that need timely, high‑resolution detail—such as renewable‑energy grid management, aviation routing, and local disaster preparedness.
Google DeepMind’s WeatherNext 3 tackles both problems head‑on. Instead of waiting for a six‑hour‑old NWP analysis, the model ingests a live global geostationary satellite mosaic every hour, allowing it to re‑initialize 24 times a day. This eliminates the data‑lag bottleneck and ensures the model starts from the most current atmospheric picture.
Resolution is also sharpened dramatically. WeatherNext 3 outputs three tiers from a single forward pass: 0.05° (~5 km) station‑trained temperature and dew point, 0.1° (~10 km) gridded surface variables (wind, pressure, sea‑surface temperature, cloud, solar radiation, precipitation), and 0.25° (~25 km) atmospheric fields across 13 pressure levels. By training directly on raw weather‑station observations rather than smoothed reanalysis grids, the model’s low‑level fields reflect what instruments actually measure, preserving the fine‑scale variations caused by local topography.
Probabilistic skill is boosted with a 64‑member ensemble. Synoptic cycles at 00, 06, 12, 18 UTC produce 15‑day forecasts, while interim hourly runs cover 48 hours, giving operators both long‑range outlook and short‑term, high‑frequency updates for fast‑developing convection.
Precipitation forecasts benefit from training on multiple sources—ECMWF reanalysis, NASA’s IMERG satellite retrievals, and Google’s own satellite‑radar reanalysis—yielding up to a 60 % CRPS improvement over IMERG at early lead times and sharper rain‑band edges that reduce the “smeared blob” effect of coarser models.
For clean‑energy stakeholders, the model supplies 100 m wind speed, full cloud‑distribution profiles, and both solar irradiance components (SSRD and FDIR), delivering the variables grid operators need to balance wind and solar generation against demand.
Access to the forecast data is available now via BigQuery, Earth Engine, and Cloud Storage after an allow‑list request; the model weights remain proprietary, but the output can be integrated directly into operational workflows.
#AI #WeatherForecasting #MachineLearning #DeepLearning #RenewableEnergy #TechInnovation