Blog
AI Weather Data: From Forecast to Operations
AI weather data is moving from an information service to a direct operational input. Google introduced WeatherNext 3 on September 3. According to the company, the model generates global forecasts every hour using current geostationary satellite data and weather-station observations.
What changes technically
Google specifies a grid spacing of up to 5 kilometres for temperature and moisture, 10 kilometres for other surface variables, and 25 kilometres for atmospheric variables such as wind speed. According to Google, WeatherNext 2 used a 25-kilometre grid and six-hour update intervals.
The data is available globally in Google products. Companies can also query it through BigQuery and Earth Engine or obtain it from Google Cloud Storage. Variables such as wind speed at 100 metres, cloud cover, and solar radiation explicitly address renewable-energy planning.
Accuracy is not an operating model
A more frequent and detailed forecast can improve dispatching, energy use, field service, or supply chains. It also increases the temptation to translate a single model output directly into a process action.
ECMWF therefore describes operational machine-learning weather prediction as more than a model: it requires datasets, infrastructure, shared standards, monitoring, and integration into a complete forecasting chain. The WMO additionally stresses that AI should support national weather services and human expertise rather than replace them. Google likewise refers users to local meteorological agencies for official warnings.
Five controls for DACH companies
Treat a weather API as an external data product with measurable service quality:
- Freshness: How old are the observation, model run, and retrieved forecast?
- Resolution: Does the grid fit the location, particularly in valleys, near lakes, or in dense urban areas?
- Uncertainty: Which range or ensemble informs the decision?
- Fallback: Which source takes over during outages, drift, or conflicting signals?
- Decision rights: Who may change thresholds, stop automated actions, and override official warnings?
Start with a reversible decision, such as scheduling a field-service day. Record the forecast version, threshold, triggered action, and actual outcome. Only when false alarms, missed events, and economic value are visible together should the data stream enter more critical processes.
The decisive question is not which weather model leads a benchmark. It is whether your operation can explain which forecast triggered which decision, and when.