EPT-2

EPT-2 is a family of AI weather forecasting models from Jua, built on its Earth Physics Transformer architecture and designed for energy markets. Wind and solar output, demand and power prices all move with the weather, while conventional numerical weather prediction typically updates four times a day and reaches users hours after its observations. The family includes the deterministic EPT-2 (global, about 9 km, hourly steps to 10 days and six-hourly to 20 days, four runs a day), EPT-2 Early (available about 2.5 hours earlier), the EPT-2e ensemble (out to 60 days), EPT-2 RR (a global rapid-refresh model updated 24 times a day), the high-resolution European models EPT-2 HRRR and EPT-2.1 Europa (updated hourly), EPT-2 Reasoning, and EPT-2.1 Helios (a European solar model updated every 30 minutes). EPT-2 RR is initialised directly from real-time satellite radiances and temperature and humidity profiles, and Helios from a short history of satellite imagery combined with atmospheric background fields. Jua's data pipeline ingests satellite imagery from GOES, Meteosat and Himawari alongside station and radar observations. Outputs include temperature, wind speed at heights from 10 to 200 m, humidity, pressure, cloud cover, precipitation and solar radiation, with hindcasts back to 2023 for backtesting. Jua reports that EPT-2 outperforms ECMWF IFS HRES and Microsoft Aurora when verified against more than 10,000 ground stations with its open-source StationBench method. Jua says customers with Alpine assets, including MFT and Statkraft, run EPT-2 HRRR alongside the global EPT-2, and Volue uses Jua's AI weather input in its hourly-updated wind and solar forecasts for Europe and Japan. Forecasts are delivered through the Jua web platform, a REST API and a Python SDK.