EPT-2 Technical Report on AI Weather Forecasting

This technical report presents EPT-2, the latest version of the Earth Physics Transformer (EPT) models developed by Jua.ai for weather-dependent energy trading. EPT-2 demonstrates significant advancements over its predecessor, EPT-1.5, particularly in predicting energy-relevant variables such as wind speed, temperature, and solar radiation across a 0 to 240-hour forecast horizon. The report details the model's architecture, operational specifications, and performance benchmarks against leading weather models, including Microsoft Aurora and the ECMWF IFS HRES. Additionally, it introduces EPT-2e, a perturbation-based ensemble model that enhances probabilistic forecasting capabilities. The report outlines the model's ability to provide forecasts at any lead time, its operational advantages, and the methodologies used for performance evaluation, including grid and point benchmarks. The findings indicate that EPT-2 achieves lower root mean square error (RMSE) and continuous ranked probability score (CRPS) compared to existing models, marking a notable improvement in the field of AI-driven weather forecasting.