Global 10-minute meteorology via climate-aware large language model for renewable energy modelling
Discuss this preprint
Start a discussion What are Sciety discussions?Listed in
This article is not in any list yet, why not save it to one of your lists.Abstract
High-temporal-resolution meteorological data are essential for renewable energy modelling, yet global reanalyses remain at hourly resolution, introducing systematic biases in renewable modelling. Here, we propose a large-language-model-driven imputation framework that reconstructs global 10-minute fields from hourly inputs using climate-aware prompts incorporating Köppen climate, IPCC regions, and geolocation priors. Fine-tuned on 10-minute satellite observations, the model improves normalized mean absolute error by 55% over state-of-the-art baselines and reduces it by 40–70% in extreme regions such as the Andes, Tibetan Plateau and Kalahari Desert. Global renewable-energy modelling demonstrates that hourly inputs can misestimate project-scale generation by up to tenfold for solar PV and twentyfold for wind in cloud-transient, complex-terrain regimes. Prospective global forecast reveals persistent annual biases of 20–55 TWh in global generation, comparable to Scotland’s annual electricity demand, highlighting the critical need for sub-hourly meteorology. The open global 10-minute dataset bridges a key gap for fine-grained renewable energy modelling.
