CLARA: probabilistic weather forecasts trained solely on weather station data, without the use of numerical models or reanalysis data

With around 28,000 parameters and the ability to be trained on a CPU, this model simultaneously predicts five surface variables and outperforms all others in the 60 comparisons carried out across six continents.

To assess compound weather risks, such as extreme heat associated with dry air, forecasts are needed that account for the relationships between variables. Researchers Chaeyeon Yi and Yun Am Seo present CLARA (Calibrated Advection-Routing Attention), a model that learns joint Gaussian distributions for five surface variables solely from station observations, without the need for numerical prediction or reanalyses.

A tiny model

With some 28,000 parameters, CLARA trains and predicts on a simple processor. Tested on 96 stations and six multi-year periods, it reduces its average energy score by 4.9% compared to a comparable learned model, and by 11 to 65% compared to reference statistical methods. Removing learned correlations significantly degrades the joint likelihood, which confirms the value of modeling the dependence between variables.

Retrained in ten regions spread over six continents, the model surpasses persistence in all 60 comparisons and a learned model of similar size in 57 out of 60 cases: an interesting avenue for regions with stations but few computing resources.