Model predicts joint weather conditions from station data alone
Learning joint probabilistic weather forecasts from station observations alone
Machine Learning
Summary
Predicting weather involves understanding how different weather variables like temperature, wind, and humidity connect to each other. The authors created a model called CLARA that learns these connections using only data from weather stations, without relying on complex weather simulations. This model can predict several weather variables together more accurately across many locations worldwide. It shows better performance than other methods that ignore relationships between variables or use simpler models. CLARA could help improve forecasts that depend on how multiple weather factors combine.
What this means in practice
- •For meteorological agencies: Generate improved joint weather forecasts using only local station data without complex numerical models.
- •For renewable energy operators: Produce more accurate short-term weather predictions across multiple variables to optimize energy production and grid management.
Authors
Chaeyeon Yi, Yun Am Seo
Abstract
Assessing compound weather risks requires forecasts representing dependence between variables. CLARA (Calibrated Advection-Routing Attention) learns joint Gaussian predictive distributions of five surface variables from station observations alone, without numerical weather prediction or reanalysis; the approximately 28,000-parameter model supports CPU training and prediction. Across six multi-year folds on 96 stations, its lead-mean energy score is 4.9% lower than that of a learned comparator with matched temporal inputs (4.7% with a similar parameter count) and 11-65% lower than those of statistical baselines. Holding marginal variances fixed, removing learned correlations worsens joint negative log-likelihood by 1.0-2.8 nats per station. A covariance-scale estimator, proved consistent under stated assumptions, improves short-lead calibration but over-corrects at long leads. Synthetic interventions show an attention-bias coefficient alone does not measure forecast influence. Retrained in ten regions on six continents, CLARA outperforms persistence in all 60 multi-year region-lead comparisons and a similarly sized learned model in 57 of 60.