MeteoSwiss: Temperature Forecasting with Graph Neural Networks

The challenge

Traditional numerical weather prediction models are computationally expensive and can struggle with local-scale temperature variations. MeteoSwiss needed a data-driven approach to complement their operational forecasting, leveraging the dense SwissMetNet network of 302 stations.

Our solution

We implemented and compared 4 architectures with increasing levels of spatial modeling: TCN (baseline without graph), RNN with node embeddings, STGNN with graph convolutions, and STGNN-Att with learned graph structure via attention. The spatial graph was based on geographic distance between stations (Gaussian kernel + K-nearest neighbors). Training used Energy Score (CRPS) loss for probabilistic calibration.

Key results

• STGNN model achieves mean error of 1.44°C, 17.4% less than TCN baseline
• At 1h horizon, our models beat ICON-CH1-EPS numerical predictions by 54%
• Spatial message-passing improves performance by 6.9% (MAE) and 7.1% (Energy Score)
• Benchmark across 1h to 24h horizons, direct comparison with MeteoSwiss operational NWP

Related case studies