核心概念
Conformer introduces continuous attention to model spatio-temporal weather dynamics effectively.
摘要
Conformer addresses the limitations of NWP models and data-driven approaches by incorporating continuous attention to capture evolving weather features. It outperforms existing models at all lead times, showcasing its effectiveness in weather forecasting.
統計資料
Conformer outperforms some existing data-driven models at all lead times.
Training Conformer takes about 5 days, significantly faster than other methodologies.
Inference time for Conformer is less than 20 seconds using a single GPU.