Keskeiset käsitteet
MSE-based models struggle with extreme weather prediction, but Exloss and ExEnsemble improve accuracy.
Tiivistelmä
- Data-driven ML models outperform traditional physics-based models in weather forecasting.
- Most ML models underestimate extreme values due to MSE loss.
- Exloss corrects underestimation bias by using asymmetric optimization.
- ExEnsemble enhances extreme weather forecast robustness at the pixel level.
- ExtremeCast achieves SOTA performance in extreme value metrics while maintaining overall accuracy.
- Ablation experiments show the importance of Exloss, Diffusion, and ExEnsemble modules.
- Case analysis demonstrates accurate predictions for typhoons and heatwaves.
- Overall forecast accuracy of ExtremeCast is competitive with top ML models.
1. Introduction:
- ML models excel in weather forecasting but struggle with extreme values.
2. Related Work:
- Data-driven ML models show potential for medium-range weather forecasts.
3. Method:
- Md maps input to output, followed by Mg enhancing details, and ExEnsemble integrates outputs.
4. Experiment:
- Dataset: ERA5 reanalysis data used for training and testing.
- Network Structure: Md uses Swin-Transformer, Mg uses U-Transformer for conditional generation.
- Baseline Models: Pangu, GraphCast, FengWu, FuXi, FuXi-Extreme, ECMWF-IFS compared.
5. Conclusions:
Exloss and ExEnsemble improve extreme value prediction in global weather forecasting.
Tilastot
MSE損失による予測は極端な天候を過小評価する。
ExlossとExEnsembleが精度向上に貢献。