EnsDiff: Ensemble Precipitation Nowcasting with Diffusion

TMLR Paper5163 Authors

Abstract: Operational Numerical Weather Prediction (NWP) precipitation nowcasting usually considers forecast reliability by utilizing an ensemble of model forecasts. Existing data-driven methods often optimize MSE deterministically or resort to probabilistic forecasting with generative models. However, they only emphasize the optimization of the point forecast metrics, which makes it challenging to balance the trade-off between the optimization of accuracy and uncertainty. Human experts can hardly make an appropriate decision with an ensemble forecast when forecast calibration and sharpness are not considered. In this paper, we propose EnsDiff, which models the probability distribution to produce ensemble diffusion predictions. Not only does it outperform the existing models on a proper scoring rule, Continuous Ranked Probability Score (CRPS), but it also outperforms others on the deterministic metrics. Extensive experiments show that EnsDiff can enhance probabilistic, deterministic skills, and perceptual quality, outperforming state-of-the-art models.

Submission Length: Regular submission (no more than 12 pages of main content)

Assigned Action Editor: ~Russell_Tsuchida1

Submission Number: 5163