Quantifying the uncertainty of precipitation forecasting using probabilistic deep learning

Abstract. Precipitation forecasting is an important mission in weather science. In recent years, data-driven precipitation techniques could complement numerical prediction, such as nowcasting, monthly projection and extreme event identification. forecasting, the predictive uncertainty arises mainly from data model uncertainties. Current deep learning methods parametric by random sampling parameters. However, usually ignored process derivation of incomplete. this study, input uncertainty, target are jointly modeled a framework to estimate uncertainty. Specifically, estimated priori propagated forward through weights according law error propagation. The considered parameters coupled with uncertainties objective function during training process. Finally, produced propagating testing experimental results indicate that proposed joint modeling for exhibits better accuracy (improving RMSE 1 %–2 % R2 %–7 on average) relative several existing methods, reduce ?28 approach Loquercio et al. (2020). incorporation changes distributions method can slightly smooth weights, leading reduction improved incorporating reducing precipitation. developed be regarded general applications.