Prior knowledge-constrained deep learning for probabilistic precipitation downscaling

Precipitation downscaling, particularly at convection-permitting scales (less than 4 km), is highly uncertain. This is especially pronounced in mountainous regions due to the interplay of complex topography and atmospheric dynamics. It impedes reliable estimation of variability and risks in localized extreme rainstorms. Deep learning-based downscaling methods have gained increasing attention but have primarily focused on deterministic prediction, which fails to capture uncertainty. Here we developed a novel Probabilistic High-resolution Precipitation Downscaling Network (P-HRDNet) with prior knowledge of key precipitation characteristics to design its loss function and model architecture. This knowledge includes data imbalance, skewed distribution, heteroscedasticity, and spatial and temporal dependencies of precipitation. P-HRDNet was tested in the southeastern Tibetan Plateau, a mountainous region lacking high-resolution precipitation data. Ten-year WRF simulations with nested domains provided coarse (9 km) and fine resolution (1 km) daily precipitation to train P-HRDNet. Compared with a baseline model SRCNN, P-HRDNet achieved greater accuracy in terms of root mean square error, mean absolute error, and Pearson correlation coefficient. Besides, it offers better uncertainty coverage and narrower uncertainty widths. This superiority is particularly evident in the extreme values. Our study highlights the importance of incorporating prior knowledge of precipitation characteristics into deep learning, and has a potential to physically constrain Artitifical-Intelligience (AI) based weather forecasting models. Furthermore, our WRF-AI framework offers an efficient solution for obtaining reliable high-resolution precipitation estimates in poorly gauged regions.