Combining Artificial Intelligence and Multifractals for Precipitation Nowcasting: the UM-GAN Example.

Intermittency is a defining characteristic of rainfall, yet it is largely overlooked in most IA nowcasting models. We emphasise its theoretical significance at various stages of the prediction process, from training to assessing its accuracy, including its dispersion relative to the intrinsic limits of predictability. Specifically, we develop a hybrid framework based on:- The generative adversarial network (GAN), a recently developed technique for training IA models through an adversarial process; Universal multifractals (UM), stochastic models of intermittency that are physically based on the cascade paradigm. They are universal in the sense that they are statistically attractive to other processes and depend only on three scale-independent parameters that are physically meaningful. In terms of physical relevance, we evaluate the nowcasting performance of the hybrid UM-GAN model and other baseline models (ConvLSTM, GAN) using continuous and categorical scores, as well as UM analysis in comparison to the observations. The results indicate that UM-GAN achieves the highest scores and accuracy, particularly demonstrating superior performance at lead times of 30 minutes and 60 minutes.