Enriching Operational High-Resolution Ensemble Forecasts with StyleGAN-2

Abstract The main approach to account for uncertainty in numerical weather prediction (NWP) is the use of ensemble prediction systems (EPSs), which run several simulations in parallel. However, the huge cost of these systems, especially for high-resolution modeling, severely limits their implementation, and in particular the number of ensemble members. This size constraint can have significant impact on the accuracy and robustness of the distributions sampling, as well as on the utility of the EPS in low-predictability situations. This study introduces an innovative hybrid forecasting system, combining the strengths of NWP and generative deep learning. The proposed method leverages the properties of the StyleGAN latent space to create a large number of new ensemble members at a reduced cost. The approach is developed using a dataset of forecasts produced by the kilometer-scale AROME-EPS operational at Météo-France. We first illustrate the capability of the generative adversarial network (GAN) to generate realistic samples with a reasonable physical consistency. It is then shown that a GAN-enriched ensemble with a few hundreds of members is able to significantly enhance the sampling of distribution tails, in a way that is consistent with the sampling obtained with a very large AROME-EPS. Probabilistic verification scores are also promising, although they point out some weaknesses of the current setup. Existing bottlenecks and avenues for improvements are discussed. Deep generative models have the potential to transform ensemble forecasting significantly. In this paper, we initiate an exploration into the vast possibilities these methods offer.