Probabilistic forecasting using deep generative models

A fundamental problem in Numerical Weather Prediction (NWP) is the generation of ensembles to capture the probability of future states of the atmosphere. This research presents a new methodology to generate analogs using deep generative models, an emerging class of deep learning approaches. The goal is to train a deep generative model using a set of historical forecasts and associated observations, and to use it to entirely or partially replace the need to maintain a potentially very large dataset. In this research, this new methodology is compared with the Analog Ensemble (AnEn) approach, a computationally efficient solution to generate analogs. The proposed approach promises to reduce the amount of memory required to produce the probabilistic forecast by several orders of magnitude. Results show that the generative model solution is constant time without performing any search, saving a considerable amount of time even in the presence of huge historical datasets.