Neural Generative Weather Forecasting

01/01/2021 to 31/12/2023

Weather forecasting systems have been optimized by specialists over decades by hand, somewhat similar to image processing systems before the deep learning revolution. Along the lines of the successes of deep learning in other domains, we propose to develop a deep learning-based, generative, temporally predictive, neural weather forecasting system. In particular, we propose the further development of our Distributed, Spatio-Temporal graph Artificial Neural Network Architecture (DISTANA).
DISTANA encodes observable values, propagates the encoded values locally over space and time, and decodes the values to generate value predictions at the next time step—which happens simultaneously in all locally-connected nodes of a mesh. As a result, closed-loop spatio-temporal forecasting is possible. To enhance DISTANA and to scale it up, it is necessary to apply it to larger datasets, which integrate information about the weather and the weather-influencing surrounding geological factors. Accordingly, we propose to combine multiple datasets, creating several prediction benchmarks for the weather in Germany, and to enhance and apply DISTANA to these benchmarks. The superordinate
goal of this project is to further develop DISTANA into a universal forecasting system that is widely applicable to other spatio-temporal prediction tasks, such as water flow, soil erosion, or even wind park turbine power output.