Conditioning Deep Learning Weather Prediction Models On Exogenous Fields

Deep learning weather prediction (DLWP) models have recently proven to be a viable alternative to classical numerical integration. Often, the skill of these models can be improved further by providing additional exogenous fields such as time of day, orography, or sea surface temperatures stemming from an independent ocean model. These merely serve as information sources and are not predicted by the model.

In this study, we explore how such exogenous fields can be utilized by DLWP models most optimally and find that the de facto standard way of concatenating them to the input is suboptimal. Instead, we suggest leveraging existing conditioning techniques from the broader deep learning community that modulate the mean and variance of normalized feature vectors in latent space. These, so called, style-based techniques lead to consistently smaller forecast errors and, at the same time, can be integrated with relative ease into existing forecasting architectures. This makes them an attractive avenue to improve deep learning weather prediction in the future.