Weather impacts on renewables agriculture:explainable AI for hyperresolution downscalingfor S2S

Extreme weather events can wreak havoc on human health, infrastructure, and economy, accentuating the need for timely detection and accurate prediction. Current sub-seasonal and seasonal NWP forecast models, e.g. the ECMWF extended range and SEAS5 products, lack the spatial resolution necessary for applications operating at a smaller scale. Dynamic downscaling can improve spatial and temporal granularity, but is computationally expensive, whereas traditional statistical downscaling requires high-resolution target data. The application of machine learning has shown promise in weather forecast downscaling, but most methods are not transferable to new regions, disregard physical boundaries, and often oversimplify the downscaled fields, thereby diminishing the representation of extreme events.

Our proposed project idea for the Harry-Otten Prize aims at developing a novel, computationally inexpensive, transferable, and physics-aware post-processing methodology including a pre- and post-processing framework addressesing these challenges. We aim to develop a low-cost computational setup, incorporating physics-informativeness and tail-awareness, establish trustworthiness within the scientific community and the public, and assess transferability to untrained regions.

Our proposed methodology combines ideas from a statistical residual fitting approach with local information (topography, land-use and physics-informed
machine learning techniques. More specifically, we suggest a GAN-setup consisting of a ConvLSTM and a physics-aware UNET. Using specified loss
functions accounting both for physics-awareness and upper/lower tails of the distribution, aiming at improving the prediction of extreme events, as well as
handling computations more efficiently. The pre-trained model can be transferred to untrained regions and enable usage of in regions with only a few local data or few computational resources.

This project has the potential for significant societal impacts, particularly in the agricultural and renewable energy sectors. By providing more accurate, highresolution predictions, it supports decision-making for optimizing harvests and reliable energy production in region that are severely affected by extreme weather and climate events. The transferable, low-cost methodology allows for easy adaptation to specific regions and applications, which is very valuable in regions with limited access to HPC resources. To this end, our framework can help enhancing societal resilience to adverse weather conditions worldwide.

Expected outcomes include an open-source deep learning framework for postprocessing and downscaling sub-seasonal and seasonal weather predictions, an improved accuracy in representation of extreme weather events, and an evaluation of the benefits of dynamical downscaling. The proposed solution is unique in its ability to increase spatial resolution to convection permitting scale and its transferability, both in terms of application to different regions and computational platforms.