Polar Weather Modelling with Graph Neural Networks

Abstract

Accurately predicting wind speed is essential for assessing the viability of new offshore wind farm locations. However, forecasting wind patterns in regions such as the Norwegian Arctic presents significant challenges due to complex terrain and variable atmospheric conditions. Graph Neural Networks (GNNs), which are well-suited for learning from spatially structured and interconnected data, offer a promising alternative to traditional modeling approaches in wind resource assessment. This project continues the exploration of GNNs for offshore wind forecasting by integrating high-resolution Synthetic Aperture Radar (SAR) data from Sentinel-1 with the Copernicus Arctic Regional Reanalysis (CARRA) dataset. A previously developed GNN architecture was adapted and evaluated for its ability to model wind fields. Building on this foundation, a new and more sophisticated GNN architecture was designed to further enhance performance, with particular attention to capturing the spatial variability present in SAR observations. The results demonstrate encouraging progress toward the integration of GNN-based models in offshore wind resource analysis, although there is still room for improvement.