Offshore Wind Prediction with Graph Neural Networks

Abstract

Offshore wind energy, especially in regions like the Norwegian Arctic, is a promising source of renewable energy. Graph Neural Networks (GNNs) have shown potential in modeling complex systems like weather, making them suitable for improving wind resource assessments. This thesis investigates the use of GNNs for predicting offshore wind patterns, utilizing high-resolution Synthetic Aperture Radar (SAR) data from Sentinel-1 and the Copernicus Arctic Regional Reanalysis (CARRA) data. The research begins with a thorough exploration of wind data sources to evaluate their reliability. The findings indicate that the SAR-based wind retrieval method offers superior spatial resolution and detail compared to traditional reanalysis products and in situ observations while maintaining an accurate representation of long-term wind resources despite its poor temporal resolution. Experiments with several graph and GNN architectures were conducted to assess their effectiveness in predicting wind fields. Simple GNN architectures generated reasonable two-dimensional wind fields but struggled to capture the detailed variations observed in SAR data. This suggests the need for more sophisticated architectures and additional data inputs to improve accuracy. Key findings highlight the importance of incorporating long-range spatial dependencies, refining performance evaluation methods, and expanding the training dataset with more comprehensive data sources. This thesis represents a first step toward integrating GNNs into offshore wind resource assessments and identifies areas for further exploration.