Integrating Wind Power into Graph-Based Limited-Area Weather Forecasting Models

Le résumé fourni par la source

Machine learning-based limited-area models (LAMs) have been shown to rival or even outperform conventional numerical weather prediction models at local, high-resolution forecasting tasks. This study investigates how the Encoder-Processor-Decoder architecture, which has been successfully employed in numerous applications, can be adapted for wind power prediction. Leveraging the Anemoi framework developed by the European Centre for Medium-Range Weather Forecasts (ECMWF) and various national weather services, we implement graph-based neural networks over a spatial domain encompassing the North Sea region. Different weather models, including standard graph neural networks and attention-based methods, are trained using high-resolution weather data from the Copernicus Regional Reanalysis for Europe (CERRA). We explore several strategies for incorporating wind power at different stages of the training pipeline, including training weather models jointly with wind power data from scratch, as well as finetuning pretrained weather models specifically for wind power forecasting. Training and verification are performed utilizing the publicly available wind power production data from the European Network of Transmission System Operators for Electricity (ENTSO-E). The impact of input feature selection and architectural design choices on forecast skill is evaluated. In addition, the resulting wind power forecasts are benchmarked against those obtained from conventional physics-based methods and state-of-the-art data-driven approaches. This comparison provides insight into the benefits and limitations of end-to-end learning frameworks for renewable energy forecasting and their operational applicability.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.