Keywords :
Offshore Wind Farms, Foundation Models, Multimodal Learning, Continual Learning, Time-series Forecasting
Abstract :
[en] Offshore wind farms (OWFs) face complex, interdependent challenges that span power forecasting, maintenance, control, and grid integration. Current AI approaches address these challenges in isolation, requiring separate models for each task. We introduce OffWindFM, a unified multimodal foundation model (FM) that integrates heterogeneous data streams, such as SCADA measurements, meteorological forecasts, spatial farm layouts, and maintenance records, through specialized sub-encoders and a shared transformer-based fusion backbone. Pre-trained via a masked multi-horizon forecasting objective and fine-tuned with lightweight task heads, OffWindFM delivers end-to-end power-and-load forecasts and is readily extensible to maintenance, control optimization, market bidding, and provides a scalable foundation for comprehensive wind farm digital twins.
