As offshore wind farms expand globally, understanding how wakes from one installation affect neighboring farms becomes increasingly important for yield forecasting and operational optimization. Researchers have developed and compared two machine learning surrogate models designed to predict wake deficits across multiple wind farms with greater speed than traditional computational fluid dynamics simulations.
The study evaluated a convolutional neural network architecture (ARU-Net with attention residual U-Net design) against a graph neural network approach (graph neural operator, or GNO). Both models were trained using multi-fidelity transfer learning, a technique that leverages low-fidelity engineering simulations as initial training data before refinement with higher-resolution Reynolds-averaged Navier–Stokes simulations. This training strategy reduces computational burden while improving accuracy.
The results revealed distinct advantages for each approach. The GNO achieved lower overall prediction error (RMSE of 0.024 versus 0.028 m/s), making it more suitable for applications requiring high numerical precision in wake deficit magnitude. Conversely, the ARU-Net demonstrated superior boundary detection, capturing the physical edges of wake regions more accurately with an F1 score of 0.98 compared to 0.91 for the GNO. This strength is valuable for understanding wake extent and interactions in complex farm geometries.
Transfer learning proved essential for the CNN-based model, substantially reducing training time and improving convergence. The GNO benefited more modestly from the multi-fidelity approach, though still showed measurable gains. Both models were tested across diverse scenarios including varying farm sizes, turbine spacing, wind speeds, and ambient turbulence conditions, demonstrating generalization capability.
These findings suggest that practical implementations may benefit from ensemble or hybrid strategies combining both architectures. The CNN's edge-detection prowess complements the GNO's numerical precision, offering wind energy engineers flexible tools for rapid, accurate inter-farm wake assessment—a capability that will become increasingly valuable as offshore wind portfolios mature and farm density increases in limited marine areas.
