Convolutional versus graph-based surrogate models for inter-farm wake prediction using multi-fidelity transfer learning

Accurate prediction of wind farm wake interactions is important for energy yield assessment, as offshore wind farms increasingly operate in close proximity to one another. This work presents a systematic comparison of two neural network surrogate models for inter-farm wake deficit prediction: a convolutional neural network (CNN)-based attention residual U-Net (ARU-Net) and a graph neural network (GNN)-based graph neural operator (GNO). Both architectures are trained using multi-fidelity transfer learning, pre-training with low-fidelity engineering model simulations, and then fine-tuning with high-fidelity Reynolds-averaged Navier–Stokes actuator wind farm (RANS-AWF) data. The models are evaluated on procedurally generated wind farm layouts spanning diverse farm sizes, turbine spacings, wind speeds, and ambient turbulence intensities. Both architectures achieve high prediction accuracy but exhibit complementary strengths: evaluated over the wake region (δw≥10-3 m s−1) and averaged across both evaluation grids, the GNO achieves a lower root mean squared error (RMSE) (0.024 vs. 0.028 m s−1), while the ARU-Net attains a higher F1 score (0.98 vs. 0.91), reflecting its superior wake boundary capture. Transfer learning substantially benefits the ARU-Net, while the GNO shows a smaller but measurable improvement, achieving its best performance with transfer learning.