Speaker
Description
Machine learning surrogate models for wind farm wake prediction often require computationally expensive, full-farm training data and lack physical interpretability. To address this, a graph-attention surrogate model is proposed for farm-scale wake and power prediction, trained exclusively on a 3-turbine configuration. By representing turbines as nodes and wake interactions as edges, the model learns local flow physics, enabling it to generalize to large-scale wind farms without retraining. To ensure physical consistency, the model embeds direction-dependent wake overlap areas into the attention mechanism as a logit bias, yielding physically interpretable attention weights. Additionally, each node tracks velocity deficit and turbulence intensity, allowing wakes to depend on local inflow, which is crucial for deep arrays. Validated against held-out farm-scale Floris simulations, the surrogate accurately predicts velocity deficits and total array power in a fraction of the time, demonstrating robust scaling behavior in deep array configurations.