Graph neural operator for wind farm wake flow

Abstract

Wind farm flow simulations are computationally expensive. However, numerous simulations are often required to account for wake effects from multiple neighboring farms, which motivates the development of data-driven surrogate models. Such surrogates may also enable the consideration of neighboring farm wakes during layout optimization. Most existing data-driven surrogate approaches rely on fixed algebraic superposition principles, which may limit flexibility when extending to higher-fidelity data sources that capture nonlinear wake interactions. We propose a novel method that embeds a trainable and scalable superposition principle within a graph neural operator (GNO) architecture, enabling the model to learn complex wake combinations directly from data. The model consists of two sequential graph neural network (GNN) layers: the first encodes turbine–turbine interactions into a latent representation, while the second combines these latent turbine states to predict the wind speed at a desired location. The GNO is trained on a large dataset generated with PyWake, a steady-state engineering wake model. More than 2000 unique layouts were procedurally generated for training, with an additional 998 for testing. The GNO accurately identifies regions of strong wake interaction, although the spatial extent of wakes is slightly underestimated compared to the simulated values in cases with pronounced wake effects. Overall, the proposed GNO represents a methodological advancement in data-driven wind farm flow surrogates, introducing a new conceptual framework inspired by established engineering wake modeling principles.