A Graph Neural Network approach for wake prediction in offshore wind farms using Large Eddy Simulations
C.W. van den Berg (TU Delft - Technology, Policy and Management)
R.A. Verzijlbergh – Mentor (TU Delft - Technology, Policy and Management)
Stefan Pfenninger-Lee – Mentor (TU Delft - Technology, Policy and Management)
Amineh Ghorbani – Mentor (TU Delft - Technology, Policy and Management)
More Info
expand_more
Other than for strictly personal use, it is not permitted to download, forward or distribute the text or part of it, without the consent of the author(s) and/or copyright holder(s), unless the work is under an open content license such as Creative Commons.
Abstract
The large-scale development of offshore wind is a complex socio-technical challenge in which technological performance, economic viability, government policy, energy markets, and the decisions of multiple actors are strongly interconnected. Developers and investors must make long-term design, bidding, and investment decisions under considerable uncertainty. Technical modelling uncertainty can therefore propagate beyond engineering calculations into financial assessments, tender strategies, and ultimately the viability of offshore wind projects. Improving the information available to decision-makers is consequently important for individual projects and the wider transition towards a low-carbon electricity system.
Wake prediction represents one important source of uncertainty. Wind turbines reduce downstream wind speeds, affecting neighbouring turbines and the expected energy yield of a wind farm. Engineering wake models such as Jensen are computationally efficient but rely on simplifying assumptions, while high-fidelity Large Eddy Simulations (LES) provide more detailed representations of wake behaviour at much greater computational cost. This creates a trade-off between prediction accuracy and computational effort, particularly when many layouts or scenarios must be evaluated.
This thesis investigates whether a Graph Neural Network (GNN) trained on high-fidelity LES data can improve offshore wind-farm wake predictions while maintaining computational efficiency compared with traditional engineering wake models. A turbine-interaction GNN was developed in which turbines are represented as nodes and potential wake interactions as directed edges. Through message passing, the model learns how upstream turbines influence downstream wind-speed deficits. The architecture was first verified using synthetic Jensen data and subsequently evaluated using the WINS50 LES dataset. The experiments investigated heterogeneous inflow information, turbulence intensity, atmospheric-stability indicators, graph construction, and alternative message-passing architectures. Finally, a multi-farm GNN was trained on 25 wind farms and evaluated on the unseen future_052 layout, using Jensen/NOJ as the principal engineering-model baseline.
The experiments demonstrate that the GNN can successfully learn turbine–turbine wake relationships. Additional model information and complexity did not automatically improve prediction accuracy. Local wind speed provided useful information about heterogeneous conditions, while local wind direction offered limited additional benefit. Turbulence intensity generally improved performance, and exploratory experiments indicated that atmospheric-stability information may provide further predictive value. Wider interaction corridors improved predictions, whereas full connectivity offered only a small additional improvement at substantially greater computational cost. An attention-based architecture also did not outperform the edge-conditioned NNConv model, indicating that a relatively simple, physically informed graph representation can efficiently capture relevant wake interactions.
The strongest result was obtained in the cross-farm experiment. When trained on 25 wind farms and evaluated on the excluded future_052 layout, the GNN achieved a mean absolute error of 0.7888 m/s, compared with 1.0859 m/s for Jensen, representing an approximately 27% reduction in prediction error. The GNN also exhibited virtually no mean prediction bias. Including graph construction, evaluation of the complete test set was approximately 4.5 times faster than the Jensen implementation. These results demonstrate that wake relationships learned from computationally expensive LES simulations can be transferred to an unseen wind-farm layout while retaining computational efficiency.
An illustrative valuation case translated predicted wind speeds into annual energy production (AEP) and discounted lifetime revenues. For the considered wind farm, the LES reference resulted in an AEP of approximately 5318 GWh. The GNN overestimated AEP by approximately 0.6%, corresponding to a discounted lifetime revenue error of approximately EUR 22.8 million under the case-study assumptions. The Jensen model produced a substantially larger energy-production and revenue error. Although this was not a complete financial valuation, it demonstrated how technical modelling errors can propagate into economically significant differences in project assessment.
Overall, LES-trained turbine-interaction GNNs provide a promising link between high-fidelity physical simulation and the computational efficiency required for repeated wind-farm assessment. Improved prediction accuracy and faster evaluation could help developers and investors compare layouts, expected revenues, investment alternatives, and tender strategies more reliably, while enabling more scenarios to be assessed under uncertainty. With further validation, these models could provide a more reliable technical basis for wind-farm design and investment decisions, supporting the large-scale development of offshore wind capacity.