This study introduces a transferable Graph Neural Network (GNN)-based model for potential power prediction. By encoding the wind farm as a graph, the GNN captures complex spatial relationships and aggregates information from neighboring turbines. The model learns localized patterns (i.e., subgraphs) across multiple wind farms, enabling it to generalize and adapt to different wind farm layouts. Using Supervisory Control and Data Acquisition (SCADA) data from two offshore wind farms, we evaluate three GNN-based models: a single-farm baseline, a transfer model trained on a source farm, and a multi-task model leveraging data from both farms. Our results demonstrate significant improvements in prediction accuracy. The GNN models achieve relative reductions in Mean Absolute Error (MAE) ranging from 23.3% to 32.4% and in Mean Absolute Percentage Error (MAPE) from 24.3% to 33.5%, compared to the power curve binning method. Furthermore, the wind-direction-frequency-weighted average of energy ratio prediction errors improves by 42.4% to 57.6%. The multi-task model, trained on data from both farms, emerged as the top performer, showcasing the benefits of multi-task learning in enhancing robustness and generalization. Importantly, even the transfer model, trained solely on data from the source farm, outperformed traditional methods when applied to an unseen wind farm. These findings highlight the potential of GNNs to advance power prediction accuracy, particularly for new installations with limited data, and suggest promising avenues for further exploration of multi-task learning in wind farm applications.
Daenens, S, Verstraeten, T, Daems, P-J, Nowe, A & Helsen, J 2025, 'Power Prediction in Offshore Wind Farms using Transferable Multi-Task Graph Neural Networks', Journal of Physics: Conference Series, vol. 3016, no. 1, 012021. https://doi.org/10.1088/1742-6596/3016/1/012021
Daenens, S., Verstraeten, T., Daems, P.-J., Nowe, A., & Helsen, J. (2025). Power Prediction in Offshore Wind Farms using Transferable Multi-Task Graph Neural Networks. Journal of Physics: Conference Series, 3016(1), Article 012021. https://doi.org/10.1088/1742-6596/3016/1/012021
@article{2aa7e79a852e499faf2c24216a1539bc,
title = "Power Prediction in Offshore Wind Farms using Transferable Multi-Task Graph Neural Networks",
abstract = "This study introduces a transferable Graph Neural Network (GNN)-based model for potential power prediction. By encoding the wind farm as a graph, the GNN captures complex spatial relationships and aggregates information from neighboring turbines. The model learns localized patterns (i.e., subgraphs) across multiple wind farms, enabling it to generalize and adapt to different wind farm layouts. Using Supervisory Control and Data Acquisition (SCADA) data from two offshore wind farms, we evaluate three GNN-based models: a single-farm baseline, a transfer model trained on a source farm, and a multi-task model leveraging data from both farms. Our results demonstrate significant improvements in prediction accuracy. The GNN models achieve relative reductions in Mean Absolute Error (MAE) ranging from 23.3% to 32.4% and in Mean Absolute Percentage Error (MAPE) from 24.3% to 33.5%, compared to the power curve binning method. Furthermore, the wind-direction-frequency-weighted average of energy ratio prediction errors improves by 42.4% to 57.6%. The multi-task model, trained on data from both farms, emerged as the top performer, showcasing the benefits of multi-task learning in enhancing robustness and generalization. Importantly, even the transfer model, trained solely on data from the source farm, outperformed traditional methods when applied to an unseen wind farm. These findings highlight the potential of GNNs to advance power prediction accuracy, particularly for new installations with limited data, and suggest promising avenues for further exploration of multi-task learning in wind farm applications.",
author = "Simon Daenens and Timothy Verstraeten and Pieter-Jan Daems and Ann Nowe and Jan Helsen",
note = "Publisher Copyright: {\textcopyright} Published under licence by IOP Publishing Ltd.",
year = "2025",
month = jun,
day = "12",
doi = "10.1088/1742-6596/3016/1/012021",
language = "English",
volume = "3016",
journal = "Journal of Physics: Conference Series",
issn = "1742-6596",
publisher = "IOP Publishing Ltd.",
number = "1",
}