Presentations

Abstract

FOR WIND FARM DEVELOPMENTS WITHIN LARGER WIND FARM CLUSTERS OF TOMORROW WE EXPECT GREATER WAKE AND BLOCKAGE IMPACTS. THE ASSOCIATED UNCERTAINTIES INCREASE THE COST OF FINANCE, AND MAY BREAK THE BUSINESS CASE FOR SPECIFIC PROJECTS. It is well known that atmospheric stability and the site-specific boundary layer significantly impact wake and blockage losses, but these factors are not commonly included in established wake models. To model the impact of these drivers DNV's RANS CFD model has been upgraded to consider site specific boundary layers predicted by the Weather Research and Forecasting (WRF) model. With these considerations, WRF-RANS-CFD has been shown to accurately predict wind farm flows across large offshore wind farm clusters. [1][2] However, RANS CFD is computationally expensive, requiring high performance computing, so DNV have developed CFD.ML as a solution. A Graph Neural Network (GNN) is trained to emulate the RANS CFD predictions of turbine interaction loss factors, producing a model fast enough to use in a layout optimisation process [3]. CFD.ML has previously been shown to model blockage related patterns of production well. However, long range wake impacts were underpredicted when using a GNN trained on a neutral-only offshore boundary layer. [4] METHOD We train a new "multi-stability" GNN to emulate all DNV's WRF-RANS-CFD turbine interaction loss factor predictions, considering 7 new input parameters that describe the atmospheric boundary layer, including turbulence, shear, temperature, boundary layer height, and the Coriolis effect. At validation sites entirely outside of the training and test data, we demonstrate CFD.ML's skill at replicating RANS predictions, including a large offshore wind farm cluster. The Multi-stability CFD.ML model is then tested against SCADA power measurements for skill at predicting long-range wakes and blockage impacts. Multi-stability CFD.ML predictions are made for discrete atmospheric boundary layers inputs, and compared to corresponding to atmospheric stability class binned SCADA power results. RESULTS We show how average SCADA derived power production patterns on the front row of an offshore wind farm within the wake of an upstream farm have been found to fall within the envelope of Multi-Stability CFD.ML predictions for unstable and stable offshore boundary layers. In this study we will investigate further how best to use the WRF time series output to define a time varying boundary layer profile and test how different modelling schemes (e.g. time series vs frequency domain) perform in comparison to the SCADA power measurements. The variation with atmospheric stability of the CFD.ML predicted pattern of production due to blockage will also be investigated. CONCLUSIONS AND DISCUSSION The upgraded "multi-stability CFD.ML" model considering atmospheric boundary inputs can be used to better inform wake and blockage loss factor estimates for large wind farms. We discuss how we may determine the extra inputs that describe the atmospheric boundary layer, from WRF or otherwise, and the limitations of the CFD.ML model. [1] Christiane Montavon, 2023, WESC, Glasgow [https://zenodo.org/records/8000511/files/CMontavon_et_al_Blockage_Cluster_interactions_from_dual_scanning_lidars_DNV_EnBW_WESC2023.pdf?download=1] [2] Elizabeth Traiger, 2023, ACP R&T, Austin [https://acp2023rt.eventscribe.net/posters/posterWall.asp] [3] James Bleeg 2020 J.Phys.: Conf.Ser. 1618 062054 [https://iopscience.iop.org/article/10.1088/1742-6596/1618/6/062054] [4] Karol Mitraszewski, WindEurope technical workshop, Lyon [https://myworkspace.dnv.com/download/public/renewables/windfarmer/docs/AI%20for%20turbine%20interactions%20-%20WindEuropeTech%202023_ver30.05.pdf]