Effect of wake model uncertainties on power output estimates for wind farms with wake steering control

Abstract

Because of technical and economical constraints that limit the space where turbines can be sited, wind farms experience significant power losses due to aerodynamic interactions between their wind turbines. In recent years, yaw-based wake steering control, in which upstream turbines are yawed to deflect the wakes away from downstream turbines, has been established as a promising wind farm control method to reduce wake losses. This thesis investigates the feasibility of training a generalised datadriven model capable of predicting the power of a Vestas V112-3.3 MW regardless of the size, layout and wake steering strategy of the wind farm in which it is located. The data to train the data-driven model is generated with a low-fidelity engineering wake model for a variety of designed wind farm layouts, and its accuracy is validated with SCADA data of the wind farm of Mont Heudelan. The engineered features to parameterize a layout and a wake steering strategy are described in detail, and the transfer-learning ability of the model is demonstrated by testing it with data of Mont Heudelan. The Mont Heudelan specific surrogate model performs with a MAE of 0.44%, a RMSE of 0.95%, and a R2 of 99.95%, whereas the generalised model achieves a MAE of 0.69%, a RMSE of 1.31%, and a R2 of 99.88%, hence leading to the conclusion that model generalisation for the mentioned conditions is feasible. Then, a surrogate-based wind farm power optimization with wake steering control for Mont Heudelan is carried out. It is estimated that a promising 1.4% Annual Energy Production gain could be obtained. Conclusions regarding the uncertainty of the calculated 1.4% gain are drawn by carrying out uncertainty propagation of relevant wake model parameters and inherent model and procedure uncertainties. It is identified that the highest source of uncertainty is due to the binning method used to discretize the wind speed, wind direction, and turbulence intensity resulting in a lower expected AEP gain of 1.35% and a standard deviation of 0.16%. Finally, appropriate recommendations for future research work are offered based on the drawn conclusions and on the limitations to which the project has been subjected.