Average power losses due to wind turbine wakes are of the order of 10 to 20% of total power output in large offshore wind farms. Accurately quantifying power losses due to wakes is, therefore, an important part of overall wind farm economics. The focus of this research is to compare different types of models from computational fluid dynamics (CFD) to wind farm models in terms of how accurately they represent wake losses when compared with measurements from offshore wind farms. The ultimate objective is to improve modelling of flow for large wind farms in order to optimize wind farm layouts to reduce power losses due to wakes and loads.The research presented is part of the EC-funded UpWind project, which aims to radically improve wind turbine and wind farm models in order to continue to improve the costs of wind energy. Reducing wake losses, or even reduce uncertainties in predicting power losses from wakes, contributes to the overall goal of reduced costs. Here, we assess the state of the art in wake and flow modelling for offshore wind farms, the focus so far has been cases at the Horns Rev wind farm, which indicate that wind farm models require modification to reduce under-prediction of wake losses while CFD models typically over-predict wake losses. Further investigation is underway to determine the causes of these discrepancies.

There is an urgent need to develop and optimize tools for designing large wind farm arrays for deployment offshore. This research is focused on improving the understanding of, and modeling of, wind turbine wakes in order to make more accurate power output predictions for large offshore wind farms. Detailed data ensembles of power losses due to wakes at the large wind farms at Nysted and Horns Rev are presented and analyzed. Differences in turbine spacing (10.5 versus 7 rotor diameters) are not differentiable in wake-related power losses from the two wind farms. This is partly due to the high variability in the data despite careful data screening. A number of ensemble averages are simulated with a range of wind farm and computational fluid dynamics models and compared to observed wake losses. All models were able to capture wake width to some degree, and some models also captured the decrease of power output moving through the wind farm. Root-mean-square errors indicate a generally better model performance for higher wind speeds (10 rather than 6 m s−1) and for direct down the row flow than for oblique angles. Despite this progress, wake modeling of large wind farms is still subject to an unacceptably high degree of uncertainty.

No abstract

Previous research has revealed the need for a validation study that considers several wake quantities and code types so that decisions on the trade-off between accuracy and computational cost can be well informed and appropriate to the intended application. In addition to guiding code choice and setup, rigorous model validation exercises are needed to identify weaknesses and strengths of specific models and guide future improvements. Here, we consider 13 approaches to simulating wakes observed with a nacelle-mounted lidar at the Scaled Wind Technology Facility (SWiFT) under varying atmospheric conditions. We find that some of the main challenges in wind turbine wake modeling are related to simulating the inflow. In the neutral benchmark, model performance tracked as expected with model fidelity, with large-eddy simulations performing the best. In the more challenging stable case, steady-state Reynolds-averaged Navier-Stokes simulations were found to outperform other model alternatives because they provide the ability to more easily prescribe noncanonical inflows and their low cost allows for simulations to be repeated as needed. Dynamic measurements were only available for the unstable benchmark at a single downstream distance. These dynamic analyses revealed that differences in the performance of time-stepping models come largely from differences in wake meandering. This highlights the need for more validation exercises that take into account wake dynamics and are able to identify where these differences come from: mesh setup, inflow, turbulence models, or wake-meandering parameterizations. In addition to model validation findings, we summarize lessons learned and provide recommendations for future benchmark exercises.

No abstract