Wind turbine wakes: from numerical modeling to machine learning
Date
2018
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
University of Delaware
Abstract
Two main topics are studied in this research. First, the importance of compressibility effects of large horizontal-axis wind turbines are systematically assessed using the Blade Element Moment (BEM) method and unsteady Reynolds-Averaged Navier-Stokes (RANS) simulations. Second, a deep neural network (NN) with transfer learning ability are proposed for efficient wind farm power estimation. ☐ The tips of large horizontal-axis wind turbines can easily reach high speeds, thus raising the concern that compressibility effects may influence turbine wakes and ultimately power production. All past studies have assumed that these effects are negligible. In Chapter \ref{chap2}, compressibility effects are assessed in terms of blade aerodynamic properties and variable density separately. Using the BEM method, we find that under normal operating conditions (i.e., wind speed $<\sim$15 m s$^{-1}$ and tip speed ratio TSR $<\sim12$) aerodynamic corrections to the lift and drag coefficients of the blades have a minimal impact, thus the incompressible coefficients are adequate. In Chapter \ref{chap3}, compressibility effects are assessed in terms of variable-density, numerical simulations of a single turbine and two aligned turbines, modeled via the actuator line model with the default aerodynamic coefficients, are conducted using both the traditional incompressible and a compressible framework. The flow field around the single turbine and its power performance are affected by compressibility and both show a strong dependency on TSR. Wind speed and turbulent kinetic energy (TKE) differences between compressible and incompressible results origin from the rotor tip region but then impact the entire wind turbine wake. Power production is lower by 8\% under normal operating conditions (TSR$\sim$8) and 20\% lower for TSR$\sim$12 due to compressibility effects. When a second turbine is added, the front turbine experiences similar effects as the single-turbine case, but TKE differences are enhanced while wind speed differences are reduced after the second turbine in the overlapping wakes. These findings suggest that compressibility effects play a more important role than previously thought on power production and, due to the acceptable additional computational cost of the compressible simulations, should be taken into account in future wind farm studies. ☐ In Chapter \ref{chap4}, a deep neural network is trained and validated using three years of one-minute observations of wind speed, direction, and power generated at the offshore Lillgrund wind farm (Sweden). In its traditional form, the NN is used to generate a new two-dimensional power curve, which predicts with high accuracy (error $\sim2\%$) the power of the entire Lillgrund wind farm based on wind speed and direction. By contrast, manufacturers only provide one-dimensional power curves (i.e., power as a function of wind speed) for a single turbine. The second innovative application is the use of a geometric model (GM) to calculate two simple geometric properties to replace wind direction in the NN. The resulting GM-trained NN has the powerful feature of being applicable to any wind farm, not just Lillgrund. A validation at the onshore N{\o}rrek{\ae}r wind farm in Denmark demonstrates the high accuracy (error $\sim6\%$) and transfer-learning ability of the GM-trained NN.
Description
Keywords
Earth sciences, Blade element, CFD, Compressibility effects, Neural network, Wake