Hybrid methods for wind farm simulation and control
- Wind farm simulation and control involve a trade-off between accuracy and computational speed. I explore how hybrid methods combine the best attributes of each constituent method to achieve both accuracy and speed. First, I consider high-fidelity simulation of a wind farm. These simulations inform design and verification of farm layout optimization and control strategies and models. Typically simulations represent only a subset of the relevant physics, which includes the atmospheric boundary layer (ABL), blade near-wall turbulence, and turbine-turbine interactions. This hinders optimization and control development. I modify a hybrid between two fluid models, one Reynolds- Averaged Navier Stokes (RANS) and one large eddy simulation (LES), that is well suited to capture all three effects. I focus on modifying first the RANS closure, SST k-omega, and then the hybrid approach, Active Model Split (AMS), to accurately simulate the ABL, including the Coriolis effect and buoyancy. I show that the modified SST k-omega provides comparable accuracy to the RANS model more commonly used for this problem, k-epsilon. AMS improves upon SST k-omega, achieving accuracy comparable to that of LES. I then extend these results to a diurnal cycle with buoyancy. The second part of the work focuses on wind farm control, specifically wake steering. Wake steering yaws upstream wind turbines to deflect their wakes from downstream turbines, increasing the total power produced by the wind farm. Typically wake steering doesn't adapt to turbine status, which leads to lost power when turbines are inactive. I develop a hybrid between model- and learning- based control, differentiable control (DC), for adaptive wake steering. This includes implementing a control-oriented wind farm model in a learning package so that it is automatically differentiable and compatible with learning tools. It also requires identifying a neural network architecture which can learn the discontinuous functions that arise in wake steering. I show that DC accomplishes adaptive wake steering with comparable accuracy to the most common wake steering method, which is purely model-based, while reducing the offline run-time by over 90%. Furthermore, DC handles not only varying wind speed but also varying wind direction, an extension which is challenging for purely learning-based methods.
|Type of resource
|electronic resource; remote; computer; online resource
|1 online resource.
|Adcock, Christiane Marie Otten
|Degree committee member
|Degree committee member
|Stanford University, School of Engineering
|Stanford University, Institute for Computational and Mathematical Engineering
|Statement of responsibility
|Christiane Marie Otten Adcock.
|Submitted to the Institute for Computational and Mathematical Engineering.
|Thesis Ph.D. Stanford University 2023.
- © 2023 by Christiane Marie Otten Adcock
- This work is licensed under a Creative Commons Attribution Non Commercial 3.0 Unported license (CC BY-NC).
Also listed in
Loading usage metrics...