Multi-fidelity deep learning for wake modeling of wind turbines

Wake prediction of wind turbines is one of the challenging problems in wind farms due to due to the complex unsteady nature of interactions of turbine wake with other wakes as well as atmospheric turbulence. The engineering wake model should be sufficiently accurate and computationally cheaper to be employed for tasks like wind farm layout optimization and wind farm controls. In this talk, we explore an application of a composite neural network framework to learn the wake model from large samples of low-fidelity data along with very few samples of high-fidelity data. The composite framework consists of a neural network to learn the low-fidelity data coupled with two neural networks to learn the linear and nonlinear correlation between low and high fidelity data. In particular, we train a composite neural network using data generated from hierarchies of physical models to predict the three-dimensional velocity field in turbine wakes. The prediction from the composite neural network matches well with the high-fidelity data compared to a neural network trained solely using the high-fidelity data. This works opens up possibilities for data-efficient construction of surrogate models for wake prediction that can be utilized to study the influence of wind speed, yaw angles, and layout configuration on wind farm power production.