PhyWakeNet: a dynamic wake model accounting for aerodynamic force oscillations

Abstract

Advanced wind energy technologies require predictions of the dynamic behaviour of wind turbine wakes. In this work, we present a dynamic wind turbine model, PhyWakeNet, a physics-integrated generative adversarial network-convolutional neural network (GAN-CNN) model for wind turbines under aerodynamic force oscillations. The model combines three interconnected submodels for the time-averaged wake, wake meandering, and small-scale wake turbulence. The time-averaged wake model derives from mass and momentum conservation based on the concept of momentum entrainment, which is computed based on the wake meandering and small-scale wake turbulence models. The wake meandering is captured through conditional GAN-reconstructed spatial modes and a neural-network-enhanced dynamic system for temporal evolution, while the small-scale wake turbulence is generated via a CNN based on the time-averaged wake, wake meandering, and inflow turbulence. The test cases show that the PhyWakeNet model accurately predicts the wake statistics, with the error of the time-averaged velocity deficits, the variance of the streamwise velocity fluctuations, and the wake meandering amplitude to be less than 1 %, 10 %, and 15 %, respectively. Moreover, the model also accurately captures the large-scale temporal variations of instantaneous wake centres and velocity deficits, enabling applications in wake management to mitigate aerodynamic loads and power fluctuations in wind farms.