PhyWakeNet: A Physics-Integrated Machine Learning Model for Spatiotemporal Wind Turbine Wake Predictions

Advanced wind energy technologies require predictions of wind turbine wakes' transient behaviour. The turbulent nature of these wakes makes them extremely difficult to model - let alone predict.In this work, we present a PhyWakeNet model, a physics-integrated generative adversarial network-convolutional neural network (GAN-CNN) model for broad-spectrum wind turbine wake predictions. The model combines three interconnected submodels for time-averaged wake components, and coherent and incoherent turbulence fluctuations. The time-averaged wake flow model derives from mass and momentum conservation with entrainments across the wake boundary computed based on the coherent and incoherent wake flow models. The coherent wake flow is captured through conditional GAN-reconstructed spatial modes and neural network-enhanced dynamic system for temporal evolution, while the incoherent wake flow is generated via a CNN based on the time-averaged wake, coherent wake flow, and upstream measurements. Validation under active wake control demonstrates the model's capability to predict forcing frequency-dependent wake responses, flow recovery, and turbulence kinetic energy. The model accurately captures broadband wake fluctuations and temporal variations of key characteristics like instantaneous wake center and velocity deficit, enabling potential applications in wake management to mitigate aerodynamic loads and power fluctuations in wind farms.