A three-dimensional dynamic wake prediction framework for multiple turbine operating states based on diffusion model

Abstract

The modeling of wind turbine wakes is critical for turbine control, layout optimization, and power prediction, yet achieving both high accuracy and efficient computation remains a challenge. This study proposes a machine learning (ML)-based three-dimensional dynamic wake prediction framework consisting of a freestream field generator, a diffusion model, and an analytical wake model. The framework employs an iteration-independent prediction method to reconstruct wake fields directly from inflow data and turbine states, making prediction errors independent of the time-marching prediction iterations. The framework seamlessly integrates a diffusion model for enhanced prediction of transient wake characteristics, and an analytical model ensuring adaptability to various turbine operating strategies. The performance of the proposed framework was evaluated under various turbine operating strategies, including greedy, wake-steering, and partially-operating. With an 8476$\times$ speedup over Large Eddy Simulation (LES), the framework delivers high-accuracy predictions, showing 3.9% transient and 0.7% time-averaged errors relative to the average freestream velocity. Additionally, the rotor-effective speed derived from the predicted wake fields aligns closely with simulation-derived results, confirming the framework’s accuracy. To the best of our knowledge, this work presents the first ML-based framework capable of 3-D dynamic wake prediction, offering an accurate and efficient solution for wind turbine wake modeling.