Data-driven modeling of wind farm wake flow based on multi-scale feature recognition

Accurate and efficient predictions of wind flow developments with wake effects accounted are crucial for wind farm layouts and power forecasting. Existing methods can be broadly classified as physical measurement, numerical simulations, physics-based modeling, and data-driven modeling. The first two is of high cost in terms of time and resources, the third suffers from low accuracy due to limited physics modeled, while the last one takes advantage of the large amount of high-quality data available and has become increasingly popular. This study proposes a rapid data-driven modeling method for wind farm wake flow, inspired by video frame interpolation and based on the principle of similarity, which utilizes a multi-scale feature recognition technique. The method transforms wind farm field data into images and predicts wake flow by identifying, matching, and interpolating features from a limited set of wake flow images using the Scale-Invariant Feature Transform (SIFT) and Dynamic Time Warping (DTW) approaches. To demonstrate the effectiveness of the proposed method, six representative cases were evaluated, encompassing mini wind farms with varying turbine spacings, different turbine sizes, combinations of spacing and size variations, different numbers of turbines, and various degrees of wind direction misalignment. A Mean Absolute Percentage Error (MAPE) ranging from 0.68% to 2.28% is achieved. Due to its ability to flexibly compute both 2D and 3D wake flow fields, the proposed method offers unique computational efficiency advantages over Large Eddy Simulation (LES) and Meteodyn WT in scenarios where two-dimensional wake flow fields are sufficient to meet industrial requirements. Therefore, this method can be employed for the extension of the wake flow database serving wind farm design, power prediction, etc., as an alternative to measurements, numerical simulation, and physics-based modeling, balancing efficiency and accuracy.