Investigating data-driven surrogates for the Ainslie wake model

Abstract

Abstract Efficient and accurate calculation of wind turbine wakes are central to the design and analysis of wind farms. We present the promising results of an investigation into machine learning techniques for creating computationally efficient surrogates of the Ainslie wake model, as a computationally accessible stepping stone to more complex models. We have compiled an extensive dataset comprising 8600 wake fields, and explored a diverse range of techniques. Among these, the most promising techniques are regression decision trees (RDTs) and Fourier-expanded multi-layer perceptrons (FEMLPs) which have proven to be very effective for image applications. Even with smaller training sets, acceptable results are achieved, offering a feasible balance between size and performance. Both RDTs and FEMLPs excel in interpolation, while FEMLPs particularly showcase robust extrapolation capabilities. Visual comparisons show that RDTs outperform FEMLPs in terms of the smoothness of the generated wake fields. All techniques create surrogates that are at least six orders of magnitude faster than the Ainslie model implementation. The investigation results in a flexible pair of turbine wake surrogate models which are extremely fast and achieve high performance with a reasonable training time. Their generic nature makes them promising candidates for creating effective surrogates for more complex wake models.