Comparative Study of Surrogate Models for Aeroelastic Wind Turbine Blade

Abstract

This study compares four metamodeling techniques for approximating the aeroelastic response of a wind turbine blade subjected to one-way fluid-structure interaction (FSI). The evaluated methods include second-order polynomial response surfaces (RSM), Kriging, artificial neural networks (ANN), and genetic aggregation. An experimental design based on Latin hypercube sampling (LHS) was used to generate 27 training points for 5 input variables, supplemented by 5 independent points for validation. The results demonstrate that second-order polynomials offer the best generalization robustness for this problem, closely followed by Kriging. Genetic aggregation shows intermediate performance, while neural networks prove inadequate for capturing physical complexity in a low-sampling context. This study highlights the relevance of polynomial approaches and Kriging for aeroelastic optimization when computational cost limits the size of available data. These findings provide a practical framework for wind turbine designers to select cost-effective surrogate models, enabling rapid structural optimization and reliability assessment while significantly reducing the computational overhead of high-fidelity FSI simulations.