A novel approach to inverse design of wind turbine airfoils using tandem neural networks

Abstract

The performance of a wind turbine and its efficiency majorly depends on wind-to-rotor efficiency. The aerodynamic design of the wind turbine blades using high-fidelity tools such as adjoint-computational fluid dynamics (CFD) is accurate but computationally expensive. It becomes impractical when the number of design variables increases for multidisciplinary optimization (MDO). Low-fidelity tools are computationally cheaper but are not accurate, especially in regions of adverse pressure gradient and reverse flows. Surrogate modeling has been used in many aerodynamic problems. We develop and apply a recent architecture of the deep learning module, tandem neural networks (T-NNs) for the inverse design of wind turbine airfoils. The T-NNs trained on CFD data for fully turbulent cases predict not only the performance parameters for the given airfoil geometry but also the airfoil geometry for a given design objective. This framework uses the entire performance polar for inverse design which ensures that the airfoil optimization is not a single-point optimization problem which is essential for practical design problems. The T-NNs are also optimized to include multiple constraints like maximum thickness and trailing edge (TE) thickness which is a novel contribution in the field of inverse design using surrogate models. A statistical analysis is also performed to predict a family of airfoil geometries.