Dynamic stall modeling of wind turbine blade sections based on a data-knowledge fusion method.

Dynamic stall often leads to unsteady load and performance degradation in horizontal axis wind turbines. Therefore, accurate modeling of dynamic stall is crucial. However, due to the large variations of the blade aerodynamic profiles and the complexity of dynamic stall flow, numerical simulation and wind tunnel experiment are costly. On the other hand, widely used semi-empirical models have limited accuracy. Hence, this paper proposes a data-knowledge fusion method that incorporates physical knowledge into a neural network to improve its accuracy and generalization ability. Firstly, the force components of the Leishman–Beddoes model are incorporated into the network. An efficient dynamic stall model for the S809 airfoil is thus established with a small amount of high-precision experimental data. It achieves extrapolation predictions of reduced frequency and angle of attack with only 1/5 of the samples in the database to train. Moreover, to make full use of the accumulated existing airfoil data to assist in modeling other airfoils, the obtained S809 model is fused in the network to predict the aerodynamics of S810 and S814. The average relative error of the prediction cases is nearly 10%. Comprehensively, this paper provides a new paradigm for assessing the dynamic stall of the wind turbine blade. [Display omitted] • Develop a dynamic stall modeling neural network by integrating knowledge. • Create an efficient model that incorporates the force components of the L-B model. • Utilize accumulated aerodynamic data to model the other cross-sections of the blade. • The data fusion strategy reveals the importance of interpretable basis functions. [ABSTRACT FROM AUTHOR]