A novel neural network method of aerodynamic performance prediction for wind turbine airfoils under glaze ice conditions

Ice accretion on the blade surface, particularly the glaze ice with protruded horns, causes a severe decrease in the wind power utilization. The characteristics research on the airfoil under glaze ice conditions is the basis of exploring the anti-icing method in view of the important role of an airfoil in the aerodynamic shape of the blade. The protruded horn shape of glaze ice complicates the meshing process of airfoils and increases the number of grids, whereas more grid cells make the amount of aerodynamic computation huge. In the present study, a novel neural network method is developed for the prediction of the aerodynamic performance of wind turbine airfoils with glaze ice. The biases of nodes, connection weights between nodes, and parameters of link switches are chosen as the design variables. The social learning is adopted to modify the potential well center update mode of non-optimal particles, and the Lévy flight is combined with greedy algorithm to identify the position of optimal particles for Quantum Particle Swarm Optimization (QPSO) algorithm. The testing error is introduced to interfere with the selection of global optimal particle, and the optimizer, based on the SLLQPSO (with SLLQPSO denoting improved QPSO) algorithm combined with Binary Particle Swarm Optimization (BPSO) algorithm, seeks the solutions minimizing the training error of neural network. A new neural network, namely SLLQPSO-BPNN, is established after being trained by Back Propagation (BP) algorithm, and predicts the lift and drag coefficients of S809 and NACA64618 airfoils with glaze ice. Significant performance improvements are achieved for QPSO algorithm and neural network, confirming that a novel neural network method with high accuracy is provided for the aerodynamic performance analysis of wind turbine airfoils.