Data-Based Prediction of Unsteady Aerodynamic Forces Induced by Free-Stream Turbulence

Accurate prediction of the aerodynamic forces induced by free-stream disturbance has been a challenge for flight safety. In this paper, a novel data-based approach to model the online unsteady and nonlinear response of aircraft, i.e., aerodynamic drag and lift coefficients from inflow disturbance, which can be measured in practical flights by Light Detection and Ranging (LiDAR), is established and tested. Numerical simulations with the NACA 0012 airfoil were performed to collect samples in order to train a neural network. Each sample consists of the time series of the inflow disturbance and the aerodynamic coefficients, both transformed to the Fourier space to reduce training cost and the degree of overfitting. The impacts of the number of samples and their distributions on the prediction were analyzed. Four inflow profiles with increasing complexity were tested. Various hyperparameters were first investigated, and it was found that the neural network trained with activation function TanH and optimized scaled conjugate gradient algorithm had the best performance. Finally, neural networks for both drag and lift coefficients were trained with a total of 1,300 randomly distributed samples, and a mean error of 2.1% was achieved in the test.