Parametric modeling and analysis of transonic aeroelastic systems based on incremental learning

Abstract

The aerodynamic reduced order model (ROM) is an effective tool for predicting unsteady aerodynamics and aeroelastic responses with high accuracy and low computational costs. However, traditional aerodynamic ROMs with respect to different flow conditions are lack of robustness. How to enhance dramatically the generalization capability of aerodynamic ROMs should be further studied. This paper presents three parametric ROMs based on the least squares support vector regression algorithm (LS-SVR) and two incremental learning algorithms. LS-SVR is used to establish the relationship between aerodynamic inputs and outputs based on training data from high-fidelity flow simulations. And the Mach number is considered as an additional system input to account for varying flow conditions. The main contribution of incremental learning algorithms based on LS-SVR is that it does not need to reconstruct the ROM when adding sample data. For incremental learning algorithm with forgetting mechanism, training samples are added in the manner of a time-window to improve training efficiency. To illustrate the performance of these ROMs, the aerodynamic and aeroelastic responses of a NACA 0012 airfoil with two degrees of freedom are investigated. The simulation results show that both unsteady aerodynamic results and aeroelastic responses computed by using the three ROM-based models agree well with the results of the high-fidelity simulation. The three proposed ROMs have better generalization capacity and modeling efficiency.