Lseq2seq: A new reduced-order model for unsteady aerodynamic force identification

Abstract

Identifying unsteady aerodynamic forces is a crucial and challenging task in aerodynamics. It is also a critical research foundation for other subjects such as aeroelasticity, aircraft design, and flight dynamics. The two mainstream methods used to identify unsteady aerodynamic forces are Computational Fluid Dynamics (CFD) and experiments. However, these methods have their limitations, such as lengthy computational expense and high resource consumption. This article proposes a new reduced-order model called Long Sequence to Sequence (Lseq2seq) based on deep sequence generation models to predict unsteady aerodynamic forces in an efficient way. The Lseq2seq model is then applied to determine the hysteresis loop for the NACA0012 airfoil and the unsteady aerodynamic force of the two-freedom oscillation of the NACA64A010 airfoil in transonic flow. The results are compared with other prevalent time-sequential networks, such as Sequence to Sequence (Seq2seq) and Gated Recurrent Unit (GRU). The proposed Lseq2seq model presents better precision and generalization ability for identification. Additionally, this article explores a combined predictor–corrector method called GRU-Lseq2seq to predict the flutter response of the NACA64A010 airfoil, and the results demonstrate that the combined model could achieve better prediction accuracy than the GRU model and could be used in flutter boundary prediction.