03.4 - Applied Aerodynamics
FLIGHT-TEST DATA-DRIVEN HIGH-ANGLE-OF-ATTACK DYNAMICS MODELING VIA LSTM-POLYNOMIAL HYBRID APPROACH
Z. Zhang¹, H. Zhao¹, W. Zhang¹, J. He¹, Z. Gao¹; ¹Northwestern Polytechnical University, China
To address the significant accuracy degradation of traditional aerodynamic models in the stall/near-stall regions due to strong nonlinearities and unsteady effects, this paper proposes a hybrid modeling method combining a polynomial model and a Long Short-Term Memory (LSTM) network. Based on rigid-body dynamic equations, flight-test data were filtered, and aerodynamic coefficients were inversely identified to construct a high-quality dataset for modeling. The proposed framework employs a polynomial model with explicit physical meaning as the backbone to capture the main nonlinear and quasi-static trends of aerodynamic forces. Concurrently, an LSTM network is cascaded as a residual corrector to compensate for hysteresis effects and higher-order unsteady characteristics that are difficult for traditional models to characterize, by learning from historical state sequences. This architecture is designed to structurally balance the physical foundation of polynomial models with the high-accuracy advantage of data-driven models. Validation was performed by integrating the identified model into a six-degree-of-freedom (6-DoF) flight simulation environment. Results demonstrate that the hybrid model achieves weighted relative prediction errors of 0.35% for angle of attack, 0.47% for normal load factor, and 2.65% for pitch rate during stall maneuvers. This work provides an effective technical pathway for high- fidelity unsteady aerodynamic modeling, combining physical interpretability with data-driven precision.
DOI: 10.71945/icas2026_0269