Due to significant nonlinear and unsteady effects， it is difficult to accurately simulate the maneuver flight characteristics of aircraft at high angle of attack by existing wind tunnel experiments and numerical methods. To improve the accuracy of maneuver flight simulation at high angle of attack， the physical-model-embedding ensemble neural network was developed to accurately model the unsteady aerodynamics of aircraft at high angle of attack， and the aircraft motion equation were further coupled in time domain to realize maneuver flight simulation at high angle of attack. Taking a typical fighter as the research object， the open-loop broadband excitation， open-loop harmonic excitation and post-stall maneuver flight data of longitudinal flight at high angle of attack are utilized as sample data for aerodynamic modeling. Three types of aerodynamic models are constructed and compared， including the traditional dynamic derivative model， the black-box neural network model and the ensemble neural network model. Furthermore， the flight characteristics of coupled simulation are further compared， and the idea of using the flight simulation method to test the robustness of the aerodynamic model is proposed.Results show that the lift coefficient error of aerodynamic modeling of the physical-model-embedding ensemble neural network is 57% lower than that of the traditional dynamic derivative model， and the robustness and stability in the coupling process are better. The aircraft response error is 63% lower than that of the black-box neural network model， which proves the advantages and engineering potential of the proposed modeling framework for small-sample flight data identification.