【摘要】 飞行器的空气动力学参数具有强非线性和非定常特性，利用人工智能方法建模可以避开复杂的空气动力学机制，不仅可以降低技术门槛，还可以提高建模效率。因此，本文提出一种基于量子粒子群优化-长短时记忆（QPSO-LSTM）神经网络的非定常气动力建模方法。以NACA0012翼型俯仰运动的非定常气动特性为研究对象，通过翼型的飞行状态参数预测翼型在运动过程中所受到的气动力。在建模的过程中采用LSTM神经网络为基础模型，然后利用QPSO算法优化LSTM神经网络的超参数，如层神经元个数、历史数据长度和训练批次大小。研究结果表明，QPSO算法能较好地搜索LSTM神经网络超参数的全局最优解；QPSO-LSTM模型相比常规循环神经网络（RNN）和LSTM模型，在内插和外插预测气动力系数时具有更高的精度和更好的泛化能力，该方法可被用于航空航天领域的非定常气动力预测。更多还原

【Abstract】 Aircraft aerodynamic parameters are highly nonlinear and exhibit significant unsteady characteristics,making traditional modeling approaches complex and technically demanding. Leveraging artificial intelligence(AI)methods can bypass these complexities, lower the technical barriers, and enhance modeling efficiency. This paper proposes a method of predicting unsteady aerodynamic force by using QPSO-LSTM neural network. The QPSOLSTM model is constructed by initially employing the LSTM algorithm as the base neural network model, followed by the application of QPSO algorithm to globally optimize the neural network hyperparameters. The hyperparameters include the number of neurons per network layer, historical length of training data, and batch size during the training process. To validate the effectiveness of the modeling approach, the neural network is trained by using aerodynamic force data obtained from numerical simulations of NACA0012 airfoil under various periodic pitching motion conditions.The results indicate that QPSO algorithm is a good choice for optimizing LSTM neural network hyperparameters,which could effectively search for the global optimal solution and avoid human factors to influence the results in setting the hyperparameters. The QPSO-LSTM neural network also demonstrates the capability of precisely predicting unsteady aerodynamic force coefficient across various flight conditions, solely relying on limited flight input parameter. This feature could make the modeling method to be conveniently deployed. Furthermore, compared to conventional RNN and LSTM models, the QPSO-LSTM model demonstrates superior accuracy and enhances generalization capabilities in predicting unsteady aerodynamic force coefficients in both interpolation and extrapolation scenarios. This approach holds significant potential for applications in unsteady aerodynamic force prediction within the aerospace sector.更多还原