Aircraft wake vortex prediction method based on bayesian optimization with CNN-BiLSTM

This paper proposes a Bayesian optimization-based CNN-BiLSTM method is proposed for predicting aircraft wake vortex evolution under complex atmospheric conditions. Traditional wake vortex inversion models suffer from limited predictive accuracy in such environments, and single machine learning methods often fail to adequately capture the essential features of wake vortex evolution. In this work, we formulate wake vortex evolution as a multi-input multi-output time-series regression problem and develop a hybrid model that integrates convolutional neural networks (CNN) and bidirectional long short-term memory (BiLSTM) networks. Bayesian optimization is employed to adaptively tune the key hyperparameters of the model, and Monte Carlo dropout (MC Dropout) is introduced to achieve probabilistic wake vortex prediction and quantify predictive uncertainty. Experiments on an independent test set show that the proposed CNN-BiLSTM model consistently outperforms CNN-LSTM, BiLSTM, LSTM, CNN, and the Kelvin-Helmholtz (K-H) model, reducing the mean absolute error and root mean square error reaching 12%～16% and 17%～21%, respectively. Additional validation using DLR water tunnel experimental data demonstrates the good generalization capability of the proposed model. Furthermore, the 95% prediction intervals constructed based on MC Dropout can effectively characterize the uncertainty in the evolution of wake vortex circulation and height, with prediction interval coverage probabilities of 62.10% and 59.34% on the independent test set, respectively. These results indicate that the proposed method can effectively improve the accuracy of wake vortex evolution prediction and provide technical support for the formulation and implementation of dynamic wake separation procedures.