Abstract:

Flight trajectory prediction plays a crucial role in ensuring safe and efficient air traffic operation. In order to consider the implicit correlations between flight trajectory characteristics, the encoding and decoding operations of the prediction framework in the transformer model were enhanced. Firstly, the convolutional block was improved, and ordinary convolutions were applied to capture the correlations between neighboring time series trajectory characteristics, and dilated convolutions were added to capture correlations between non-neighboring time series trajectory characteristics. Secondly, multi-head self-attention was employed to perform calculation based on the spatiotemporal features of the flight trajectory combined with the importance of attention scores. Thirdly, probabilistic sparse method was designed to reduce the computational complexity of the multi-head self-attention and improve the model's computational efficiency. Finally, an experimental platform was established to verify the flight trajectory prediction framework. The results show that compared to the traditional transformer model and the other three neural network models, the improved transformer model shows a 14.4% improvement in time performance. By using root mean square error(RMSE) and mean absolute error(MAE) as evaluation metrics, the average prediction deviations of the improved transformer model for trajectory features such as longitude, latitude, and altitude are 0.027 and 0.021, respectively. These deviations are reduced by 0.072 and 0.063 compared to the traditional transformer model's average prediction deviations of 0.099 and 0.084. Sensitivity analysis on the lengths of prediction sequences indicates that the improved transformer model is more stable than the baseline models.

Key words: improved Transformer model, 4D flight trajectory, trajectory prediction, deep learning, dilated convolution, attention mechanism