A study on trajectory forecasting approach optimized by PSO-based TCN-attention mechanism

Abstract

In recent years, the civil aviation transportation volume has experienced rapid growth, and airspace resources have become increasingly scarce. To solve the problems of insufficient prediction accuracy of aircraft four-dimensional (4D) trajectories in complex airspace and the difficulty in global hyperparameter optimization of deep learning models, this research constructs a hybrid PSO-TCN-Attention model enhanced by the particle swarm optimization (PSO) algorithm. First, data cleaning operations are performed on raw flight trajectory data to generate standardized data samples; then, Temporal Convolutional Networks (TCN) are applied to capture long-distance temporal correlations in trajectory sequences through dilated convolutions and residual connections; in addition, a multihead self-attention mechanism is integrated to improve the ability of representational learning for key flight movement patterns; at the same time, a PSO method with dynamically adjustable inertia weights and learning factors is adopted to conduct global optimization on the key hyperparameters of the model, thereby achieving a better balance between search efficiency and convergence accuracy. Test results based on real trajectory datasets show that the proposed model outperforms traditional benchmark models such as TCN, CNN, and LSTM in terms of Mean Absolute Error (MAE) and Root Mean Square Error (RMSE), with high adaptability and effectiveness.