Joint prediction of multi-aircraft trajectories in terminal airspace: A Flight Pattern-Guided Social Long-Short Term Memory network

Abstract

Aircraft trajectory prediction is a critical foundation for tasks such as conflict detection and resolution, and monitoring of abnormal aircraft behavior, thus making it a key technology for the next generation of air traffic systems. Airport terminal areas, serving as the convergence points of the entire air transportation network, accommodate the highest density of aircraft within the aviation system. Among these densely packed aircraft, there are inherent and potential interactions that are invisible yet influential, posing a significant challenge for trajectory prediction in the terminal area. To address this issue, we propose a Flight Pattern-Guided Social Long-Short Term Memory (FPG-SLSTM) network to jointly predict multiple aircraft trajectories. Firstly, it introduces a flight pattern matching method based on classification concepts to assign a flight pattern to each trajectory. Following this, the flight intent trajectory is generated based on that flight pattern. Then, the loss function is improved by incorporating the flight intent trajectory as prior physical information, and the Social Long-Short Term Memory (Social LSTM) neural network is employed to model the trajectory prediction problem for multiple aircraft, with social pooling operations to integrate the mutual influences among aircraft. A real-world dataset was constructed to validate the proposed approach. Experimental results show that the method achieved a 14.5% improvement in horizontal prediction accuracy and a 29.5% improvement in height prediction accuracy on average for a 6-minute horizon, compared to Binary Encoding Representation for Flight Trajectory Prediction (FlightBERT). These findings highlight that the proposed framework outperforms other baseline models in terms of prediction accuracy, particularly in complex traffic environments during busy periods, demonstrating the model’s ability to further enhance both the precision and robustness of trajectory prediction.