AAGNN: Adaptive Airport Graph Neural Network for flight sequence prediction

Flight sequence is essential for aviation's secure and orderly operation. Accurate Flight Sequence Prediction (FSP) gives air traffic controllers (ATCs) an overview of the overall air traffic situation during the pre-tactical phase, allowing them to modify flight schedules when airspace capacity is reduced. The internal structure of flight sequences is compact, and then the effective extraction of the spatiotemporal dependence features of flight sequences determines the accuracy of the prediction. The flight scheduling approach treats the FSP as an optimization problem with the aircraft already known, but when considering more airlines, the problem becomes complicated to solve. Causal inference methods for solving FSP have struggled to extract implicit interaction features between flight sequences. For the above bottlenecks, we construct flight sequences from the perspective of sequence contents and transform the FSP problem into a session-based recommendation task by proposing a novel Adaptive Airport Awareness Graph Neural Network (AAGNN). Firstly, we lead in all the sequences to form an airport network, and then we augment the graph topology of the flight sequences with network structure information. Secondly, we propose an airport-aware adaptive representation graph model to obtain the implicit interaction features between flight sequences at the same airport. Finally, we develop a hybrid embedded session representation incorporating recent and overall flight sequence preferences for session recommendation. The experimental results on two real-world datasets of flight sequences in China demonstrate that the proposed AAGNN approach outperforms the numerous baseline approaches. The case study demonstrates the approach's generalizability to various specific operational scenarios.