FlightDBN: A deep Bayesian network for flight trajectory prediction in takeoff and landing

Accurate trajectory prediction during aircraft takeoff and landing is essential for flight safety, air traffic management, and decision support. However, the departure and destination constraints and the diversity of control strategies of aircraft takeoff and landing phases result in strong state-sequence dependencies in control commands. Traditional model-based trajectory prediction methods struggle with these dependencies, resulting in modeling errors accumulating due to invalid prior assumptions. Deep learning methods can automatically learn predictive mappings from data, but often lack the physical constraints and interpretability provided by model-based knowledge and require large amounts of training data, resulting in poor performance in data-incomplete or complex motion scenes. To address these challenges, this paper proposes FlightDBN, a novel deep Bayesian prediction framework derived from the aircraft dynamics model. It integrates prior model knowledge with offline data through Bayesian inference within an encoder-decoder architecture, ensuring both interpretability and predictive accuracy. First, FlightDBN encodes the input trajectory as sufficient statistics of the historical state sequence to learn higher-dimensional implicit features. Then, during decoding, FlightDBN passes the trajectory encoding results at each prediction step, uses an attention mechanism to extract features that are significantly related to the current decoded state, and designs a recursive update memory mechanism to selectively inherit and update such features. Finally, based on this memory mechanism, FlightDBN performs state updates by predicting the control commands of the aircraft dynamics model and applies Bayesian inference to compensate for modeling errors and quantify prediction uncertainties, thereby refining trajectory predictions. Therefore, FlightDBN provides both point prediction and uncertainty estimation for future trajectories, which is a practical advantage over deterministic prediction methods. Experiments on two real-world flight trajectory datasets demonstrate that FlightDBN outperforms baseline methods in prediction accuracy, covariance estimation, and robustness to data scarcity across different scenarios and conditions.