Publication
Abstract
The growing complexity of air traffic management systems requires delay-prediction models that provide not only accurate point estimates but also reliable uncertainty information for operational decision support. Bayesian Neural Networks (BNNs) offer a natural framework for probabilistic prediction, but their use for arrival-delay and flight-time-deviation prediction remains limited compared with deterministic neural-network models and related trajectory-prediction applications. This paper evaluates BNNs for predicting arrival delay (AD) and flight time deviation (FTD) on the high-frequency SBRF–SBGR Brazilian domestic route using meteorological, demand, temporal, and operational information. The study compares loss-selected and coverage-selected BNNs with artificial neural networks, adds probabilistic baselines under random and chronological splits, and evaluates a temporal-adaptation experiment using 2023–2024 data with conformal residual calibration. The random-split results show that coverage-selected BNNs preserve point-error performance close to deterministic neural networks while providing substantially better interval coverage. The chronological split shows that temporal generalization is more difficult for the original BNN configuration, motivating the adapted BNN ensemble. Under the same 2023–2024 chronological protocol used by the probabilistic baselines, the temporally adapted BNN framework provides the best AD accuracy–sharpness balance and competitive FTD performance with substantially sharper intervals than the Deep Ensemble on this route. SHAP-based feature-importance analysis and threshold-risk examples further illustrate how calibrated predictive distributions can support interpretable tactical delay assessment. These results characterize a single-route case study and should not be generalized to other routes, airports, or the broader Brazilian air traffic network without further multi-route validation.
Index terms: air traffic management; Bayesian inference; delay prediction; machine learning; neural networks.
Reproducibility
The companion repository documents the paper-faithful stochastic experiment, including the 23-input executable feature set, seeded split/scaling procedure, BNN training logic, Monte Carlo evaluation sequence, exact route-dataset checksum, and verified reference metrics.
BibTeX
Please cite the published IEEE Access article rather than the repository when referring to the research results.