A Hybrid Probabilistic Spatio-Temporal Model for Flight Delay Exceedance Prediction
Faculty Mentor
Yao Xu
Location
Russell Union Ballroom
Type of Research
On-going
Session Format
Poster Presentation
College
Allen E. Paulson College of Engineering & Computing
Department
School of Computing
Abstract
Accurate flight delay forecasting is critical for improving airport operations, airline efficiency, and passenger reliability. However, prediction remains challenging due to complex interactions among weather conditions, congestion, aircraft rotations, and network-wide delay propagation. This study proposes a hybrid probabilistic spatio-temporal (HPST) framework that integrates gradient-boosted tabular learning with recurrent modeling of recent system dynamics. The approach adopts a two-stage strategy: a calibrated classifier first estimates the probability that arrival delay exceeds 15 minutes, followed by a conditional regression model that predicts the magnitude of exceedance. All experiments are conducted under a strict chronological protocol, using training data from January to May 2024 and evaluating on an unseen June hold-out set to preserve operational realism. The architecture combines rich static descriptors, medium-term historical behavior summarized over the preceding week, and short-term congestion patterns observed within the previous 24 hours. A boosting-based prior anchors baseline predictions, while gated recurrent units adapt to evolving temporal conditions. Empirical results demonstrate consistent improvements over standalone LightGBM and neural baselines, reducing mean absolute error from over 12 minutes to approximately 7 minutes and improving robustness to extreme delays. Overall, the findings suggest that coupling probabilistic classification with spatio-temporal regression provides a practical and scalable solution for deployment-ready flight delay exceedance prediction across large air transportation networks.
Program Description
.
Start Date
4-23-2026 2:00 PM
End Date
4-23-2026 4:00 PM
Recommended Citation
Afrane, Mary Dufie; Xu, Yao; and Li, Lixin, "A Hybrid Probabilistic Spatio-Temporal Model for Flight Delay Exceedance Prediction" (2026). GS4 Student Scholars Symposium. 171.
https://digitalcommons.georgiasouthern.edu/research_symposium/2026/2026/171
A Hybrid Probabilistic Spatio-Temporal Model for Flight Delay Exceedance Prediction
Russell Union Ballroom
Accurate flight delay forecasting is critical for improving airport operations, airline efficiency, and passenger reliability. However, prediction remains challenging due to complex interactions among weather conditions, congestion, aircraft rotations, and network-wide delay propagation. This study proposes a hybrid probabilistic spatio-temporal (HPST) framework that integrates gradient-boosted tabular learning with recurrent modeling of recent system dynamics. The approach adopts a two-stage strategy: a calibrated classifier first estimates the probability that arrival delay exceeds 15 minutes, followed by a conditional regression model that predicts the magnitude of exceedance. All experiments are conducted under a strict chronological protocol, using training data from January to May 2024 and evaluating on an unseen June hold-out set to preserve operational realism. The architecture combines rich static descriptors, medium-term historical behavior summarized over the preceding week, and short-term congestion patterns observed within the previous 24 hours. A boosting-based prior anchors baseline predictions, while gated recurrent units adapt to evolving temporal conditions. Empirical results demonstrate consistent improvements over standalone LightGBM and neural baselines, reducing mean absolute error from over 12 minutes to approximately 7 minutes and improving robustness to extreme delays. Overall, the findings suggest that coupling probabilistic classification with spatio-temporal regression provides a practical and scalable solution for deployment-ready flight delay exceedance prediction across large air transportation networks.
