Temporal graph neural network for flight delay predictions
Χρονικό γραφικό νευρωνικό δίκτυο για την πρόβλεψη καθυστερήσεων σε πτήσεις

Master Thesis
Author
Giogakis, Dimitrios
Γιογάκης, Δημήτριος
Date
2025-12View/ Open
Keywords
Temporal graph neural network ; Delays predictionsAbstract
Flight delays significantly impact the efficiency and reliability of air transportation network,
making accurate delay prediction an essential task. The below Master thesis introduces a data
driven approach for predicting flight arrival delays using a Temporal Graph Neural Network
(TGNN). The air transportation network is modeled as a dynamic graph, where airports act as
nodes and flights as temporal edges that evolve over time. The TGNN framework is designed to
capture both the spatial relationships between airports and the temporal patterns that influence
flight performance. By processing sequential snapshots of the network, the model can learn how
operational conditions, historical data, and scheduled flight characteristics evolve and affect future
delays. The TGNN incorporates embeddings for node and edge features and applies temporal
message passing to extract meaningful temporal representations without relying on future or
unavailable information. The proposed approach is trained as a temporal edge regression task,
using only the data available at prediction time, such as scheduled departure and arrival times,
historical performance statistics, and temporal context. Evaluation on the most recent temporal
segments is conducted using MAE, MSE, and R² metrics to assess prediction accuracy and model
robustness. Results show that the TGNN captures essential dependencies within the evolving
transportation network and delivers reliable delay predictions. The findings highlight the potential
of temporal graph-based learning methods to enhance operational planning and decision-making
processes in large-scale air transport systems.


