International Journal of Advanced Research in Computer and Communication Engineering

Abstract: This article proposed a spatial-temporal deep learning architecture for network-wide flight delay prediction that operates within the confines of federated learning to ensure data privacy is respected among various aviation stakeholders. Due to regulatory, commercial and privacy limitations, it is not possible to use traditional centralized delay prediction models because airline operators, airports and air traffic control authorities have access to sensitive operational, passenger and meteorological data which cannot be aggregated in a centralised form. In order to tackle this fundamental problem, we propose the Hybrid Federated Delay Learning Network (HFDL-Net), which employs a spatio-temporal graph neural network with gated recurrent units (GRU) at each client node, in conjunction with an hierarchical federated aggregation approach at the central server to learn together from distributed datasets without direct data sharing. Using our architecture, we model the air transportation network as nodes with edges (aeroportunities) and routes with convolution (temporal evolution) via graph modeling and recurrent layers to capture spatial delay propagation patterns. Real-world experiments on multi-airport flight datasets spanning three major hub networks demonstrate that HFDL-Net achieves mean absolute error (MAE) improvements of 12–15% over non-federated baseline models while maintaining prediction accuracy within 3% of fully decentralized training approaches. In addition, the use of a hierarchical aggregation reduces communication overhead by 40% when compared to traditional FedAvg implementations through adaptive client selection and gradient compression techniques. The suggested scheme effectively manages non-IID data distributions among multiple clients, exhibits resilience to client dropout scenarios, and adapts well to airport networks spanning more than 100 participants. Additionally, This evidence supports federated spatial-temporal modeling as a practical, scaleable and privacy-preserving approach for network-wide flight delay prediction in real-life aviation scenarios where data sovereignty and regulatory compliance are critical requirements.

Keywords: Flight Delay Prediction, Spatial-Temporal Modeling, Federated Learning, Graph Neural Networks.

Downloads: | DOI: 10.17148/IJARCCE.2026.15227