Prediction and Evaluation Methods for Large-Scale Flight Delay Propagation Based on Spatiotemporal Network

The highly coupled nature of flight operations in spatial and temporal dimensions leads to the widespread propagation of large-scale delays across multiple airports. This issue is addressed through using dynamic network analysis to explore the patterns of air traffic delay propagation. To accurately capture the dynamics of delay spread, a spatiotemporal network is constructed where airports are nodes, departing flights are edges, and a temporal resolu-tion is 5 min. In constructing the flight delay spatiotemporal graph, edge weights are improved. Initially estimated using statistical averages of delay times or simple empirical rules, the weights are predicted via deep learning. For the flight delay prediction task, a multi-task NR-DenseNet model is employed to simultaneously predict flight delay duration (regression) and delay occurrence (classification), enhancing the accuracy and timeliness of the weights. By comparing performances across different network depths, experiments demonstrated that a 16-layer NR-DenseNet achieved optimal performance in both tasks. The regression prediction yielding a mean squared error (MSE) of 58.30 and a mean absolute error (MAE) of 3.28, while classification accuracy reached 94.8%. Regarding metric evaluation, single metrics are found to be insufficient for fully assessing the complexity of air traffic delay propagation. Therefore, three evaluation metrics, intensity, propagation rate, and velocity, are established to quantita-tively analyze the multidimensional characteristics of delay spread. Using domestic data from the East China Air Traffic Management Bureau as the study, the results indicated that the proposed method effectively reveals the spa-tiotemporal details of large-scale delay propagation within the scheduled flight timetable.