Airport flight delay prediction based on tree-structured multi-scale temporal and spatial correlation learning

Airport flight delay prediction is crucial for air traffic management and airline scheduling. However, existing delay prediction methods struggled to model the complex correlations of external-temporal-spatial factors in delay propagation due to their limited correlation learning capability, leading to insufficient delay prediction accuracy. To address this, this study proposes a Tree-structured Multi-scale Temporal and Spatial Correlation (TMTSC) learning method for airport flight delay prediction. It fuses historical delay and weather data embedding, and designs Temporal-Tree Net and Spatial-Tree Net to generate multi-scale receptive fields to learn the complex correlations of external-temporal-spatial factors. A feature decomposition and interactive learning mechanism is employed in the tree nodes, to deeply capture the temporal and spatial correlations in flight delay propagation and improve the accuracy of correlation learning. Besides, it integrates multiple diffusion graphic convolutional networks with prior knowledge of airport network topology to comprehensively learn the spatial correlations. Experimental results on two large real-world datasets from China and the U.S. showed that compared with 12 baseline methods, the prediction mean absolute error of our TMTSC was reduced by 43.61% in average on the China dataset and 15.31% on the U.S. dataset, proving the excellence of our method. The implementation codes are available at https://github.com/sculiang/FlightDelayPrediction_TMTSC.