Dynamic Spatial-Temporal Propagation Neural Network for Airport Delay Forecasting Via Traffic Flow and Delay Embedding

Research on airport delay forecasting is essential for quickly identifying key bottleneck points within the airport network, which enables controllers to promptly assess traffic conditions and take appropriate actions. However, there are two major challenges that should be sincerely considered: (1) modeling the cascading effects of airport delay is often incomplete in existing prediction methods; (2) there is an underlying relationship between traffic flow and delays, but it is difficult to measure how flow changes impact airport delays. To address these problems, we propose a dual-path deep learning framework called the dynamic spatial-temporal propagation neural network (DSTPNN) to forecast airport delays. Specifically, the spatial-temporal convolutional embedding (STCE) module is designed to capture delay propagation patterns by considering dynamic and static spatial dependencies, together with multi-scale temporal features. To model the evolutionary correlations between traffic flow and delays, the auxiliary temporal embedding (ATE) module is first constructed based on causal and dilated convolution to learn high-level feature representations of traffic flow. Based on the output of both components, we design a traffic situation awareness attention (TSAA) mechanism that incorporates both convolution and cross-attention techniques to mine the potential causal relationships between traffic features (flow and delays) at the airport level. Experiments on a real-world dataset from the Bureau of Transportation Statistics (BTS) indicate that the proposed DSTPNN outperforms the baselines, obtaining relative improvements of at least 1.5% in MAE, 2.3% in RMSE, and 9.7% in $R^2$. Furthermore, ablation and analysis experiments demonstrate that each proposed technical block is instrumental in achieving the desired performance enhancements.