Abstract: Flight delays in airport networks exhibit pronounced spatiotemporal propagation characteristics. Their temporal dependency scales vary with operational conditions, while inter-airport delay propagation is jointly influenced by route connectivity and network operating conditions. Existing spatiotemporal forecasting methods typically employ fixed temporal receptive fields, limiting their ability to capture delay patterns at different temporal scales as operational conditions change. Meanwhile, methods based on static adjacency relationships struggle to characterize dynamic variations in inter-airport influence, whereas spatial attention without topological constraints may introduce long-range associations inconsistent with the actual route network. To address these challenges, this study proposes an adaptive spatiotemporal graph learning method for airport network delay prediction, termed DAD-GWN. For temporal modeling, an Adaptive Dilation Routing Temporal Convolutional Network (ADR-TCN) module constructs a state representation based on the overall delay level and fluctuation intensity within the input window. A routing mechanism adaptively combines temporal convolution branches with different dilation rates, enabling state-adaptive modeling of temporal dependencies. For spatial modeling, a Route-Constrained Dual-Path Spatial Aggregation (RC-DPSA) module integrates bidirectional diffusion graph convolution with route-topology-constrained spatial attention. The former extracts multi-hop neighborhood information, while the latter dynamically learns inter-airport influence weights within a Ka-hop reachable neighborhood. Experiments use hourly operational data from 40 airports covering the entire year of 2024, obtained from the U.S. Bureau of Transportation Statistics, with the preceding 12 h used to predict the subsequent 6 h. Model performance is evaluated through baseline comparisons, multi-horizon forecasting, module ablation experiments, and mechanism analysis. The proposed method achieves an MAE of 7.400 min and an RMSE of 24.704 min across the six-hour prediction horizon and outperforms the main baseline models at each forecast horizon. These results demonstrate the effectiveness of combining adaptive temporal-scale modeling with route-topology-constrained spatial aggregation. The method can provide predictive support for airport network delay early warning, operational situation assessment, and collaborative traffic flow management, helping identify potential delay risks and inform operational resource allocation and the development of delay mitigation measures.

摘要： 机场网络中的航班延误具有显著的时空传播特征，其时间依赖尺度随运行状态变化而改变，而机场间延误传播受到航线连接关系及网络运行状态的共同影响。现有时空图预测方法通常采用固定时间感受野，难以根据运行状态变化有效捕捉不同时间尺度的延误特征；同时，基于静态邻接关系的方法难以刻画机场间影响强度的动态变化，而缺乏拓扑约束的空间注意力可能引入与实际航线结构不一致的远距离关联。针对上述问题，提出一种面向机场网络延误预测的自适应时空图学习方法（DAD-GWN）。在时间建模方面设计自适应膨胀路由时间卷积模块（ADR-TCN），根据输入时间窗口的整体延误水平和波动强度构建状态表示，通过路由机制自适应融合不同膨胀率的时间卷积分支，实现时间依赖尺度的状态自适应建模；在空间建模方面设计航线约束双路空间聚合模块（RC-DPSA），将双向扩散图卷积与航线拓扑约束的空间注意力相结合，利用前者提取多阶邻域信息，在Ka跳可达范围内动态学习机场间影响强度。基于美国运输统计局2024年全年40个机场的小时级运行数据，采用过去12 h预测未来6 h的实验设置，通过基线对比、多步预测、模块消融及机制分析评估模型性能。实验结果表明，所提方法在6 h综合预测任务中取得7.400 min的MAE和24.704 min的RMSE，并在各预测时距下取得优于主要基线模型的预测性能，验证了自适应时间尺度建模与航线拓扑约束空间聚合的有效性。可为机场网络延误预警、运行态势研判及协同流量管理提供预测依据，辅助识别潜在延误风险，为运行资源调配和延误缓解措施的制定提供参考。