A Physics Informed Graph Attention Transformer with Weather Driven Delay Propagation for Air Traffic State Predictio

Abstract

Accurate prediction of air traffic states, including airport queue lengths and cumulative terminal-area travel times, is essential for efficient air traffic management and congestion mitigation. Forecasting remains challenging due to complex spatiotemporal dependencies, and data-driven models often overlook the underlying physics of queue formation and dissipation. Although recent physics-informed models such as PIGAT improve prediction accuracy by incorporating queue dynamics, they do not explicitly model how weather induced delays propagate through interconnected airport networks. Delays rarely stay local: an aircraft delayed at one airport carries that delay to its next destination, and connected airports inherit disruption even under otherwise normal conditions. We introduce a novel weather-driven delay propagation equation as an additional physics-informed loss with the PIGAT framework. The proposed propagation model combines local queue congestion, weather conditions, and inbound delays from connected airports. These interactions are weighted using PIGAT’s blended flow-based and learned adjacency matrix, enabling the model to capture cascading delays across the national air transportation network rather than treating airports independently. Computational results demonstrate improved prediction accuracy over the baseline PIGAT model, particularly during periods of severe weather and network congestion.