PIGWN: Physics-Informed Graph WaveNet for Airport Flight Traffic Flow Prediction

The continuous growth of global travel demand leads to the continuous rise of airport flight flow, leading a rising need to accurately predict the airport flight traffic flow. Current research predominantly employs Graph Convolutional Neural Networks (GCN) to model intra-airport traffic changes, but overlooks comprehensive modeling of air traffic routes and temporal dimensions. Furthermore, existing models rely on data-driven approaches lacking a deep understanding of traffic change mechanisms, resulting in poor performance in new scenarios. To solve these problems, we propose a deep learning network framework Physics-Informed Graph WaveNet (PIGWN) based on multi-scale feature data fusion and constructs two new airport domain data sets. We integrated a physics-based module prior to GraphWaveNet to learn and integrate multi-scale spatio-temporal information of airports, embedding physical formulas reflecting airport traffic changes. The physics learning module fuses multi-scale spatio-temporal feature data considering the time delay and traffic conservation relationship between different airports. The Graph WaveNet (GWN) structure learns and predicts multi-scale spatio-temporal feature data, and uses the time-space convolution block model to capture and model spatio-temporal features. PIGWN model has certain theoretical significance and practical application value in the field of airport flight flow prediction, which realizes the accurate prediction of airport flight flow and provides certain decision support and reference for air traffic management. We have made the code publicly available at https://github.com/yzc0912/PIGWN.