Sector entry flow prediction based on graph convolutional networks

• Main Authors: Ma, Chunyao, Alam, Sameer, Cai, Qing, Delahaye, Daniel
• Other Authors: School of Mechanical and Aerospace Engineering
• Format: Conference or Workshop Item
• Language: English
• Published: 2022
• Subjects: Engineering::Aeronautical engineering::Aviation; Air Traffic; Flow Prediction
• Online Access: https://hdl.handle.net/10356/160332; https://www.icrat.org/
• Institution: Nanyang Technological University

Improving short-term air traffic flow prediction can help forecast demand and maximize existing capacity by tactical air traffic flow management. Most existing studies in flow prediction lacks consideration of the dynamic, structural, and interrelated nature of air traffic flows in the airspace. Therefore, this paper proposes to predict sector entry flows based on graph convolutional networks, which consider the dynamic spatial-temporal features of air traffic from a graph perspective. First, we specify a sector entry flow based on its upstream and downstream sectors. Then, each entry flow is denoted as a node in a graph. The weighted edges between the nodes are learned from a Word2vec model based on air traffic flows among the nodes. With the weighted graph constructed and the temporal flows on the nodes extracted from the flight trajectories, an Attention-based Spatial-Temporal Graph Convolutional Network (ASTGCN) module is adopted to capture spatial-temporal features of recent, daily-periodic, and weekly-periodic flows in the graph. Finally, The outputs from the ASTGCN module based on the three features are fused to generate the final prediction results. The proposed method is applied on 164 sectors of French airspace for one-month ADS-B data )from Dec 1, 2019, to Dec 31, 2019) which includes 158,856 flights. Results show that, the proposed method outperforms the well established Long short-term memory (LSTM) model, and demonstrates better capability in predicting rapid changes in traffic flow and has relatively smaller decrease in prediction accuracy as the prediction time-window increases.