Variable Taxi-Out Time Prediction Using Graph Neural Networks

Airport Collaborative Decision Making (ACDM) is an important initiative that aims at more efficient and optimised use of airport resources. Variable taxi time prediction is one of the key elements in ACDM, supporting the tactical planning needed to ensure smooth traffic flow and optimal use of taxiway resources. This paper presents a mesoscopic data-driven model for the prediction of variable taxi-out times together with the associated data processing stages. To support operational implementation, the model utilises features that are readily available as part of the pre-departure tactical planning phase – namely, information on the intended taxi route and milestone time estimates. By using a Graph Neural Network (GNN) framework, each trajectory can be represented as a sub-graph of the airport taxi network, and GNN convolution operations be performed on this subgraph to extract meaningful features. Both impeded and unimpeded taxi time predictions from the GNN model are compared against standard methodologies by the Federal Aviation Authority (FAA) and EUROCONTROL, as well as against predictions made by Gradient Boosted Machines (GBM), a popular tree-based machine learning technique. Results show that both GNN and GBM models outperform standard FAA and EUROCONTROL methods (with RMSE and MAE of the former group lower by 40% to 60% relative to the latter), and the novel GNN model slightly outperforms the GBM model by around 2 seconds, or a 2% to 4% improvement in model performance.