Stochastic Departure Metering with Machine-Learning Taxi-Time Predictors

In busy airports, departure metering (DM) can reduce unnecessary runway-area queueing by holding aircraft at their stands, but its effectiveness is highly sensitive to the accurate estimation of taxi time. Existing studies generally pursue two separate approaches to improve the effectiveness of DM decisions: (1) improving the accuracy of taxi-time prediction, and (2) incorporating taxi-time uncertainty to make more robust decisions. However, prediction errors remain unavoidable, and uncertainty sets constructed directly from historical data of multiple flights fail to capture the heterogeneity between flights, so the previous approaches always result in suboptimal decisions. Thus, this research develops a learning-based two-stage stochastic DM that constructs the uncertainty sets from the outputs of machine-learning predictors and makes robust decisions under uncertainty. This framework can capture the uncertainty set for each flight more accurately and effectively reduce the influence of prediction errors on DM decisions. The DM model is formulated as a two-stage stochastic program integrating first-stage pushback decisions and second-stage runway sequencing, solved by the sample average approximation and scenario generation method. The proposed framework is evaluated using real-world data from Tokyo International Airport (RJTT). The results show that the framework can effectively reduce the overall queueing level and right-tail risk of runway-area queueing time compared with the certainty model, with reduced expected queueing time during peak hours. Additional equity and sensitivity analyses further show that these improvements are realized without being achieved at the expense of a small subset of flights and remain stable under schedule perturbations. This study demonstrates that learning-based uncertainty representations can enhance the robustness and operational quality of DM decisions.