Generative Imitation Learning Enhanced XGBoost Approach for Increasing Airport Bridge Lending Rate

Increasing the probability of inbound and outbound aircraft docking at close aircraft stands, also referred to as bridge lending rate, can effectively reduce airport operating costs and enhance passenger experience. However, there are two significant limitations that seriously impair the bridge lending rate. First, the inaccuracy of the estimated time of arrival (ETA) due to the lack of complete flight trajectory data and poor generalization performance. Second, existing aircraft stand allocation modeling cannot meet the requirements under the complex composite aircraft stand scenarios. To address these limits, we first converted the issue of data expansion and generalization into a transferable generative imitation learning task. We proposed Generative Imitation Learning Enhanced XGBoost (GILEX), which raised the prediction accuracy of ETA by about two minutes. Secondly, we conducted a composite aircraft stand allocation (CASA) model. Finally we used the Gurobi optimizer to solve the composite aircraft stand allocation problem. We first test GILEX in physics-based control tasks from OpenAI Gym. Then we combined test GILEX and CASA modeling pipeline using the actual historical operational trajectories of different fights and aircraft stand data of Tianfu Airport. Experimental results suggest that, the proposed pipeline can increase the aircraft bridge lending rate by about 9 percentage points at most compared to the existing aircraft stands allocation methods.