Gradient-based smart predict-then-optimize framework for aircraft arrival scheduling problem

This paper introduces a gradient-based Smart Predict-then-Optimize (SPO) framework for solving the Aircraft Arrival Scheduling Problem (ASP) in Terminal Maneuvering Area. Traditional approaches to ASP typically separate arrival time prediction from scheduling optimization, potentially leading to incomplete solutions. We address this limitation by developing an end-to-end learning framework that directly integrates prediction with optimization objectives. Our methodology introduces the concept of traffic instances for simultaneous prediction of multiple aircraft arrival times, coupled with a Mixed Integer Programming (MIP) model for scheduling optimization. We evaluate our approach using real-world data from London Gatwick Airport, analyzing 47452 arrival flights from June to September 2024, organized into 2404 traffic instances. The framework incorporates comprehensive weather data through the ATMAP algorithm, considering factors such as wind, visibility, precipitation, and dangerous phenomena. Experimental results demonstrate that the MLP+SPO+ framework shows particular effectiveness in adapting to adverse weather conditions, strategically balancing transit times with operational efficiency. While the minimum time window is required, the MLP+SPO+ will reach around 85.0% and 43.4% lower costs compared with the First-Come-First-Serve (FCFS) cost and optimized true cost, respectively. These findings suggest significant potential for improving arrival scheduling efficiency through integrated SPO approaches.