Decision-focused learning for gate assignment to reduce potential conflicts under arrival time uncertainty

Abstract

Flight delays frequently disrupt airport operations, making pre-planned gate assignments infeasible and requiring real-time adjustments. Traditional ‘Predict-then-Optimize’ methods train prediction models to minimize statistical errors (e.g. MSE), but they often ignore how different prediction errors affect the final schedule. This study proposes a Decision-Focused Learning (DFL) framework that combines a deep learning arrival-time predictor with a differentiable gate assignment optimization model. By integrating the optimization problem into the training process, the model learns decision-aware predictions using a differentiable surrogate of model-defined operational regret. Validated on real-world operational data from a major international hub, our approach produces fewer potential conflicts in the initial gate plans compared to standard baselines. Notably, the proposed method reduces potential gate conflicts in the initial Stage 1 plan under true arrivals before recourse by up to 30.9% compared to the standard predict-then-optimize strategy.