Predictive and prescriptive analytics for robust airport gate assignment planning in airside operations under uncertainty

With the increasing demand for air transport, numerous airports have exceeded their available capacity, resulting in more frequent congestion and disruptions. Therefore, airport gate assignment plans must prioritise robustness to alleviate congestion, absorb disruptions, and maintain high service levels. Given the uncertainties in airside operations, providing robust decisions is challenging. To address this issue, we employ two prescriptive analytics approaches to develop airport gate assignment plans. These approaches leverage historical data, auxiliary data, and machine learning (ML) methods to enhance decision effectiveness and robustness. Initially, we adopt a predict-then-optimise approach, utilising ML methods to predict aircraft arrival times. These predictions are then used as input for a deterministic model of the airport gate assignment problem (AGAP). Subsequently, we explore an estimate-then-optimise approach. In this approach, we first estimate the distribution of uncertain aircraft arrival times using ML methods. Then, we solve the two-stage stochastic programming model for the AGAP based on the estimated distribution. Given the complexity of the estimate-then-optimise approach, we develop an effective scenario selection strategy, the cluster-based scenario reduction (CSR) method, to maintain tractability while ensuring decision performance. Concurrently, we develop an efficient exact solution method, the Benders-based branch-and-cut (BBC) method, to effectively handle larger and more complex test instances. Numerical experiments using real-world data from Xiamen Gaoqi International Airport demonstrate the effectiveness of the CSR and BBC methods. The CSR method performs better with a smaller sample size, while the BBC method significantly enhances computational performance compared to commercial solvers. These proposed methods improve the tractability and scalability of the estimate-then-optimise approach. Notably, the estimate-then-optimise approach outperforms the predict-then-optimise approach driven by the same ML method. Furthermore, we find that estimate-then-optimise approaches, supported by well-performing ML methods and scenario selection strategies, provide superior performance compared to other optimisation approaches.