全球约80\%的货物依赖海上运输，而港口作为海运关键节点，其运营效率直接影响到全球贸易的发展。在港口运营中，来港船只需停靠在指定泊位，并由岸桥对船上集装箱装卸,以此实现完整的港船交互。因此，泊位的高效分配与岸桥的科学调度共同决定了港口的作业效率。然而，优化这两个关键环节面临着巨大的挑战：一方面，泊位与岸桥之间存在着复杂的相互制约关系，优化一个环节往往会影响到另一个环节的运行；另一方面，不确定性因素的干扰也增加了优化难度，尤其是船只到港时间的不确定性，使得调度工作难以精确掌控。 为有效应对这些挑战，本文设计并提出了一种创新的分布鲁棒（DRO）泊位和岸桥联合调度问题（BACAP）模型。不同于传统模型，本文将船只到达延误视为随机变量，并采用基于K聚类的机器学习方法构建模糊集，对船舶到达延误进行有效分类的同时保留了数据的整体特性。在精确求解方面，通过采用Benders分解策略，本模型被分解为确定性主问题和不确定性子问题：主问题处理岸桥分配和船舶相对位置，而子问题则决策船舶实际停泊位置和到离港时间。然而，在Benders迭代过程中，主问题的计算复杂度会逐步增加。为加速主问题求解，本文引入了列与约束生成（C\&CG）割，并结合问题的特性构建了多个有效不等式，从而显著提升了求解效率。在子问题层面，本文通过样本平均近似（SAA）方法处理模型中期望值的计算，进而将其转化为可解的二阶锥规划（SCOP）模型。由于子问题的求解效率受到SAA场景数的影响，场景数越多，求解速度越慢。针对这一挑战，本文提出了一种基于在线凸优化（OCO）的启发式算法加速子问题求解。该算法具有较佳的通用性，可拓展至任意SAA模型。 在数值实验环节，本研究基于常州某集装箱港口的真实数据，构建了不同规模的算例。大量实验结果表明OCO算法在求解时间上平均缩短了94\%，效率显著；在求解质量上，OCO算法在中小算例中能取得最优解，在大算例中虽有细微差距，但仍在可接受范围。此外，本文比较了多种相关模型，展示了DRO模型的调度优势，并通过对比实验验证了有效不等式在加速求解中的有效作用。最后，通过灵敏度分析，本文探讨了不同机器学习聚类方法构建的模糊集以及SAA场景数对模型性能的影响。
Approximately 80% of global goods are transported by sea, making ports a critical node in maritime transportation. The efficiency of port operations directly impacts global trade flow and costs. In port operations, arriving ships must dock at designated berths, and quay cranes are used to load and unload containers from the ships, thus achieving complete port-ship interaction. Therefore, efficiently allocating berths to arriving vessels and scheduling quay cranes are crucial. These processes form a complete port-ship interaction flow, affecting the overall efficiency of port operations. However, optimizing these interaction processes faces significant challenges due to the complex interplay between processes and the inherent uncertainties of maritime transport, especially the unpredictability of ship arrivals.This thesis designs and proposes an innovative Distributionally Robust Optimization (DRO) model for the Berth Allocation and Crane Assignment Problem (BACAP) to address these challenges effectively. Unlike traditional models, this thesis considers ship arrival delays as random variables. It employs a K-cluster-based machine learning method to construct an ambiguity set, effectively classifying ship arrival delays while preserving the overall characteristics of the data. For exact algorithms, the model is cleverly decomposed into a deterministic master problem and an uncertain subproblem using the Benders decomposition strategy: the master problem deals with crane allocation and the relative positioning of ships, while the subproblem focuses on the actual berthing positions and the arrival and departure times of ships. Several valid inequalities based on the problem‘s characteristics and column-and-constraint generation (C&CG) constraints are proposed, which significantly improve the solution efficiency. Specifically, the thesis applies the Sample Average Approximation (SAA) method for the subproblem to transform it into a solvable Second-Order Cone Programming (SCOP) model. Even so, the solution efficiency of the subproblem is affected by the number of scenarios in the SAA process, with more scenarios leading to slower solutions. To address this issue, the thesis proposes a heuristic algorithm based on Online Convex Optimization (OCO) to accelerate the solution of the subproblem. Notably, this algorithm demonstrates excellent versatility, making it extensible to any SAA model. In the numerical experiment section, this thesis uses real data from a medium-sized container port in Changzhou to generate instances of different scales. Extensive experimental results show that the OCO algorithm reduces the solution time by an average of 94%, demonstrating its high efficiency. Regarding solution quality, the OCO algorithm achieves exact solutions for small- and medium-scales and remains within an acceptable range for big-scale. Furthermore, the thesis compares various models, showcasing the scheduling advantages of the DRO model. Also, this thesis validates the key role of valid inequalities in accelerating the solution through comparative experiments. Finally, through sensitivity analysis, the thesis explores the impact of different machine learning clustering methods used to construct the ambiguity set and the number of SAA scenarios on the model‘s performance.
