The continuous rise in global container traffic has necessitated more efficient operational strategies at container terminals (CTs). However, developing effective strategies requires mitigating delays in handling equipment operations. Yard trucks (YTs) are pivotal for transporting containers between quay cranes (QCs) and yard cranes (YCs). A YT shortage can cause operational delays for both QCs and YCs, extending the makespan of main-vessel operations at the berth. Conversely, deploying an excessive number of YTs eliminates these delays but incurs higher operational costs. Consequently, determining the optimal number of YTs constitutes a bi-objective optimization problem involving a tradeoff between minimizing the operational makespan and associated YT deployment costs. This study introduces a machine learning (ML)-enhanced non-dominated sorting genetic algorithm 2 (NSGA-2) approach designed to address this challenge by minimizing both the makespan and deployment costs. The proposed method uses ML models to predict fitness scores and efficiently guide the search toward optimal solutions, while also assisting in generating, selecting, and mutating promising solutions during the optimization process. The effectiveness of ML-enhanced NSGA-2 was validated using empirical data from Busan Port Terminal in South Korea. The results highlight the superior performance and practical applicability of the proposed approach, which achieved a significant reduction of 16% in deployed YTs compared to existing strategies.

As global supply chains grow increasingly complex and cargo volumes continue to rise, the accurate prediction of vessel turnaround time (VTT) has become essential for enhancing port operational efficiency. Previous studies primarily utilized static features, such as vessel specifications and container volumes, often failing to capture the dynamic nature of real-world operations. To address this limitation, we propose a forecasting framework that incorporates dynamic operation indicators (OIs) from a two-stage queuing model into time-series models. Leveraging operational data from Busan Port, our approach achieves of up to 28% improvement in mean absolute percentage error compared with existing techniques. Explainability analysis using Shapley additive explanations (SHAP) further demonstrates that OIs rank among the most influential features. By providing more reliable VTT predictions, port operators can enhance berth planning, optimize resource allocation, minimize scheduling delays, and enhance overall port efficiency. These results highlight the practical value of the framework for data-driven decision-making in port operations.

Forecasting the maritime economics index, including container volume and Baltic Panamax Index, is essential for long-term planning and decision-making in the shipping industry. However, studies on container volume prediction are not sufficient, and the bulk freight index has highly fluctuating characteristics, which pose a challenge in long-term prediction. This study proposes a new hybrid framework for the long-term prediction of the maritime economics index. The framework consists of time-series decomposition to break down a time-series into several components (trend, seasonality, and residual), a two-stage attention mechanism that prioritizes important variables to increase long-term prediction accuracy and a long short-term memory network that predicts and combines all components to derive the final predictive outcome. Extensive experiments are conducted using the container volume data, bulk freight index data, and various external variables. The proposed framework achieved a better predictive performance than existing time-series methods, including conventional machine learning and deep learning-based models, in the long-term prediction of container volume and the Baltic Panamax Index. Hence, the proposed method can help in decision-making through accurate long-term predictions of the maritime economics index.

Time-series forecasting (TSF) is a traditional problem in the field of artificial intelligence, and models such as recurrent neural network, long short-term memory, and gate recurrent units have contributed to improving its predictive accuracy. Furthermore, model structures have been proposed to combine time-series decomposition methods such as seasonal-trend decomposition using LOESS. However, this approach is learned in an independent model for each component, and therefore, it cannot learn the relationships between the time-series components. In this study, we propose a new neural architecture called a correlation recurrent unit (CRU) that can perform time-series decomposition within a neural cell and learn correlations (autocorrelation and correlation) between each decomposition component. The proposed neural architecture was evaluated through comparative experiments with previous studies using four univariate and four multivariate time-series datasets. The results showed that long- and short-term predictive performance was improved by more than 10%. The experimental results indicate that the proposed CRU is an excellent method for TSF problems compared to other neural architectures.

Binary classification (BC) is a practical task that is ubiquitous in real-world problems, such as distinguishing healthy and unhealthy objects in biomedical diagnostics and defective and non-defective products in manufacturing inspections. Nonetheless, fully annotated data are commonly required to effectively solve this problem, and their collection by domain experts is a tedious and expensive procedure. In contrast to BC, several significant semi-supervised learning techniques that heavily rely on stochastic data augmentation techniques have been devised for solving multi-class classification. In this study, we demonstrate that the stochastic data augmentation technique is less suitable for solving typical BC problems because it can omit crucial features that strictly distinguish between positive and negative samples. To address this issue, we propose a new learning representation to solve the BC problem using a few labels with a random pair cross-distance learning mechanism. First, by harnessing a few labeled samples, the encoder network learns the projection of positive and negative samples in angular spaces to maximize and minimize their inter-class and intra-class distances, respectively. Second, the classifier learns to discriminate between positive and negative samples using on-the-fly labels generated based on the angular space and labeled samples to solve BC tasks. Extensive experiments were conducted using four real-world publicly available BC datasets. With few labels and without any data augmentation techniques, the proposed method outperformed state-of-the-art semi-supervised and self-supervised learning methods. Moreover, with 10% labeling, our semi-supervised classifier could obtain competitive accuracy compared with a fully supervised setting.