Machine Learning Applications for Seaside Port Operations Planning and Scheduling

Aplicaciones de aprendizaje automático para la planificación y programación de operaciones portuarias marítimas

Maria Boluda-Prieto¹, Ana Esteso², M.M.E. Alemany³, Angel Ortiz⁴

Received: 22/1/2026 | Accepted: 18/2/2026

Abstract

The surge in containerised trade has intensified the need for efficient resource allocation in port operations, particularly in berth allocation (BAP), quay crane assignment (QCAP) and quay crane scheduling (QCSP) problems. While mathematical programming and metaheuristic approaches have traditionally been used to solve these problems, their scalability and adaptability remain limited. Recent advances in Machine Learning (ML) offer new optimisation possibilities. This paper conducts a systematic literature review of approaches that apply machine learning techniques to seaside port operations (BAP, QCAP, QCSP, and their integrated variants). Operations such as quayside transport planning, landside or yard activities are outside the scope of this review. The review was conducted following PRISMA guidelines and was based on publications indexed in Scopus and Web of Science. The selected works are analysed and classified across nine dimensions. The results highlight a predominant focus on economic optimisation, with BAP being the most frequently addressed problem. Earlier studies broadly applied regression techniques and predictive approaches. At the same time, more recent research has shifted towards deep reinforcement learning and prescriptive solutions, opening new opportunities for sustainable and adaptive port operations. Finally, it has identified key trends and opportunities in ML applications for port operations.

Keywords: berth allocation, quay crane assignment, quay crane scheduling, artificial intelligence.

Resumen

El aumento del comercio contenerizado ha intensificado la necesidad de una asignación eficiente de recursos en las operaciones portuarias, especialmente en los problemas de asignación de atraques (Berth Allocation Problem, BAP), asignación de grúas de muelle (Quay Crane Assignment Problem, QCAP) y programación de grúas de muelle (Quay Crane Scheduling Problem, QCSP). Aunque tradicionalmente se han utilizado enfoques basados en programación matemática y metaheurísticas para resolver estos problemas, su escalabilidad y adaptabilidad siguen siendo limitadas. Los avances recientes en aprendizaje automático (Machine Learning, ML) ofrecen nuevas posibilidades de optimización. Este artículo realiza una revisión sistemática de la literatura sobre enfoques que aplican técnicas de aprendizaje automático a las operaciones portuarias en el lado marítimo (BAP, QCAP, QCSP y sus variantes integradas). Quedan fuera del alcance de esta revisión operaciones como la planificación del transporte en el muelle, las actividades del lado terrestre o las operaciones de patio. La revisión se llevó a cabo siguiendo las directrices PRISMA y se basó en publicaciones indexadas en Scopus y Web of Science. Los trabajos seleccionados se analizan y clasifican en nueve dimensiones. Los resultados muestran un claro predominio de la optimización económica, siendo el BAP el problema abordado con mayor frecuencia. Los estudios iniciales aplicaban principalmente técnicas de regresión y enfoques predictivos. Por su parte, las investigaciones más recientes han evolucionado hacia el aprendizaje profundo por refuerzo y soluciones prescriptivas, abriendo nuevas oportunidades para unas operaciones portuarias más sostenibles y adaptativas. Finalmente, se identifican las principales tendencias y oportunidades en la aplicación del aprendizaje automático a las operaciones portuarias.

Palabras clave: asignación de atraques, asignación de grúas de muelle, programación de grúas de muelle, inteligencia artificial.

1. Introduction

Maritime transport is a fundamental pillar of international trade, as it carries the largest portion of both the volume and value of global transactions. It is estimated that more than 80% of global trade (by volume) and over 70% (by value) depend on seaports (Oudani et al., 2023). These infrastructures function as strategic nodes within an intermodal transportation network that connects maritime traffic with overland systems via railroads and highways (Castilla-Rodríguez et al., 2020). The growth in maritime traffic has driven the need to optimise port operations, requiring a more efficient management of resources and enhanced sector competitiveness (Le et al., 2024). Within this context, ensuring a swift and efficient flow of containers at the lowest possible cost has become a priority for port terminals. This is because loading and unloading times directly impact the duration of the vessel's stay at port, thereby affecting both terminal productivity and operational profitability (Kolley et al., 2021).

To enhance efficiency in port operations, it is essential to address the various interrelated challenges in seaside terminal operations, a specific subset of port operations that includes operational planning for vessel berthing and quay crane assignment and scheduling (Bierwirth & Meisel, 2010). Berth Allocation Problem (BAP), determines where and when vessels berth; the Quay Crane Assignment Problem (QCAP), defines the assignment of quay cranes to vessels, including both the number of cranes allocated and the specific cranes involved; and the Quay Crane Scheduling Problem (QCSP), defines the sequence of crane operations (Correcher et al., 2024). Beyond these individual problems, integrated approaches such as the Berth and Quay Crane Assignment Problem (BACAP) and the Berth Allocation and Quay Crane Assignment and Scheduling Problem (BACASP) seek to optimise the resource allocation and scheduling simultaneously (Bierwirth & Meisel, 2010). In contrast, planning problems related to internal transport (e. g., quayside horizontal transport) and yard or landside operations are outside the scope of this work.

Traditionally, these problems have been tackled using mathematical programming techniques for small-scale instances. However, due to their high computational complexity, larger problem instances typically require heuristics or metaheuristics to obtain near-optimal solutions within reasonable computational times. Despite their effectiveness, these approaches often struggle with real-time adaptability and the increasing complexity of modern port operations. These limitations have driven interest in AI-based methods, which have demonstrated their effectiveness in other scheduling problems (Velastegui et al., 2023).

In this context, Machine Learning (ML) has emerged as a potential tool, either as a complement to traditional solution methods or as the primary resource for optimisation and decision-making. Depending on the specific problem and context, ML has been applied with various objectives. Its main uses include resource allocation optimisation, data prediction and analysis, and automated decision-making in dynamic environments. While research on ML applications in port operations exists, a systematic classification of these studies remains scarce. However, in other domains, prior work confirms the growing role of AI in decision-making (Mateo-Casalí et al., 2025). In this line, Herrera (2025), highlights the importance of explainability and human-artificial intelligence collaboration for supporting trustworthy decision-making processes.

For example, Mekkaoui et al. (2020), provide a general classification of ML applications in port operations, organising the literature by port functional areas (e. g., seaside, yard, landside, etc.), by the nature of the learning task addressed, and by the types of techniques employed. Complementarily, Mansoursamaei et al. (2023) conduct a systematic review focused on environmental sustainability in ports, structuring the evidence around environmental impact, the functional area of application within the port, and the ML techniques used. Taken together, these reviews provide a general map of ML use in the port domain. However, despite this broad mapping, they are not oriented towards a detailed synthesis of planning and scheduling problems in the seaside area when the objective is to compare approaches at the level of specific decision subproblems (BAP, QCAP, QCSP, and their integrated formulations). This approach is highly valuable, but it makes it more difficult to draw specific conclusions about operational decisions.

As a consequence, although these reviews may include studies related to seaside operations, they do not provide a synthesis specifically structured around the subproblems inherent to this area (BAP, QCAP, QCSP and their integrated formulations) that would allow the evidence to be consolidated in a homogeneous manner and in a way that is directly useful for decision-making. Therefore, despite their value as general reviews, a more specific, granular review remains necessary to systematise how ML is being applied to these problems; not only which techniques are used, but also for what purpose, under which operational assumptions and with what types of data the models are developed.

Filling this gap, Boluda-Prieto et al. (2025) propose a preliminary review of ML applications to seaside port operations. Building on that foundation, this paper extends the previous review in several ways. First, it provides a more detailed characterisation of the analysed corpus. Second, it includes a bibliometric analysis of publication sources and their impact at the time of publication. Third, it offers a structured overview of each selected study to contextualise the problem addressed, the proposed approach and the main results. Finally, it conducts a cross-dimensional analysis to identify methodological patterns, underexplored combinations, and structural research gaps with improved traceability.

To ensure reproducibility and enable systematic comparison across studies, the review is structured and analysed across nine dimensions: (1) the type of problem addressed, (2) the solution approach adopted, (3) the ML techniques used, (4) the purpose of ML in the study, (5) the type of data used for ML models, (6) the objective pursued in each study, (7) the operational context considered, (8) whether replanning strategies are incorporated, and (9) the technological paradigm under which the study was developed. As a result, this work addresses the following research questions:

RQ1. How has ML been applied to planning and scheduling in seaside port operations?

RQ2. What research gaps, trends, and opportunities emerge in applying ML to seaside port operations?

By addressing these questions, this study makes three main contributions to the field. First, it provides a systematic review focused on ML applications to seaside port operations planning and scheduling. Second, it conducts a comprehensive analysis of existing applications across nine dimensions, examining not only individual aspects but also their interrelationships, enabling the identification of patterns, trends, and gaps in the literature. Third, it outlines future research directions in this field. Taken together, these contributions advance both the academic understanding of ML in port operations and its practical implications for developing more efficient, sustainable and resilient maritime terminals.

The remainder of this paper is structured as follows: Section 2 details the research methodology, Section 3 develops the literature review through five steps: material collection, descriptive analysis, the design of the classification framework, the systematic classification of the reviewed studies, and the cross-dimensional analysis. Finally, Section 4 highlights the main findings and the most promising future research directions.

2. Research methodology

The methodology followed to perform this literature review is structured into five steps and is inspired by the methodologies proposed by Esteso et al. (2018), Lorente-Leyva et al. (2024), and Seuring & Müller (2008).

• Material collection: The scope of this review was defined to studies applying machine learning to seaside port operations planning and scheduling problems, with particular emphasis on the berth allocation problem (BAP), quay crane assignment problem (QCAP), quay crane scheduling problem (QCSP), and their integrated variants BACAP and BACASP, which are the most widely recognised and standardised formulations in the literature. In contrast, planning problems related to internal horizontal transport, yard operations and landside interfaces were outside the scope of the review.

Bibliographic searches were performed in Scopus and Web of Science, covering all records available until 2025. No temporal restrictions were applied. The search focused on the title, abstract, and keyword fields.

The query combined two elements: (i) the problem of interest (berth allocation, quay crane assignment, and quay crane scheduling), and (ii) the applied technology (machine learning). Since the search was restricted to title, abstract, and keywords, it was observed that some works employing machine learning may not explicitly use this term in those fields but instead refer to broader or related concepts such as artificial intelligence or deep learning. To avoid missing such contributions, these terms were also included in the query string.

Based on these criteria, the following query was formulated:

TITLE-ABS-KEY (Berth Allocation OR “Quay Crane Assignment” OR “Quay Crane Scheduling” OR “BACAP” OR “BACASP”) AND TITLE-ABS-KEY (AI OR “Artificial Intelligence” OR “Deep Learning” OR “Machine Learning”)

The following inclusion criteria were applied:

o The paper must be written in English.

o The full text must be accessible.

o The study must apply machine learning techniques.

o The paper must specifically address BAP, QCAP, QCSP or integrated combinations of these problems.

Exclusion criteria were defined as the converse of these rules. Accordingly, studies were excluded if they were not written in English, if the full text was not accessible, if they did not apply ML techniques, or if they did not address the problems under study. The review followed the PRISMA guidelines to ensure transparency and reproducibility.

• Descriptive analysis: The selected studies were characterised according to: (i) year of publication, (ii) type of publication (journal article or book chapter), (iii) publication source, (iv) impact of the source in the year of publication, considering both Journal Citation Reports (JCR) and Scimago Journal Rank (SJR).

• Classification framework: To support a systematic analysis of the selected studies, a classification framework was designed with the objective of identifying the problems addressed with ML, the techniques employed and their functional purposes, as well as other operational and contextual aspects that help to understand existing approaches and to identify the state of the art and gaps in the literature. It was structured around nine dimensions: (i) type of problem addressed, (ii) solution approach adopted, (iii) machine learning techniques applied, (iv) purpose of using machine learning, (v) type of data used for ML models, (vi) objective pursued in each study, (vii) operational context considered, (viii) inclusion of replanning strategies, and (ix) technological paradigm under which the study was developed.

• Classification and analysis: Based on the framework, the studies were to be systematically classified across the defined dimensions. This classification allows for structured comparisons and facilitates the identification of patterns, gaps, and opportunities.

• Cross-dimensional analysis: The interrelations between dimensions were explored. This cross-dimensional perspective enabled the identification of methodological consistencies and underexplored combinations, enriching the understanding of how ML has been incorporated into port operations and pointing to avenues for future research.

3. Literature review

This section reviews the literature on machine learning for planning and scheduling in seaside port operations. It covers the material collection, the descriptive analysis, study-by-study summaries, the classification framework used in this review and the cross-dimensional analysis.

3.1. Material collection

The search was carried out in Scopus and Web of Science (WoS), yielding a total of 120 records. After removing duplicates, 77 unique articles remained. During screening by title and abstract, 49 records were excluded: 1 was not written in English, 3 had no accessible full text, 31 did not apply machine learning techniques, and 14 did not address problems in seaside port operations. Applying the previously described exclusion criteria, this set was reduced to 28. During eligibility, 15 articles were excluded because the full text confirmed the absence of ML application. Following a full-text assessment, 13 studies were finally included in the review.

Figure 1 presents the PRISMA flow diagram, which summarises the identification, screening, eligibility, and inclusion phases of the review process.

3.2. Descriptive analysis

The review covers a total of 13 articles published between 2017 and 2025. No temporal restrictions were applied in the search process; therefore, the limited number of studies and their recent dates highlight that the use of machine learning in seaside port operations planning and scheduling is still an emerging field of research. The earliest publication dates from 2017, while most contributions are concentrated in 2021, 2023, and 2024, accounting for 69% of the reviewed works.

As shown in Figure 2, which displays the annual distribution of publications, the linear fit has a low coefficient of determination (R² = 0.165), indicating that annual production has not followed a stable growth pattern but has fluctuated. However, the moving average suggests an upward trend over the analysed period, in line with growing interest in applying machine learning techniques to the port sector. Overall, the descriptive analysis indicates that this is an emerging field, characterised by low scientific output but undergoing consolidation.

The main bibliometric characteristics of the reviewed studies, including the year of publication, type of source (journal article or book chapter in conference proceedings), and source impact indicators in the year of publication, are analysed. The indicators included are the JCR category and quartile of the source, as well as the SJR subject area, category, and quartile when available. A total of 62% of the reviewed studies were published in scientific journals, while the remaining 38% appeared as book chapters in international conference proceedings. To improve readability, the bibliometric characteristics are presented separately for journal articles (Table 1) and for book chapters in international conference proceedings (Table 2).

Among the eight journal papers, seven were published in journals indexed in Q1 of the JCR in the year of publication, and one in Q2. The JCR categories covered are diverse, including Operations Research & Management Science (50% of journal papers), Computer Science – Artificial Intelligence (38% of journal papers), Engineering, Electrical & Electronic (38% of journal papers) and Oceanography (38% of journal papers), among others. This evidence both supports the multidisciplinary nature of the topic and its positioning in highly ranked journals. In parallel, the SJR classification situates these journals mainly in Q1 in diverse subject areas and categories.

Regarding the five papers published as book chapters in international conference proceedings (Table 2), none of the sources is indexed in JCR. Their visibility in SJR is also heterogeneous. Two proceedings volumes are not indexed in SJR, while two others are indexed with subject areas and categories but without quartile assignment. Finally, one conference proceeding volume is indexed with quartile assignment, appearing in Computer Science (miscellaneous) (Q3) and Theoretical Computer Science (Q4).

3.3. Overview of the selected studies

The 13 selected articles explore the application of ML techniques in port operations, specifically addressing the BAP, QCAP, QCSP, and their integrated formulations.

To provide readers with a preliminary reference, Table 3 offers a concise overview of each study, outlining the problem addressed, the ML approach adopted and the main contributions and results obtained. This overview serves to contextualise the selected words before applying the classification framework introduced in the next subsection and also provides a first step towards addressing RQ1 by showing how ML has been applied across different problems in port operations.

3.4. Classification framework

To facilitate a systematic analysis of the selected studies, a classification framework comprising nine dimensions was employed. These dimensions capture different aspects of the problem and the way ML techniques have been applied, allowing comparisons across studies. Figure 3 illustrates the framework used as a reference for the classification.

The problem type dimension captures the specific problem in seaside port operations planning and scheduling that each study addresses. It distinguishes whether the focus is on the Berth Allocation Problem (BAP), the Quay Crane Assignment Problem (QCAP), the Quay Crane Scheduling Problem (QCSP) or integrated formulations such as BACAP or BACASP.

The solution approach dimension refers to the methodological strategy adopted to solve the problem. Approaches can be exact, heuristic/metaheuristic, ML as a complementary tool, or ML as the primary solution method. Exact methods aim to find the optimal solution to the problem by using mathematical models. They become computationally intractable as the problem size increases, limiting their application to smaller-scale problems (Korekane et al., 2024). For more complex cases, heuristics and metaheuristics are employed; these explore the solution space to obtain high-quality results in less time, although without guaranteeing absolute optimisation. When ML is used as a complementary tool (ML(C)), it usually enhances traditional approaches by adjusting parameters, generating instances, or making predictions. In contrast, ML as the primary solution method (ML(P)) directly supports decision-making and adapts to uncertain environments, making it useful for problems characterised by high variability and multiple interdependent factors.

The ML technique dimension involves methods that learn from data through a training process, unlike exact and heuristic/metaheuristic approaches, which explore the solution space without incorporating a training phase (Karimi-Mamaghan et al., 2022). Within these techniques, the approaches can be regression, deep learning, reinforcement learning and deep reinforcement learning. Regression (RG) is a machine learning supervised approach for predicting continuous values, which allows modelling and analysing the relationship between a dependent variable and one or more independent variables (Tyagi et al., 2022). Deep learning (DL) is a subfield of ML that uses deep neural networks (multiple layers) to learn hierarchical representations of high-dimensional data and perform prediction/classification tasks with a high capacity for automatic feature extraction (LeCun et al., 2015). In reinforcement learning (RL), the agents learn sequential decision-making policies through interaction with an environment and reward actions (Arena et al., 2024). Finally, Deep Reinforcement Learning (DRL) combines DL and RL by leveraging deep neural networks to approximate policies and value functions to address complex environments with unstructured and high-dimensional inputs (François-Lavet et al., 2018). Zhang et al. (2024) highlight it as a promising solution for dynamic planning problems due to its self-learning capability, computational efficiency, and high adaptability.

The purpose of the ML dimension identifies the functional role of ML within each study. Applications may be descriptive, predictive, prescriptive, learning-oriented, or control-oriented (Bačiulienė et al., 2023; Waller & Fawcett, 2013). The descriptive role involves analysing historical and current datasets without generating forecasts in order to understand the current situation and the relationship among different elements of the problem; the predictive role extrapolates future events from historical trends; the prescriptive role forecasts events and recommends precise actions; the learning role dynamically adapts the model to enhance performance over time; and finally, the control role autonomously executes real-time decisions without human intervention.

The type of data dimension classifies the nature of the information used to train or implement ML models. Categories include real-time data, which is stored immediately and reflects the current state of the system; historical data, comprising records that allow for trend analysis (Pereira et al., 2025); simulated data, generated through computational models that replicate the system's behaviour (Ding et al., 2023); synthetic data, produced artificially (Xie et al., 2025; Yue et al., 2025); and predictions, which represent future estimates based on patterns or predictive models.

The objective pursued dimension specifies the goals that the analysed studies aim to achieve. It is classified into six categories: economic sustainability, environmental sustainability, social sustainability, resilience, robustness and security. Economic sustainability aims to achieve optimal financial conditions and ensure overall viability; environmental sustainability focuses on reducing emissions and improving energy efficiency(Cammin et al., 2020; Kolley et al., 2021); social sustainability aims to improve working conditions, maximise positive community impact and promote staff wellbeing; resilience is distinguished by its objective to ensure stability and efficiency in the face of environmental variability and uncertainty (Kolley et al., 2023; Yu et al., 2018); and robustness refers to the ability to adapt and recover from unexpected disruptions (de León et al., 2017; Lv et al., 2024; Zhou et al., 2024). Finally, security encompasses cybersecurity, data protection and operational integrity.

The context dimension delineates the environment in which the research is conducted. A deterministic context assumes all parameters are known and predictable, whereas an uncertain context explicitly considers variability, uncertainty, and unforeseen changes (Esteso et al., 2018).

The replanning strategy dimension considers how and when adjustments are incorporated into the decision-making process. Event-based strategies update decisions in response to specific triggers (Al Samrout et al., 2024; Aljuaid et al., 2024), while periodic strategies apply adjustments at fixed intervals (Lorente-Leyva et al., 2025).

The technological paradigm dimension situates the studies within broader industrial frameworks: Pre-Industry 4.0, Industry 4.0, and Industry 5.0 (Paschek et al., 2022). At the first level, traditional methods with minimal or no digitalisation are employed, such as pure mathematical models, classical optimisation, and manual decision-making. In contrast, the industry 4.0 category utilises advanced technologies, including Artificial Intelligence (AI), the Internet of Things (IoT), automation, and data-driven decision-making. Finally, Industry 5.0 emphasises human-machine collaboration, with a particular focus on sustainability, resilience, ethics in automation, and human-centred optimisation (Guerrero et al., 2025).

Taken together, these dimensions establish the analytical characteristics that will be examined in the following section to address RQ1.

3.5. Classification and analysis of selected publications

Based on the proposed framework, Tables 4, 5 and 6 present a structured classification of the reviewed studies across the nine dimensions, ordered chronologically from the most recent to the earliest. This classification provides a systematic overview of how ML has been applied to planning and scheduling in seaside port operations, thereby addressing RQ1. At the same time, it serves as the basis for the subsequent analysis, which focuses on identifying trends, gaps and potential research directions in the literature, thus contributing to answering RQ2.

The problem type dimension shows that the BAP is the most extensively studied, with several works addressing it dynamically by adjusting berth allocation in real-time in response to vessel arrivals (Cervellera et al., 2021; Cheimanoff et al., 2021; Kolley et al., 2021, 2023; Korekane et al., 2024; Li et al., 2023; Lv et al., 2024; Zhou et al., 2024), with one of them focusing on a multi-terminal and continuous BAP (Yu et al., 2018). In contrast, significantly fewer studies explore the BACAP, with only two works (Oudani et al., 2023; Yu et al., 2018) and the QCSP, with one study addressing it (Long et al., 2025). No studies were identified that specifically address the independent QCAP, nor the fully integrated BACASP, using ML techniques. This gap in the literature suggests that research integrating the three problems with ML remains unexplored.

Turning to the solution approach dimension, 62% of the papers use ML as a complementary tool to enhance exact methods (38%) or metaheuristics (31% of papers). More specifically, these papers are focused on predicting data that is later used by traditional methods, such as the berthing time (Oudani et al., 2023), vessel arrival time (Kolley et al., 2021, 2023), quay crane productivity rate (Oudani et al., 2023), expected profit for vessel handling (Oudani et al., 2023), and delays in vessel arrivals (Cammin et al., 2020; Yu et al., 2018).

Additionally, ML is employed for hyperparameter optimisation in metaheuristic (Cheimanoff et al., 2021) and for selecting the most suitable metaheuristic to address a port operations problem (de León et al., 2017). To achieve this, these studies primarily utilise as ML technique regression-based ML methods, including Adaboost (Oudani et al., 2023), artificial neural networks (Kolley et al., 2023; Yu et al., 2018), back-propagation neural network (Yu et al., 2018), Borda’s method (de León et al., 2017), classification and regression trees (Yu et al., 2018), decision tree regressor (Kolley et al., 2023), Footrule (de León et al., 2017), k-nearest neighbour (de León et al., 2017; Kolley et al., 2021, 2023), linear regression (Kolley et al., 2021, 2023), random forest (Cheimanoff et al., 2021; Yu et al., 2018), support vector machines (Cammin et al., 2020), XGBoost (Oudani et al., 2023). In addition, another paper applies ML as a complementary tool by integrating deep neural networks with exact algorithms, optimising the selection of branching variables and significantly reducing computational times (Korekane et al., 2024).

By contrast, 38% of the analysed papers employ ML as the primary solution approach, meaning that the problem is directly solved using ML algorithms rather than traditional optimisation or metaheuristic techniques. Notably, all these studies focus on addressing a single problem independently, specifically the dynamic BAP (Cervellera et al., 2021; Li et al., 2023; Lv et al., 2024; Zhou et al., 2024) or the QCSP (Long et al., 2025).

Regarding ML techniques, four studies apply Deep Reinforcement Learning (DRL), while one study employs standard Reinforcement Learning (RL). The specific DRL algorithms used include Proximal Policy Optimisation (Long et al., 2025), Double Duelling Deep Q-Network (Li et al., 2023; Zhou et al., 2024), Deep Q-network (Lv et al., 2024), whereas the RL-based approach utilises cross-entropy optimisation (Cervellera et al., 2021).

The purpose of ML dimension reveals that prescriptive applications dominate, representing 46% of the papers, primarily used to identify or recommend optimal decision policies. Predictive ML models follow in frequency, accounting for 38% of the studies, with applications focused on forecasting key operational variables. Learning-based approaches are less common (23%) but have been explored in the context of hyperparameter tuning for metaheuristics (Cheimanoff et al., 2021) and heuristic selection (de León et al., 2017). Meanwhile, no studies explicitly employ ML for purely descriptive or autonomous control purposes, highlighting a potential gap in the literature.

Regarding the type of data used by ML, historical data is the most predominant, with a weight of 54% (Cammin et al., 2020; Kolley et al., 2021, 2023; Long et al., 2025; Oudani et al., 2023; Zhou et al., 2024) followed by using synthesised data with 38% (Kolley et al., 2021; Korekane et al., 2024; Li et al., 2023; Lv et al., 2024; Zhou et al., 2024).

The objective pursued dimension is largely dominated by economic objectives, considered in 77% of the cases, and is focused on minimising the vessel turnaround time at the port. This aspect has been considered alongside resilience and robustness. Regarding resilience, the emphasis is on adapting to variations in arrivals (Cervellera et al., 2021; Lv et al., 2024; Zhou et al., 2024). In terms of robustness, Li et al. (2023) address the integration of uncertainty to maintain stability and de León et al. (2017) highlight the capability to cope with unforeseen scenarios during training. Resilience and robustness are considered together in 46% of the cases.

The context dimension exhibits a nearly balanced distribution, with 54% of the papers adopting deterministic assumptions and 46% incorporating uncertainty. In terms of uncertain contexts, most authors consider the time of arrival of vessels at the port as uncertain (Cervellera et al., 2021; Kolley et al., 2021, 2023; Lv et al., 2024; Yu et al., 2018; Zhou et al., 2024), while Lv et al. (2024) also consider the duration of berthing uncertain.

Looking at the replanning dimension, the review highlights the predominance of event-based approaches (77%), where adjustments are triggered by disruptions. Periodic replanning strategies are rarely explored (23%).

Finally, the technological paradigm dimension reveals that all studies fall under the industry 4.0 framework, reflecting the integration of AI, and more concretely ML, in seaside port operations.

3.6. Cross-dimensional analysis

The cross-dimensional analysis provides additional insights into how ML is applied to seaside port operations by examining the interactions between dimensions. By uncovering methodological consistencies, underexplored combinations, and emerging patterns, this analysis directly contributes to answering RQ2. The most relevant interrelations are presented.

Figure 4 illustrates the relationship between the problem type and the solution approach dimensions, demonstrating that BAP predominates the use of both ML as a complementary solution method, combined with exact/heuristic approaches, and as a primary solution method. Integrated problems remain largely unexplored, with a few available contributions (16% of analysed papers) relying on ML as a complementary tool.

The relation between the solution approach and purpose of ML dimensions (Figure 5) reveals a clear division. When ML is employed as a complementary tool, it is predominantly used for predictive tasks, such as forecasting arrivals (Cammin et al., 2020; Kolley et al., 2021, 2023), berthing times, productivity or benefits (Oudani et al., 2023). Conversely, when ML acts as the primary solution method, RL and DRL consistently appear, supporting prescriptive decision-making.

The link between the ML technique and purpose of ML dimensions (Figure 6) further underlines this pattern: regression-based models are almost exclusively applied for predictive purposes, while reinforcement learning and deep reinforcement learning are consistently linked to prescriptive applications. This finding underlines how methodological choices constrain the functional role of ML in port operations.

In addition, when crossing ML technique and the type of data used by ML dimensions (Figure 7), it emerges that regression-based models rely primarily on historical data and are the only ones using real-time data, whereas RL and DRL approaches are mostly trained on synthetic datasets, complemented by historical and simulated data.

Finally, when analysing the relation between the ML technique and replanning strategy dimensions (Figure 8), a consistent pattern emerges. Regression-based approaches are mainly applied in scenarios with periodic replanning, where decisions are revised at fixed intervals. In contrast, RL and DRL techniques are predominantly associated with event-based replanning.

Taken together, these cross-dimensional insights highlight consistent patterns in how ML has been incorporated into seaside port operations. The predominance of regression-based models for predictive purposes and periodic replanning contrasts with the role of RL/DRL, which are mainly used for prescriptive applications in event-driven contexts. This methodological divide also reflects the problem focus, where ML, as a complementary tool, dominates the BAP, while prescriptive RL/DRL is only beginning to emerge in dynamic BAP and QCSP studies. The limited use of real-time data and the absence of integrated BACASP formulations further reveal gaps that constrain the potential of ML to address the complexity and uncertainty of port operations. These findings suggest promising directions for future research, including the extension of RL/DRL approaches to integrated problems, the incorporation of real-time and streaming data, and the exploration of hybrid methodologies that bridge the predictive strength of regression with the adaptive capabilities of reinforcement learning.

4. Conclusions and future research lines

This paper reviews the existing literature on the application of ML to seaside port operations planning and scheduling, specifically berth allocation, quay crane allocation and quay crane scheduling. Planning problems related to internal transport, yard operations and landside coordination are excluded from the scope of this review. The study first collects the relevant material under the PRISMA protocol and conducts a descriptive analysis to characterise the field. It then proposes a nine-dimensional classification framework, applies it to the selected studies to analyse them and finally does a cross-dimensional analysis to identify patterns and interrelations. Through this process, the study provides answers to the two research questions proposed:

• RQ1. How has ML been applied to planning and scheduling in seaside port operations?

The descriptive analysis shows ML is an emerging field, given the limited number of publications. However, the existing contributions are mainly concentrated in high-impact and multidisciplinary journals (62% of the total, 87.5% of them in Q1). In contrast, the conference papers (38%) are mostly published in proceedings with lower or inconsistent indexation.

The classification analysis shows that research has mainly focused on the BAP, with only a few contributions on the BACAP and a single one on the QCSP. Around half of the BAP studies adopt a dynamic perspective, while the rest are more static. No works apply ML to QCAP or to BACASP. In terms of solution approach, most studies use ML as a complementary tool, mainly for predicting input parameters, tuning metaheuristic parameters or selecting the best heuristic. A smaller proportion uses ML as the primary method, limited to the dynamic BAP and the QCSP, where RL and DRL dominate.

The purpose dimension shows a predominance of prescriptive applications, with predictive and learning applications being less explored. Historical data is the main source, complemented by synthetic data, while real-time data is less used. The objective pursued is predominantly economic, often combined with resilience and robustness. However, environmental and social aspects are less considered, and security is not considered.

Context assumptions are almost evenly split between deterministic and uncertain, mainly linked to arrival times. Event-based replanning dominates, with periodic strategies being rare. All studies remain framed within the Industry 4.0 paradigm, highlighting the integration of AI but also exposing the need to advance towards Industry 5.0 and Society 5.0, where efficiency must be combined with sustainability, resilience and human well-being.

• RQ2. What research gaps, trends, and opportunities emerge in applying ML to seaside port operations?
The cross-dimensional analysis shows that complementary ML is strongly linked to predictive purposes (arrival times, berthing times and crane productivity), while prescriptive applications are almost exclusively associated with RL/DRL as the main solution method. Regression models depend on historical and real-time data and are the only ones linked to periodic replanning, while RL/DRL rely mainly on synthetic and simulated data and are tied to event-based replanning. These findings highlight that regression is used for prediction and periodic adjustments, and RL/DRL is used for prescriptive decision-making in dynamic and event-driven contexts.

Based on the results of this review, several future research directions can be outlined. Firstly, there is a need to explore hybrid approaches that combine ML with exact optimisation models and advanced heuristics in order to address integrated problems such as BACASP. Such approaches could overcome the limitations of exact methods while ensuring scalability and maintaining solution quality.

Secondly, sustainability-oriented objectives should be incorporated more explicitly into models to support seaside port operations. Metrics for emissions, energy efficiency and environmental impact, as already demonstrated in recent studies, could allow decision support systems to balance operational efficiency with environmental performance.

Finally, future research should focus on extending solutions to dynamic and real-time contexts. The integration of real-time data with adaptive models capable of anticipating disruptions and readjusting decisions would enable continuous replanning, ensuring the robustness and resilience of port operations under uncertainty.

Funding

This work has been partially developed in the framework of the TALON project funded by the European Union’s Horizon program under grant agreement No 101070181. The views and opinions expressed are those of the author(s) and do not necessarily reflect those of the EU, which is not responsible.

Conflicts of interest

The authors declare no conflicts of interest.

References

ALJUAID, A. M., KOUBÂA, M., AMMAR, M. H., KAMMOUN, K., & HACHICHA, W. (2024). Mathematical Programming Formulations for the Berth Allocation Problems in Container Seaport Terminals. Logistics, 8(2). https://doi.org/10.3390/logistics8020050

AL SAMROUT, M., SBIHI, A., & YASSINE, A. (2024). An improved genetic algorithm for the berth scheduling with ship-to-ship transshipment operations integrated model. Computers & Operations Research, 161, 106409. https://doi.org/10.1016/J.COR.2023.106409

ARENA, S., FLORIAN, E., SGARBOSSA, F., SØLVSBERG, E., & ZENNARO, I. (2024). A conceptual framework for machine learning algorithm selection for predictive maintenance. Engineering Applications of Artificial Intelligence, 133, 108340. https://doi.org/10.1016/J.ENGAPPAI.2024.108340

BACIULIENE, V., BILAN, Y., NAVICKAS, V., & CIVIN, L. (2023). The Aspects of Artificial Intelligence in Different Phases of the Food Value and Supply Chain. Foods, 12(8), 1654. https://doi.org/10.3390/foods12081654

BIERWIRTH, C., & MEISEL, F. (2010). A survey of berth allocation and quay crane scheduling problems in container terminals. European Journal of Operational Research, 202(3), 615–627. https://doi.org/10.1016/J.EJOR.2009.05.031

BOLUDA-PRIETO, M., ESTESO, A., ALEMANY-DÍAZ, M. DEL M. E., & ORTIZ, Á. (2025). Systematic Review of Machine Learning Applications in Port Operations Optimization (pp. 186–192). https://doi.org/10.1007/978-3-032-10126-6_30

CAMMIN, P., SARHANI, M., HEILIG, L., & VOß, S. (2020). Applications of Real-Time Data to Reduce Air Emissions in Maritime Ports. In Design, User Experience, and Usability. Case Studies in Public and Personal Interactive Systems (pp. 31–48). https://doi.org/10.1007/978-3-030-49757-6_3

CASTILLA-RODRÍGUEZ, I., EXPÓSITO-IZQUIERDO, C., MELIÁN-BATISTA, B., AGUILAR, R. M., & MORENO-VEGA, J. M. (2020). Simulation-optimization for the management of the transshipment operations at maritime container terminals. Expert Systems with Applications, 139, 112852. https://doi.org/10.1016/J.ESWA.2019.112852

CERVELLERA, C., GAGGERO, M., & MACCIÒ, D. (2021). Policy Optimization for Berth Allocation Problems. 2021 International Joint Conference on Neural Networks (IJCNN), 1–6. https://doi.org/10.1109/IJCNN52387.2021.9533891

CHEIMANOFF, N., FONTANE, F., KITRI, M. N., & TCHERNEV, N. (2021). A reduced VNS based approach for the dynamic continuous berth allocation problem in bulk terminals with tidal constraints. Expert Systems with Applications, 168, 114215. https://doi.org/https://doi.org/10.1016/j.eswa.2020.114215

CORRECHER, J. F., PEREA, F., & ALVAREZ-VALDES, R. (2024). The berth allocation and quay crane assignment problem with crane travel and setup times. Computers & Operations Research, 162, 106468. https://doi.org/10.1016/j.cor.2023.106468

DE LEÓN, A. D., LALLA-RUIZ, E., MELIÁN-BATISTA, B., & MARCOS MORENO-VEGA, J. (2017). A Machine Learning-based system for berth scheduling at bulk terminals. Expert Systems with Applications, 87, 170–182. https://doi.org/10.1016/J.ESWA.2017.06.010

DING, Y., CHEN, K., TANG, M., YUAN, X., HEILIG, L., SONG, J., FANG, H., & TIAN, Y. (2023). An efficient and eco-friendly operation mode for container transshipments through optimizing the inter-terminal truck routing problem. Journal of Cleaner Production, 430, 139644. https://doi.org/10.1016/J.JCLEPRO.2023.139644

ESTESO, A., ALEMANY, M. M. E., & ORTIZ, A. (2018). Conceptual framework for designing agri-food supply chains under uncertainty by mathematical programming models. International Journal of Production Research, 56(13), 4418–4446. https://doi.org/10.1080/00207543.2018.1447706

FRANÇOIS-LAVET, V., HENDERSON, P., ISLAM, R., BELLEMARE, M. G., & PINEAU, J. (2018). An Introduction to Deep Reinforcement Learning. Foundations and Trends® in Machine Learning, 11(3–4), 219–354. https://doi.org/10.1561/2200000071

GUERRERO, B., MULA, J., & POLER, R. (2025). Sustainable operations management towards Industry 5.0. Dirección y Organización, (85), 85–92. https://doi.org/10.37610/85.692

HERRERA, F. (2025). Intelligent organizational engineering driven by human-AI collaboration and explainable AI to increase productivity. Dirección y Organización, (87), 5–14. https://doi.org/10.37610/87.702

KARIMI-MAMAGHAN, M., MOHAMMADI, M., MEYER, P., KARIMI-MAMAGHAN, A. M., & TALBI, E. G. (2022). Machine learning at the service of meta-heuristics for solving combinatorial optimization problems: A state-of-the-art. European Journal of Operational Research, 296(2), 393–422. https://doi.org/10.1016/J.EJOR.2021.04.032

KOLLEY, L., RÜCKERT, N., & FISCHER, K. (2021). A Robust Berth Allocation Optimization Procedure Based on Machine Learning. In Logistics Management. Lecture Notes in Logistics (pp. 107–122). https://doi.org/10.1007/978-3-030-85843-8_7

KOLLEY, L., RÜCKERT, N., KASTNER, M., JAHN, C., & FISCHER, K. (2023). Robust berth scheduling using machine learning for vessel arrival time prediction. Flexible Services and Manufacturing Journal, 35(1), 29–69. https://doi.org/10.1007/s10696-022-09462-x

KOREKANE, S., NISHI, T., TIERNEY, K., & LIU, Z. (2024). Neural network assisted branch and bound algorithm for dynamic berth allocation problems. European Journal of Operational Research, 319(2), 531–542. https://doi.org/10.1016/J.EJOR.2024.06.040

LE, T.-H., MINH, T. D., & HOA, N. T. N. (2024). Elitist Strategy Genetic Algorithm-Based Planning Optimization Deriving for Smart Port Decision Support System. In Intelligence of Things: Technologies and Applications. ICIT 2024. Lecture Notes on Data Engineering and Communications Technologies (Vol. 230, pp. 35–44). https://doi.org/10.1007/978-3-031-75596-5_4

LECUN, Y., BENGIO, Y., & HINTON, G. (2015). Deep learning. Nature, 521(7553), 436–444. https://doi.org/10.1038/nature14539

LI, B., YANG, C., & YANG, Z. (2023). Multiple Container Terminal Berth Allocation and Joint Operation Based on Dueling Double Deep Q-Network. Journal of Marine Science and Engineering, 11(12). https://doi.org/10.3390/jmse11122240

LONG, L. N. B., YOU, S. S., CUONG, T. N., & KIM, H. S. (2025). Optimizing quay crane scheduling using deep reinforcement learning with hybrid metaheuristic algorithm. Engineering Applications of Artificial Intelligence, 143, 110021. https://doi.org/10.1016/J.ENGAPPAI.2025.110021

LORENTE-LEYVA, L. L., ALEMANY, M. M. E., & PELUFFO-ORDÓÑEZ, D. H. (2024). A conceptual framework for the operations planning of the textile supply chains: Insights for sustainable and smart planning in uncertain and dynamic contexts. Computers & Industrial Engineering, 187, 109824. https://doi.org/10.1016/J.CIE.2023.109824

LORENTE-LEYVA, L. L., ALEMANY, M. M. E., & PELUFFO-ORDÓÑEZ, D. H. (2025). Optimization of Textile SCs Design and Planning: A Bibliometric Analysis. In Lecture Notes on Data Engineering and Communications Technologies (pp. 16–22). https://doi.org/10.1007/978-3-031-82334-3_4

LV, Y., ZOU, M., LI, J., & LIU, J. (2024). Dynamic berth allocation under uncertainties based on deep reinforcement learning towards resilient ports. Ocean & Coastal Management, 252, 107113. https://doi.org/10.1016/J.OCECOAMAN.2024.107113

MANSOURSAMAEI, M., MORADI, M., GONZÁLEZ-RAMÍREZ, R. G., & LALLA-RUIZ, E. (2023). Machine Learning for Promoting Environmental Sustainability in Ports. Journal of Advanced Transportation, 2023, 1–17. https://doi.org/10.1155/2023/2144733

MATEO-CASALÍ, M. Á., PABLO FIESCO, J., ANDRES, B., & POLER, R. (2025). Inteligencia Artificial para el soporte a la toma de decisiones en el ciclo de vida de los equipos industriales. Dirección y Organización, (85), 67–84. https://doi.org/10.37610/85.691

MEKKAOUI, S. E., BENABBOU, L., & BERRADO, A. (2020). A Systematic Literature Review of Machine Learning Applications for Port’s Operations. 2020 5th International Conference on Logistics Operations Management (GOL), 1–5. https://doi.org/10.1109/GOL49479.2020.9314756

OUDANI, M., SEBBAR, A., ZKIK, K., & BELHADI, A. (2023). A Prescriptive Analytics Approach for Port Logistics Planning. 2023 9th International Conference on Control, Decision and Information Technologies (CoDIT), 77–81. https://doi.org/10.1109/CoDIT58514.2023.10284173

PASCHEK, D., LUMINOSU, C.-T., & OCAKCI, E. (2022). Industry 5.0 Challenges and Perspectives for Manufacturing Systems in the Society 5.0 (pp. 17–63). https://doi.org/10.1007/978-981-16-7365-8_2

PEREIRA, M. T., ROCHA, N., SILVA, F. G., MOREIRA, M. Â. L., ALTINKAYA, Y. O., & PEREIRA, M. J. (2025). Process Optimization in Sea Ports: Integrating Sustainability and Efficiency Through a Novel Mathematical Model. Journal of Marine Science and Engineering, 13(1). https://doi.org/10.3390/jmse13010119

SEURING, S., & MÜLLER, M. (2008). From a literature review to a conceptual framework for sustainable supply chain management. Journal of Cleaner Production, 16(15), 1699–1710. https://doi.org/10.1016/J.JCLEPRO.2008.04.020

TYAGI, K., RANE, C., HARSHVARDHAN, & MANRY, M. (2022). Regression analysis. Artificial Intelligence and Machine Learning for EDGE Computing, 53–63. https://doi.org/10.1016/B978-0-12-824054-0.00007-1

VELASTEGUI, R., POLER, R., & DÍAZ-MADROÑERO, M. (2023). Aplicación de algoritmos de aprendizaje automático a sistemas robóticos multiagente para la programación y control de operaciones productivas y logísticas: una revisión de la literatura reciente. Dirección y Organización, (80), 60–70. https://doi.org/10.37610/dyo.v0i80.643

WALLER, M. A., & FAWCETT, S. E. (2013). Data Science, Predictive Analytics, and Big Data: A Revolution That Will Transform Supply Chain Design and Management. Journal of Business Logistics, 34(2), 77–84. https://doi.org/10.1111/jbl.12010

XIE, X., JI, B., & YU, S. S. (2025). A Variable Neighborhood Search Algorithm for the Integrated Berth Allocation and Quay Crane Assignment Problem. Sustainability, 17(9). https://doi.org/10.3390/su17094022

YU, J., TANG, G., SONG, X., YU, X., QI, Y., LI, D., & ZHANG, Y. (2018). Ship arrival prediction and its value on daily container terminal operation. Ocean Engineering, 157, 73–86. https://doi.org/10.1016/J.OCEANENG.2018.03.038

YUE, M., WANG, Y., GUO, S., DAI, L., & HU, H. (2025). A multi-objective optimization study of berth scheduling considering shore side electricity supply. Ocean & Coastal Management, 261, 107500. https://doi.org/10.1016/J.OCECOAMAN.2024.107500

ZHANG, C., JURASCHEK, M., & HERRMANN, C. (2024). Deep reinforcement learning-based dynamic scheduling for resilient and sustainable manufacturing: A systematic review. Journal of Manufacturing Systems, 77, 962–989. https://doi.org/10.1016/J.JMSY.2024.10.026

ZHOU, Q., CAO, X., & WANG, P. (2024). Deep Reinforcement Learning for Dynamic Berth Allocation with Random Ship Arrivals. 2024 6th International Conference on Data-Driven Optimization of Complex Systems (DOCS), 799–805. https://doi.org/10.1109/DOCS63458.2024.10704490

_______________________________

¹ Research Centre on Production Management and Engineering (CIGIP), Universitat Politècnica de València, Camino de Vera S/N, Valencia 46022, Spain. Email: mboluda@cigip.upv.es ORCID: 0009-0009-6615-1489

² Research Centre on Production Management and Engineering (CIGIP), Universitat Politècnica de València, Camino de Vera S/N, Valencia 46022, Spain. Email: aesteso@cigip.upv.es ORCID: 0000-0003-0379-8786

³ Research Centre on Production Management and Engineering (CIGIP), Universitat Politècnica de València, Camino de Vera S/N, Valencia 46022, Spain. Email: mareva@omp.upv.es ORCID: 0000-0002-0992-8441

⁴ Research Centre on Production Management and Engineering (CIGIP), Universitat Politècnica de València, Camino de Vera S/N, Valencia 46022, Spain. Email: aortiz@cigip.upv.es ORCID: 0000-0001-5690-0807