Ports are actively pursuing greater operational efficiency to handle the increasing global flow of goods, while simultaneously improving the energy efficiency of their operations to comply with new environmental regulations. As a result, innovation-leading ports have begun to recognize the potential of digital twins to monitor, coordinate, and optimize port processes, enabling energy savings and reductions in both costs and CO2 emissions. Although digital twins have gained significant momentum in other domains, such as smart manufacturing and aerospace, their adoption in ports remains challenging. This can be explained by the multi-stakeholder nature of ports and the high complexity of their interconnected processes, requiring decision-making across organizational boundaries.

Grounded in the port context, this thesis examines what constitutes a digital twin, proposes a framework to assess the maturity of existing port digital twins, and develops modeling and explainable AI-enabled decision support components for port and maritime operations. These components span seaside, quay, yard, and gate processes and can serve as building blocks of future port digital twin implementations. The thesis consists of six papers:

Paper 1 provides an in-depth literature review of digital twins across multiple domains and transfers insights from these to the port domain. The paper outlines how digital twins can enhance operational efficiency and support energy savings in ports. It also identifies the characteristics and design requirements that a port-specific digital twin must fulfill. Based on these findings, the paper proposes a tailored definition of a digital twin for the port domain.

Paper 2 discusses how digital twins’ maturity can be assessed within six maturity levels and presents milestones for their implementation. Notably, Interoperability is identified as the highest maturity level,as the numerous stakeholders and their respective digital twins must work together to reach a coordinated system of systems performance. Using this assessment demonstrates that only a few innovation-leading ports have developed sophisticated digital twinning solutions so far.

Paper 3 focuses on container retrieval, balancing two competing objectives: minimizing yard crane moves and adhering to tight truck scheduling. This reflects the conflicting perspectives of different stakeholders in the port context. The provided optimization model and heuristic algorithm demonstrate that addressing both problems simultaneously may result in reduced efficiency of the individual objectives. However, from a systems perspective, this approach leads to higher overall port efficiency.

Paper 4 examines quay cranes at the system level by developing an explainable AI framework to predict whether a quay crane will experience a breakdown during vessel operations. Using monitoring data, operational data, and weather observations, the study identifies how operational intensity, hoist-related warning patterns, and environmental conditions jointly influence the likelihood of a breakdown. This system-level predictive capability enhances situational awareness and enables early identification of disruptions.

Paper 5 builds on Paper 4 by focusing on the prediction of individual critical error events. Rather than assessing the overall likelihood of a breakdown, the model identifies which error type is likely to occur next and estimates its timing. Using eXtreme Gradient Boosting with lagged error sequences, operational data, and weather conditions, the study offers component-level insights that complement the systemlevel prediction in Paper 4 and support more targeted maintenance interventions.

Paper 6 expands the perspective beyond ports by analyzing fuel consumption in inland ferry operations using GPS-derived trip legs and journeys enriched with environmental data. Combining unsupervised clustering to uncover operational patterns with supervised learning and SHAP-based explainability, the study identifies operational speed as the dominant driver of fuel consumption and links consumption patterns to individual captains’ driving behavior. This contributes to maritime decision-making by enabling targeted interventions such as eco-driving strategies.

Together, these six papers contribute a conceptual grounding of port digital twins, provide a tool for their assessment, and provide modeling components to aid in port and maritime decision-making.

Hamnar strävar aktivt efter ökad operativ effektivitet för att hantera det ökande globala varuflödet, samtidigt som de strävar efter att förbättra energieffektiviteten. Som ett resultat har ledande hamnar börjat se potentialen hos digitala tvillingar för att skapa överblick samt koordinera och optimera processer i hamnen. Målet med användningen av digitala tvillingar är energibesparingar samt minskning av kostnader och CO2-utsläpp. Medan digitala tvillingar har använts inom andra områden såsom tillverknings-, flyg- och rymdindustrin, har införandet i hamnar varit jämförelsevist långsamt. Detta kan förklaras, bland annat, av hamnens många olika involverade aktörer och den höga komplexiteten i de ofta sammanlänkade hamnprocesserna.

Därför fokuserar denna avhandling, med utgångspunkt i hamnkontexten, vad som utgör en digital tvilling, presenterar egenskaper för olika mognadsnivåer hos befintliga digitala tvillingar, och introducerar modellerings- och beslutsstödskomponenter baserade på förklarbar AI för hamn- och maritima operationer. Dessa komponenter omfattar kustnära processer, kajoperationer, yard-processer och gate-funktioner, och kan fungera som byggstenar i framtida digitala tvillingar för hamnar. Avhandlingen består av sex artiklar:

Artikel 1 bygger på en omfattande litteraturöversikt, inom vilken digitala tvillingar för olika områden studeras ingående för att överföra insikter från dessa till hamndomänen. Detta resulterar i en presentation av vad som utgör en hamns digitala tvilling och de krav som en hamns digitala tvilling måste uppfylla, tillsammans med en diskussion om hur digitala tvillingar i hamnar kan bidra till energibesparingar.

Artikel 2 presenterar ett ramverk för hur mognaden hos digitala tvillingar kan bedömas baserat på sex mognadsnivåer och presenterar milstolpar för deras implementering. Noterbart är att interoperabilitet identifieras som den högsta mognadsnivån, eftersom de många intressenterna och deras respektive digitala tvillingar måste koordineras för att nå en fungerande system-av-systemnivå. Genom att använda denna bedömning visar det sig att endast några få innovationsledande hamnar hittills har utvecklat sofistikerade digitala tvillinglösningar.

Artikel 3 fokuserar på containerupphämtning med hänsyn till två konkurrerande mål: att minimera energikrävande kranrörelser och att hålla planerade tider för lastbilar. Detta speglar de potentiellt motstridiga perspektiven hos olika intressenter i hamnkontexten. Den utvecklade optimeringsmodellen och algoritmen visar att gemensam hantering av båda dessa mål kan leda till minskad effektivitet för de respektive individuella målen, men ökad effektivitet från ett systemperspektiv för hamnen som helhet.

Artikel 4 studerar kajkranar på systemnivå genom att utveckla ett förklarbart AI-ramverk för att förutsäga om en kajkran kommer att drabbas av ett driftstopp under ett fartygsanlöp. Genom att använda övervakningsdata från kranarna, operativa data från terminalen och meteorologiska observationer identifierar studien hur operativ belastning, hoist-relaterade varningar och väderförhållanden gemensamt påverkar sannolikheten för driftstopp. Modellen förbättrar situationsmedvetenheten och möjliggör tidigare identifiering av störningar.

Artikel 5 bygger vidare på Artikel 4 genom att fokusera på prediktion av enskilda kritiska felhändelser. I stället för att uppskatta sannolikheten för ett övergripande driftstopp förutser modellen vilken feltyp som sannolikt inträffar härnäst och när detta sker. Med hjälp av eXtreme Gradient Boosting i kombination med sekvenser av tidigare fel, aktuella operativa data och väderförhållanden tillhandahåller studien komponentnivåinsikter som kompletterar systemnivåanalysen i Artikel 4 och möjliggör mer riktade och tidskritiska underhållsåtgärder.

Artikel 6 breddar avhandlingens fokus till maritima operationer genom att analysera bränsleförbrukning i färjetrafik baserat på GPS- data och kompletterande miljödata. Genom att kombinera oövervakad klustring för att identifiera återkommande operativa mönster med övervakade prediktionsmodeller och SHAP-baserad förklarbarhet visar studien att fartygshastighet är den dominerande faktorn bakom bränsleförbrukning. Analysen kopplar också bränsleförbrukningsmönster till individuella befälhavares beteenden och möjliggör riktade åtgärder, såsom eco-driving.

Tillsammans bidrar dessa sex artiklar med en konceptuell grund för digitala tvillingar i hamnar, ett verktyg för att bedöma mognaden hos befintliga lösningar samt ett antal modelleringskomponenter som kan stödja datadrivet och förklarbart beslutsfattande i både hamn- och maritima verksamheter.

Electrification is seen as a key pathway toward more sustainable transport systems. This paper examines a Swedish island ferry conversion from diesel to battery electric propulsion by integrating quantitative and qualitative insights: (i) a route-level efficiency assessment to map energy savings across comparable service legs; (ii) an explainable gradient boosting model (SHAP) to quantify the operational and environmental drivers of trip-level energy use; and (iii) interviews with 55 passengers and three captains to capture perceived benefits, operational constraints, and adaptation strategies. The model achieves high predictive accuracy (R² = 0.97), showing that vessel speed and propulsion type dominate variation in energy intensity, while route geometry and wind contribute smaller but systematic effects. Electrification reduces energy intensity across routes by 10-87%, driven primarily by higher conversion efficiency. Interview results reveal improved onboard comfort, tighter operational margins around charging and schedule adherence, and heating loads as critical constraints in colder seasons. The combined use of EMS data, interpretable ML, and stakeholder interviews provides a rare, system-level perspective on the technical and human factors shaping electric ferry performance. Viewed through a multi level perspective, the findings indicate that electric ferries scale most effectively on short, predictable routes when infrastructure, timetables, and procurement incentives align with the operational characteristics of electric propulsion.

Enhancing fuel efficiency in ferry operations is essential for reducing emissions and advancing maritime sustainability. This study presents a data-driven framework that uses second-level GPS data enriched with operational and environmental variables to identify and explain fuel consumption patterns. Vessel movements are segmented into trip legs and journeys, and operational metrics such as speed, wind exposure, and fuel use are computed. A hybrid machine learning approach combines unsupervised clustering to detect recurring operational patterns with gradient boosting models and explainable methods to quantify feature impacts. The framework achieves strong performance, with a cluster classification accuracy of 94 percent and a coefficient of determination of 0.97 for fuel prediction. Results indicate that operational speed is the dominant driver of fuel consumption, while analysis of captain assignments reveals the influence of human factors. The proposed framework provides actionable insights for speed management and operational optimization, enabling cost-effective emission reductions in ferry services.  

Efficient port operations depend on the disruption free operation of quay cranes (QCs), which transfer containers between vessels and internal trucks. As global container through put rises, QCs face increased pressure, resulting in accelerated wear and tear. This can lead to QC downtime, which could interrupt the entire chain of port operations. Therefore, timely identification and prediction of critical errors is essential to enable timely maintenance to lower the risk of downtime. This study utilizes two years of QC monitoring data, enriched with weather conditions and terminal operational context, alongside twenty critical error events identified by the terminal operator. The goal is to predict the occurrence and timing of these critical errors through a three-stage machine learning model. The first stage predicts the type of the next critical event based on historical error patterns, warnings, and contextual data. The second stage estimates a time window in which the event will occur. The third stage refines timing predictions when more than one hour remains. The first two stages are formulated as multiclass classification problems, and the third as a regression task. All stages utilize eXtreme Gradient Boosting (XGBoost). SHapley Additive exPlanations (SHAP) are used to identify influential features. Results show that the model predicts the next critical error type with 83% accuracy and its immediacy with 71% accuracy. However, approximating the timing of events anticipated to occur beyond one hour remains challenging. These findings support proactive maintenance planning and operational adjustments, helping port operators mitigate disruptions and enhance QC reliability.

Quay cranes (QCs) play a vital role in ship-to-shore operations, enabling the seamless transfer of cargo between sea and land. However, increasing trade volumes require faster and more cost-effective container handling, exerting significant pressure on QCs and leading to greater wear on critical components such as wires, hoists, and rope clamps. While operations research has explored maintenance scheduling to improve terminal performance, comparatively little work has examined how machine learning can exploit the growing volume of QC monitoring and operational data to predict breakdowns before they occur. This study contributes to this area by integrating terminal operations data, QC monitoring logs, and meteorological observations into a unified analytical framework. We employ explainable artificial intelligence (XAI), using both global and local SHapley Additive exPlanations (SHAP) to identify the operational and environmental factors most strongly associated with QC failures and to illustrate concrete, instance-level examples of how specific conditions contribute towards breakdowns. In parallel, we develop a robust machine learning pipeline built around nested cross-validation to assess the predictive capability of multiple classifiers for forecasting QC breakdowns. Our XAI analysis reveals that breakdown risk is closely linked to QC working time, the distribution of moves across simultaneously operating QCs, hoist overload and trolley alignment warnings, and adverse weather conditions. Among the evaluated models, LightGBM achieved the highest predictive accuracy, reaching up to 83% in identifying breakdown-prone scenarios. These findings demonstrate the feasibility and value of data-driven predictive maintenance for QCs, providing insights that support safer, more reliable, and more efficient terminal operations. 

Urban transportation planning faces increasing complexity as cities seek to optimize mobility systems without extensive historical data. This paper presents CI-SGNN (City-Invariant Spatial Graph Neural Network), a novel framework for cross-city bike-sharing demand prediction that leverages Point of Interest (POI) distributions and spatial attention mechanisms. Our approach addresses the critical challenge of predicting categorical mobility demand in new urban environments by learning transferable relationships between urban amenities and travel patterns from source cities. The framework integrates OpenStreetMap POI features with GNNs, enabling zero-shot transfer learning across diverse metropolitan areas. We formulate demand prediction as a multi-class classification problem, categorizing origin-destination pairs into five demand levels. Experimental validation using real CitiBike data from Manhattan and Washington DC demonstrates superior performance, achieving 72.4% accuracy, which overperforms state-of-the-art base-lines. The attention-based spatial aggregation mechanism effectively captures inter-zone dependencies. Our results demonstrate successful zero-shot adaptation capabilities, enabling practical deployment for bike-sharing infrastructure planning in cities lacking historical mobility data using only publicly available urban features. 

Modelling of long-distance travel is often overlooked in national transport models, which is problematic, for example, when assessing the impacts of large railway investments involving a substantial share of cross-border travel. One kind of long-distance travel is visitors’ trips. In this paper, we estimate trip generation, mode and destination choice models for visitors’ long-distance trips to Sweden using survey data on incoming visitors collected at the Swedish border and major airports and ferry terminals. We describe the modelling challenges faced, such as having origin data on the country level only. We explain visitor choice of the destination zone in Sweden using point-of-interest data from open-street-map. Travel time and travel cost elasticities are calculated for a railway investment scenario. The results show that reducing door-to-door travel time by rail is more important than reducing travel costs if the goal is to switch visitors’ trips from air and car to rail.

Digital twinning of vehicles is an iconic application of digital twins, as the concept of twinning dates back to the twinning of NASA space vehicles. Although digital twins (DTs) in the automotive industry have been recognized for their ability to improve efficiency in design and manufacturing, their potential to enhance land vehicle operation has yet to be fully explored. Most existing DT research on vehicle operations, aside from the existing body of work on autonomous guided vehicles (AGVs), focuses on electrified passenger cars. However, the use and value of twinning varies depending on the goal, whether it is to provide cost-efficient and sustainable freight transport without disruptions, sustainable public transport focused on passenger well-being, or fully autonomous vehicle operation. In this context, DTs are used for a range of applications, from real-time battery health monitoring to enabling fully autonomous vehicle operations. This leads to varying requirements, complexities, and maturities of the implemented DT solutions. This paper analyzes recent trends in DT-driven efficiency gains for freight, public, and autonomous vehicles and discusses their required level of maturity based on a maturity tool. The application of our DT maturity tool reveals that most DTs have reached level 3 and enable real-time monitoring. Additionally, DTs of level 5 already exist in closed environments, allowing for restricted autonomous operation. 

In recent years, autonomous delivery robots have gained tremendous momentum in last-mile logistics operations due to their potential to reduce costs, emissions, and congestion while providing access to narrow roads. However, their operation is restricted by distance, capacity, and maintenance needs. These limitations necessitate a combination of robot and van operations, wherein vans deliver non-bulky parcels to robots while performing bulky and distant deliveries themselves. Building on insights from the Helsingbotica project in Sweden, which includes key stakeholders such as VTI, Hugo, Apotea, BEST Transport, and the City of Helsingborg, this paper aims to evaluate the potential of a joint robot-van setup by analyzing the cost and energy savings of this approach at different geographic scales and configurations. We propose a universally applicable parcel demand estimation framework that uses OpenStreetMap building density and road network data to simulate parcel demand and create meaningful service zones using K-Medoids clustering. The estimated demand and constructed zones serve as input for an integer programming model that assigns parcels to robots and vans in a cost- and energy-saving manner, considering restrictions such as distance and parcel weight. The model’s parameters are calibrated based on structured workshops with industry partners. Our results demonstrate that integrating autonomous delivery robots can reduce operational costs by up to 57% and energy consumption by up to 42%, depending on the configuration. Thus, this study concludes that integrating robots into last-mile delivery can enhance the flexibility and efficiency of logistics service providers, offering a sustainable solution for urban deliveries. 

The transition towards sustainable and intelligent collective transport systems requires new approaches to monitoring and managing vehicle cabin environments. Within this context, the TwinAIR Horizon Europe EU-funded project collaborated with the Ride the Future platform, installing sensors and delivering a Vehicle Cabin Digital Twin (VCDT). Doing so required the installation of in-cabin and peripheral sensors. However, although sensors provide essential physical cabin comfort measurements of temperature, humidity, CO₂, particulate matter, and volatile organic compounds, the sensor values alone cannot be relied upon for direct decision-making, since they are prone to noise, failures, and limited spatial coverage. Such limitations motivate the integration of computational models that replicate/augment physical sensing into digital twins (DTs). Such models are referred to as 'virtual sensors'.