Artificial Intelligence and Intelligent Transport Applications

Introduction and Context

Artificial intelligence (AI) is changing multimodal transportation from a largely planned and reactive activity into a data-driven, adaptive system. Logistics providers increasingly combine road, rail, maritime, air, terminal, warehouse and information flows. AI supports this integration by detecting patterns in large and heterogeneous datasets, estimating future conditions, recommending actions and, in some cases, executing decisions automatically.

In multimodal transport, the value of AI does not arise simply from using a sophisticated algorithm. It arises when reliable data, operational processes, human expertise and governance are connected. A highly accurate prediction that is delivered too late, cannot be explained to a dispatcher, or ignores contractual and safety constraints may have little operational value. Conversely, a moderately accurate model that is trusted, monitored and embedded in decision processes can produce substantial benefits.

This chapter examines machine learning for forecasting, estimated time of arrival (ETA) prediction, disruption detection, dynamic routing, asset utilization and predictive maintenance. It also evaluates autonomous vehicles, intelligent terminals, explainability, bias and human oversight.

Learning Objectives

By the end of this chapter, students should be able to:

1. Explain how AI and machine learning support multimodal transport decisions.
2. Select appropriate forecasting, classification, anomaly-detection and optimization approaches.
3. Interpret common performance measures for ETA, disruption and maintenance models.
4. Evaluate dynamic routing and asset-utilization applications.
5. Assess autonomous vehicles and intelligent terminals from operational, safety and governance perspectives.
6. Design an AI deployment approach incorporating explainability, fairness, cybersecurity and human oversight.

1. AI as a Decision Capability in Multimodal Logistics

AI refers broadly to computational systems that perform tasks associated with perception, learning, reasoning, prediction or decision-making. Machine learning (ML) is a subset of AI in which a model learns relationships from data rather than being programmed entirely through fixed rules.

A multimodal AI system typically contains five layers:

Transport data are often distributed across carriers, ports, customs agencies, terminal operators and customers. They also differ in frequency, quality and meaning. A vessel position may be updated every few minutes, while a rail schedule may be revised only when a formal event occurs. Consequently, data integration and semantic consistency are as important as model selection.

An AI project should begin with a decision, not with an algorithm. The design question is: Which operational decision will improve, for whom, under what constraints, and at what acceptable risk? This prevents the common error of developing a prediction without defining its use.

2. Forecasting Transport Demand and Workload

Forecasting supports capacity planning, procurement, labor scheduling, fleet allocation and terminal design. Relevant targets include shipment volumes, container arrivals, lane demand, empty-equipment requirements, gate transactions and warehouse workload.

A time-series forecast may be represented as:

Here, ŷ_{t+h} is the forecast for period t+h, y represents the historical target variable, x represents explanatory variables such as promotions, weather or economic indicators, h is the forecast horizon, and f is the estimated forecasting function. The formulation applies when historical observations and relevant explanatory variables are available and the relationship is sufficiently stable for the intended horizon.

Traditional methods such as moving averages, exponential smoothing and autoregressive models remain useful, especially when data are limited and interpretability is important. ML methods, including gradient-boosted trees and recurrent or transformer-based neural networks, can represent nonlinear relationships and interactions. However, complexity does not guarantee superior performance. Models must be compared against simple baselines using time-ordered validation rather than random splitting, which can leak future information into training data.

Common measures include:

MAE is expressed in the units of the target and gives equal linear weight to errors. RMSE has the same units but penalizes large errors more strongly. n is the number of observations, y_i is the observed value and ŷ_i is the prediction. Both measures assume paired observations and are inappropriate if missing values are handled inconsistently. Percentage measures can be misleading when actual demand is zero or close to zero.

Forecasting should include prediction intervals where decisions depend on uncertainty. A terminal may need to plan for a high-demand quantile rather than the expected value. The operational question is therefore not only “What volume is likely?” but also “What range of volumes is plausible, and what is the cost of being under- or over-prepared?”

3. ETA Prediction and Arrival-Time Reliability

ETA prediction estimates when a shipment, vehicle, vessel, train or container will reach a defined milestone. In multimodal systems, ETA is not a single physical quantity: it depends on the milestone definition, such as port arrival, berth availability, terminal release, customs clearance or final delivery.

Useful features include current position, speed, route, historical travel time, weather, congestion, appointment schedules, border conditions, terminal queues, transshipment plans and mode connections. The model should distinguish between information known at prediction time and information learned later. Otherwise, performance will be overstated.

An ETA model can produce an expected arrival time and an interval. If the predicted arrival is 16:00 and the 90% interval is 15:20–17:10, a planner can decide whether a missed connection is sufficiently likely to justify intervention. Calibration is important: among predictions claiming 90% coverage, approximately 90% of actual arrivals should fall within the stated intervals over a suitable evaluation sample.

ETA accuracy may be reported using MAE in minutes. Operational usefulness also requires measuring bias, late-arrival recall and performance by lane, carrier, mode, geography and weather condition. A model that performs well on major corridors but poorly on remote destinations may create inequitable service outcomes.

4. Disruption Detection and Anomaly Analysis

Disruptions include severe weather, labor stoppages, infrastructure failure, cyber incidents, missed connections, equipment shortages, unusual dwell times and unexplained route deviations. AI can detect disruptions earlier by combining event streams that no individual operator can continuously monitor.

Supervised classification is appropriate when historical examples of disruptions are labelled. Its output may be a probability such as P(disruption | current observations). Classification performance can be assessed using precision, recall and the F1 score:

TP, FP and FN denote true positives, false positives and false negatives. Precision answers how many alerts were meaningful; recall answers how many actual disruptions were detected. The preferred balance depends on consequences. In safety-critical monitoring, missing a serious event may be more costly than generating additional alerts.

When labelled disruptions are scarce, unsupervised or semi-supervised anomaly detection can identify observations that differ from normal patterns. Examples include unusually long terminal dwell, a refrigerated container temperature trajectory outside its normal profile, or a vehicle consuming substantially more fuel than expected. An anomaly is not automatically a disruption; it is a signal requiring interpretation.

Alert systems should use severity levels, persistence rules and escalation procedures. Excessive false alarms produce alert fatigue, while delayed alerts reduce the time available for mitigation. The system should preserve the evidence supporting an alert so that staff can validate it and later review the decision.

5. Dynamic Routing and Decision Optimization

Routing in multimodal logistics involves more than selecting the shortest path. The decision may include departure time, mode sequence, transshipment terminal, service frequency, capacity reservation, inventory consequences, emissions and contractual commitments.

A conceptual objective is:

where C is total generalized cost, T is transport time, M is monetary operating cost, R is disruption or lateness risk, E is emissions, and c_T, c_M, c_R and c_E are decision-maker weights with compatible units after normalization or monetization. This is a managerial objective rather than a universal law. The weights must reflect the organization’s priorities and constraints.

Dynamic routing combines a changing network state with optimization. Inputs may include real-time congestion, road closures, port queues, vessel schedules, weather and available capacity. AI can forecast future travel times or disruption probabilities; an optimization method then selects a feasible route. This distinction matters: prediction estimates what may happen, while optimization chooses what to do.

Hard constraints can include vehicle capacity, legal driving hours, hazardous-material restrictions, service windows, train cut-offs and terminal operating hours. A route that minimizes predicted cost but violates a hard constraint is not an acceptable solution. Robust or stochastic optimization can be used when travel times and capacities are uncertain, but it may produce more conservative plans.

A practical architecture uses a rolling horizon. The system periodically updates observations, forecasts the near future, solves the routing problem and compares the new recommendation with the current plan. Replanning should be controlled: constant route changes can create operational instability, driver confusion and additional handling costs.

6. Asset Utilization and Predictive Maintenance

AI improves utilization by matching assets to demand and reducing unproductive time. Applications include container repositioning, vehicle dispatch, chassis allocation, empty-mile reduction, berth allocation and equipment pooling. Features may include location, asset type, availability, maintenance status, expected demand, service commitments and return probabilities.

Utilization should not be interpreted as “maximize activity at any cost.” Excessive utilization may reduce resilience, eliminate maintenance windows and increase failure risk. A balanced measure considers productive use, idle time, empty movement, service reliability and asset condition.

Predictive maintenance estimates the likelihood or timing of failure. A binary model may estimate whether failure will occur within a specified horizon. A survival model estimates the probability of surviving beyond time t, often written as:

T is the random failure time and S(t) is the survival function. The model is meaningful only when the failure definition, observation period and censoring treatment are clearly specified. A component that has not failed by the end of observation is not necessarily a non-failure; its record may be censored.

Maintenance decisions should compare the expected cost of intervention with the expected cost of failure. A simple decision rule is to intervene when:

where p_f is the estimated probability of failure during the decision horizon, C_f is the relevant cost of failure, and C_m is the cost of planned maintenance. The rule is simplified: it may need to include inspection costs, downtime, safety consequences, salvage value and the effect of maintenance on future reliability. For safety-critical assets, regulatory requirements may require intervention even when the economic calculation does not.

7. Autonomous Vehicles and Automated Transport

Autonomous transport exists on a spectrum rather than as a single category. Driver assistance, supervised automation, geofenced automated yard tractors, remotely supervised vehicles and fully autonomous operation have different technical and legal requirements.

Autonomous systems generally combine:

• perception through cameras, radar, lidar, GPS and other sensors;
• localization and mapping;
• prediction of other agents’ movements;
• motion planning;
• vehicle control; and
• monitoring, fallback and emergency procedures.

In logistics, controlled environments such as ports, mines, warehouses and dedicated yards are more suitable for early deployment than mixed urban traffic. Nevertheless, restricted environments still contain pedestrians, unexpected obstacles, communication failures, weather effects and ambiguous human signals.

Evaluation must address safety cases, operational design domain, fail-safe behavior, remote intervention, cybersecurity, maintenance and responsibility allocation. A vehicle that performs well in normal conditions may fail under sensor occlusion, unusual road layouts or degraded communications. Human supervision must therefore be designed as an active operational function, not merely a nominal presence.

8. Intelligent Terminals and Control Towers

An intelligent terminal integrates physical equipment, information systems and decision support. A terminal operating system can be connected to optical character recognition, positioning systems, appointment platforms, gate cameras, crane sensors, weather feeds and customs information.

Applications include:

• predicting gate and berth congestion;
• sequencing cranes and yard moves;
• allocating storage locations;
• detecting unsafe behavior or equipment anomalies;
• coordinating truck appointments;
• synchronizing rail and vessel cut-off times; and
• providing end-to-end visibility through a control tower.

The terminal is a socio-technical system. An algorithmic yard plan may be infeasible if it ignores driver availability, equipment changeover, union rules, inspection requirements or exceptional cargo. Successful deployment therefore requires interfaces that allow operators to inspect recommendations, override them and record reasons for deviation.

Figure: An AI-enabled transport system is a feedback loop linking data, models, decisions, operations and governance.

9. Explainability, Bias and Human Oversight

Explainability concerns whether relevant users can understand how a model produced a result. Different users require different explanations. A data scientist may need feature effects and calibration plots; a dispatcher may need the principal reasons for a late-arrival alert; an auditor may need a complete versioned record of data, model and decision.

Interpretability methods include inherently simple models, feature-importance analysis, local explanations and counterfactuals. Explanations should not be mistaken for proof of causality. A feature may be predictive because it acts as a proxy for another factor, not because changing it would cause the outcome.

Bias can enter through under-representative data, historical discrimination, inconsistent labels, missing data, proxy variables and unequal error costs. Consider a maintenance model trained mainly on one vehicle type: its apparently strong average accuracy may conceal poor performance on another type. Fairness analysis should compare relevant error rates and service outcomes across groups, lanes, regions, modes or asset classes. There is no single universal fairness metric; the appropriate measure depends on the decision and its consequences.

Human oversight should specify:

1. which decisions require approval;
2. when a system must defer to a human;
3. what evidence the human receives;
4. how overrides are recorded;
5. who is accountable for the outcome; and
6. how incidents trigger model review.

“Human in the loop” is inadequate if operators lack time, authority, training or meaningful information. Automation bias may cause staff to accept incorrect recommendations merely because they appear computationally authoritative. Conversely, excessive override rights without accountability can neutralize the intended benefits of the system.

10. Governance, Security and Regulation

AI governance in transport should cover the entire lifecycle: problem definition, data acquisition, development, testing, deployment, monitoring, retirement and incident response. Important controls include data lineage, access management, encryption, model versioning, reproducible testing, change approval and audit logs.

Privacy is relevant when systems process driver identities, location traces, biometric information or video. Data minimization and purpose limitation reduce unnecessary exposure. Cybersecurity is especially important because manipulated location, cargo or maintenance data can create physical as well as financial harm.

Regulatory obligations vary by jurisdiction and application. Safety-critical automated systems may require formal risk assessment, certification, operational-domain restrictions and documented fallback procedures. Contractual governance is also significant: transport partners should define data ownership, data quality responsibilities, liability for incorrect recommendations and rights to audit models.

Model monitoring should examine:

• predictive accuracy and calibration;
• data drift and concept drift;
• performance by relevant segment;
• alert volume and override rates;
• safety incidents and near misses;
• financial, service and emissions outcomes; and
• unexpected changes in human behavior.

A model can remain statistically accurate while becoming operationally harmful if users change their behavior in response to it. Monitoring must therefore include system-level outcomes, not only model-level metrics.

11. Illustrative Case: Multimodal Cold-Chain Control Tower

Background

This illustrative case concerns NorthLink Foods, a fictional distributor using refrigerated trucks, rail and short-sea shipping. Its customers require delivery within narrow time windows, and cargo quality depends on maintaining temperature limits.

Problem

NorthLink experienced missed rail connections, uncertain arrival times and temperature excursions during transshipment. Planners received GPS, reefer sensor and carrier event data, but the information was fragmented. Emergency truck bookings were expensive and often made too late.

Analysis

The company first defined three decisions: whether to intervene on a shipment, which alternative mode to reserve, and whether a reefer required inspection. It created a common event model for departure, arrival, gate-in, loading, unloading and customs release. Historical data were separated into training and time-ordered testing periods. Performance was evaluated by lane and season rather than only as an overall average.

An ETA model used location, speed, historical segment time, terminal queue indicators, weather and planned connections. A disruption detector combined supervised labels for missed connections with anomaly rules for prolonged stops and abnormal temperature changes. The system displayed an ETA interval and a reason code rather than a single unexplained score.

Approach and Solution

The control tower used a rolling horizon. When the probability of missing a critical connection exceeded a defined threshold, the system generated alternatives ranked by cost, delivery risk and emissions. A planner approved any mode change affecting customer commitments. Temperature alerts were escalated immediately, while low-severity anomalies were grouped for review.

The maintenance component estimated failure risk for refrigeration units, but safety inspections and manufacturer requirements remained mandatory. All overrides were logged with the selected reason, and monthly reviews examined whether certain carriers, regions or equipment types received disproportionate alerts.

Outcome

In the first evaluation period, the organization reported fewer late interventions and improved visibility of connection risk. It did not treat the results as proof that AI alone caused the improvement: planners also changed escalation procedures and data quality. The company continued monitoring false alerts, interval calibration, privacy and the risk that emergency re-routing might increase empty vehicle movements.

Learning Points

The case demonstrates that value came from integrating prediction with decisions and authority. Data standardization preceded model deployment. Human approval was concentrated on consequential actions rather than every routine alert. Finally, performance was assessed using service, cost, quality, emissions and governance measures rather than predictive accuracy alone.

12. Implementation Roadmap and Best Practices

A robust implementation sequence is as follows:

1. Define the decision and success criteria. Specify the operational action, time horizon, constraints and acceptable risk.
2. Assess data readiness. Check completeness, timeliness, ownership, definitions, bias and cybersecurity.
3. Establish a baseline. Compare AI with current rules, expert judgment and simple statistical methods.
4. Design evaluation correctly. Use time-based splits, realistic scenarios and segment-level analysis.
5. Pilot in a bounded setting. Use a lane, terminal, asset group or operational domain where outcomes can be observed.
6. Integrate with workflows. Provide alerts, explanations, approval controls and feedback channels.
7. Monitor continuously. Track drift, errors, safety, fairness, overrides and business outcomes.
8. Scale with governance. Standardize documentation, training, incident response and model retirement.

Best practice is to treat AI as a capability portfolio. Some applications require high predictive accuracy; others benefit more from transparency, speed, robustness or ease of integration. The appropriate question is not whether an organization is “using AI,” but whether its decisions are becoming more reliable, resilient, safe and sustainable.

Contemporary Developments

Current developments include foundation models for multimodal event interpretation, digital twins for terminal and network simulation, edge AI for low-latency vehicle and equipment decisions, federated learning for cross-company collaboration, reinforcement learning for sequential control, and optimization that explicitly includes carbon and resilience objectives. These technologies remain subject to data, compute, safety and governance limitations.

Generative AI can assist planners by summarizing disruptions, querying operational data and drafting exception reports. It should not be allowed to invent shipment status or silently execute high-consequence changes. Retrieval from authoritative systems, role-based access, citations to source events and human approval are essential.

Chapter Summary

AI supports multimodal logistics through forecasting, ETA prediction, disruption detection, dynamic routing, asset allocation and predictive maintenance. Forecasting methods should be selected according to data, horizon, uncertainty and decision consequences; complex models are not automatically better than transparent baselines. ETA systems should define the milestone, prevent data leakage and report uncertainty. Disruption detection requires a deliberate balance between missed events and false alerts.

Dynamic routing combines predictive models with constrained optimization. Asset utilization must be balanced against resilience and maintenance needs, while predictive maintenance depends on clear failure definitions, censored observations and cost-aware decisions. Autonomous vehicles and intelligent terminals are socio-technical systems whose safety depends on operational domains, fallback procedures, cybersecurity and competent human supervision.

Explainability, bias assessment, privacy, auditability and lifecycle monitoring are central requirements rather than optional additions. The most effective deployments connect data and models to authorized decisions, measure system-level outcomes and retain meaningful human accountability.

Key Terms

Artificial intelligence: Computational capability associated with learning, perception, reasoning or decision-making.

Machine learning: Methods that estimate patterns or relationships from data.

ETA: Estimated time of arrival at a specified transport milestone.

Forecasting: Estimating future demand, workload or operating conditions.

Data leakage: Use of information in model development that would not be available at decision time.

Prediction interval: A range intended to contain a future observation with a stated level of coverage.

Disruption detection: Identification of events or conditions that threaten planned operations.

Anomaly detection: Identification of observations that differ from an established pattern.

Dynamic routing: Repeated adjustment of routes as network conditions change.

Rolling horizon: Decision method that repeatedly updates forecasts and plans over a moving future window.

Asset utilization: Productive use of vehicles, containers, equipment or infrastructure.

Predictive maintenance: Maintenance planning based on estimated future condition or failure risk.

Survival function: Probability that a failure time exceeds a specified time.

Autonomous vehicle: A vehicle capable of sensing and performing some or all driving tasks with limited human control.

Intelligent terminal: A terminal using integrated data, automation and decision support to coordinate operations.

Explainability: The extent to which model behavior and outputs can be understood by relevant users.

Model drift: Change in data patterns or relationships that reduces model validity.

Human oversight: Defined human authority to review, approve, override or stop automated decisions.

Control tower: An integrated capability for monitoring and coordinating supply-chain or transport events.

Review Questions

Short Answer Questions

1. Why should an AI project begin with an operational decision rather than an algorithm?
2. Distinguish between precision and recall in disruption detection.
3. What is the purpose of a prediction interval in ETA management?
4. Identify two hard constraints that may affect multimodal routing.
5. Why can an anomaly not automatically be treated as a disruption?

Long Answer Questions

1. Compare traditional forecasting methods with machine-learning methods for multimodal transport demand.
2. Explain how a rolling-horizon architecture can support dynamic routing while avoiding excessive plan instability.
3. Discuss the technical and organizational requirements for predictive maintenance of transport assets.
4. Evaluate the role of explainability and human oversight in intelligent terminals.
5. Analyze the principal governance risks associated with autonomous vehicles and AI control towers.

Analytical / Application Questions

1. A terminal’s disruption detector identifies nearly every actual disruption but generates many false alarms. Which metrics and operational changes should managers examine before deployment?
2. Design an ETA evaluation plan for a service using road, rail and maritime legs. Specify data splitting, performance measures, uncertainty reporting and segment-level checks.
3. A carrier wants to reroute shipments solely to minimize predicted transport time. Construct a more complete decision framework incorporating cost, reliability, emissions, capacity and human approval.

Multiple Choice Questions with Answers

Question 1

Which statement best distinguishes prediction from optimization in multimodal transport?

A. Prediction estimates future conditions, while optimization selects an action subject to objectives and constraints.

B. Prediction always uses rules, while optimization always uses neural networks.

C. Prediction is used only for demand, while optimization is used only for maintenance.

D. Prediction removes uncertainty, while optimization creates uncertainty.

Correct Answer: A. Prediction estimates future conditions, while optimization selects an action subject to objectives and constraints.

Question 2

Which metric gives the average absolute error in the same units as the forecast target?

A. F1 score

B. MAE

C. Precision

D. Recall

Correct Answer: B. MAE

Question 3

If a disruption detector has TP = 80 and FP = 20, its precision is:

A. 20%

B. 50%

C. 80%

D. 100%

Correct Answer: C. 80%

Question 4

Why is a random train-test split often inappropriate for transport time-series forecasting?

A. It prevents the model from using historical observations.

B. It may allow information from the future to influence training evaluation.

C. It always produces too few observations.

D. It can only be used with classification models.

Correct Answer: B. It may allow information from the future to influence training evaluation.

Question 5

In predictive maintenance, what does S(t) = P(T > t) represent?

A. The probability that failure occurs exactly at time t

B. The probability that the asset survives beyond time t

C. The average maintenance cost at time t

D. The proportion of defective assets repaired at time t

Correct Answer: B. The probability that the asset survives beyond time t

Question 6

Which is a hard routing constraint?

A. A planner’s preferred color for route maps

B. A vehicle’s legal driving-hour limit

C. A model’s feature-importance ranking

D. A carrier’s average historical fuel price

Correct Answer: B. A vehicle’s legal driving-hour limit

Question 7

Which practice most directly reduces alert fatigue?

A. Sending every anomaly to every employee

B. Removing all thresholds from the alert system

C. Using severity levels, persistence rules and escalation procedures

D. Hiding the evidence supporting each alert

Correct Answer: C. Using severity levels, persistence rules and escalation procedures

Question 8

Why should autonomous vehicles be evaluated within an operational design domain?

A. Performance and safety depend on the environmental and operational conditions in which the vehicle is used.

B. The domain determines the vehicle’s paint specification.

C. It eliminates the need for fallback procedures.

D. It guarantees that sensors cannot fail.

Correct Answer: A. Performance and safety depend on the environmental and operational conditions in which the vehicle is used.

Question 9

Which example best illustrates data leakage in ETA prediction?

A. Using historical travel time known before departure

B. Using weather forecasts available at prediction time

C. Using the actual terminal release time when predicting release before it occurred

D. Using the vehicle’s current GPS position

Correct Answer: C. Using the actual terminal release time when predicting release before it occurred

Question 10

What is the most appropriate view of human oversight in high-consequence AI decisions?

A. A nominal approval button with no information or authority

B. A defined process specifying review, evidence, override authority and accountability

C. A requirement that humans manually perform every calculation

D. A method for transferring all responsibility to the algorithm

Correct Answer: B. A defined process specifying review, evidence, override authority and accountability