Predicting Port Bottlenecks with AI and Natural Language

Author: Denis Avetisyan

A new approach combines graph neural networks with large language models to forecast congestion in major ports and provide clear, understandable explanations for potential disruptions.

The study area encompasses the Los Angeles/Long Beach port complex, modeled as a spatial graph to facilitate analysis of logistical networks and interconnections within the region.

This work introduces AIS-TGNN, a framework leveraging spatiotemporal analysis of AIS data with LLM-grounded explainability for improved maritime congestion prediction.

Predicting port congestion is critical for global supply chain resilience, yet existing systems often lack the transparency needed for effective operational response. This is addressed in ‘LLM-Grounded Explainability for Port Congestion Prediction via Temporal Graph Attention Networks’, which introduces a novel framework-AIS-TGNN-that combines spatiotemporal graph analysis with large language models to not only forecast congestion but also generate interpretable explanations grounded in model evidence. Experiments utilizing Automatic Identification System (AIS) data from the Ports of Los Angeles and Long Beach demonstrate improved predictive performance-achieving an AUC of 0.761-alongside explanations with 99.6% directional consistency. Could this approach pave the way for more auditable and actionable AI systems in maritime logistics and beyond?

Decoding the Chaos: Predicting Port Congestion

The intricate web of global supply chains faces escalating vulnerability due to port congestion, a phenomenon with far-reaching economic consequences. Modern just-in-time inventory management, while efficient, leaves little buffer for delays; consequently, even localized congestion can trigger ripple effects across continents. Recent years have demonstrated this acutely, with major ports experiencing backlogs that disrupted manufacturing, increased consumer prices, and contributed to inflationary pressures. These disruptions aren’t limited to specific goods; everything from raw materials and agricultural products to finished consumer items are susceptible. The increasing scale of container ships, coupled with landside infrastructure limitations and labor shortages, further exacerbates the issue, turning isolated incidents into systemic risks for the global economy. Ultimately, the efficiency of international trade hinges on the ability to move goods seamlessly through these critical maritime hubs.

Current approaches to forecasting port congestion frequently rely on lagging indicators and historical averages, proving inadequate in the face of increasingly dynamic and unpredictable global events. These methods often fail to account for the complex interplay of factors – from geopolitical instability and unexpected surges in demand to logistical bottlenecks and labor disputes – which can rapidly transform normal operations into crippling delays. Consequently, a shift towards proactive, data-driven solutions is essential. Such solutions must not only predict the likelihood of congestion but also offer clear, interpretable explanations of the contributing factors, enabling stakeholders to anticipate disruptions, optimize resource allocation, and implement effective mitigation strategies within the intricate maritime logistics network.

The ability to accurately forecast port congestion and, crucially, explain those predictions is becoming paramount for resilient maritime logistics. Beyond simply knowing a disruption is likely, stakeholders – from shipping companies to inland transport providers and warehouse operators – require insight into why a bottleneck is forming. This understanding allows for proactive mitigation strategies, such as rerouting vessels, pre-positioning resources, and adjusting schedules to minimize delays and associated costs. Without interpretable forecasts, the potential for cascading failures throughout the supply chain increases significantly, hindering efficient resource allocation and ultimately impacting global trade. Effective prediction, coupled with transparent explanations of contributing factors – like vessel bunching, equipment imbalances, or labor shortages – transforms reactive crisis management into a proactive, optimized network.

AIS-TGNN: A Framework for Seeing Through the Fog

The AIS-TGNN framework integrates a Temporal Graph Attention Network (TGAT) with a Large Language Model (LLM) Explanation Module to address the challenges of both predicting and interpreting port congestion. The TGAT component processes Automatic Identification System (AIS) data, constructing a dynamic graph representing vessel interactions and locations over time. This allows the model to capture complex dependencies crucial for understanding congestion patterns. Following the TGAT’s predictive output, the LLM Explanation Module analyzes the graph’s learned representations to generate human-readable explanations regarding the factors contributing to predicted congestion levels, thereby moving beyond simple forecasting to provide actionable insights.

The Temporal Graph Attention Network (TGAT) within the AIS-TGNN framework utilizes Automatic Identification System (AIS) data, which broadcasts vessel position, speed, heading, and identification details, to construct a dynamic graph representing maritime traffic. This graph’s nodes are individual vessels and its edges represent their proximity and interaction. The TGAT then applies attention mechanisms to weigh the influence of neighboring vessels over time, effectively capturing both spatial relationships – how close vessels are to each other – and temporal dependencies – how vessel movements influence each other’s future positions. By analyzing these relationships, the TGAT can model the complex interplay of vessel traffic and identify patterns indicative of potential congestion events, going beyond static spatial analysis to account for the evolving dynamics of port activity.

The AIS-TGNN framework demonstrated a test Area Under the Curve (AUC) of 0.761 when predicting daily congestion-escalation events at the Port of Los Angeles and Long Beach. This performance metric indicates the model’s ability to distinguish between days with escalating congestion and those without. Crucially, the framework extends beyond simple predictive capability by incorporating an LLM Explanation Module, enabling the identification of contributing factors to predicted congestion. This allows for analysis of why a particular congestion event is likely, providing actionable insights for port authorities and stakeholders beyond a basic forecast.

The area under the receiver operating characteristic curve (AUC-ROC) demonstrates validation performance across training epochs for LR, GCN, and TGAT models.

Decoding the Signals: Natural Language Explanations

The LLM Explanation Module functions by translating quantitative data from the predictive model – specifically attention weights and standardized feature values, known as Feature Z-Scores – into human-readable natural language. Attention weights indicate the relative importance of each input feature in the model’s decision-making process, while Feature Z-Scores represent the degree to which each feature deviates from its historical average. This conversion allows users to understand why a particular prediction was made, identifying the key factors influencing the outcome rather than simply receiving a prediction score. The module’s output is a textual report detailing these influential features and their corresponding impact on the prediction.

The LLM Explanation Module utilizes GPT-4o-mini to translate model outputs – specifically attention weights and standardized feature values (Feature Z-Scores) – into human-readable reports. These reports detail the key factors influencing predicted congestion levels, providing insight into the model’s reasoning. GPT-4o-mini’s role is to synthesize these numerical data points into concise, informative narratives, enabling users to understand why a particular congestion prediction was made, rather than simply receiving a prediction itself.

Rigorous evaluation of the congestion-prediction framework at the Port of Los Angeles and Long Beach yielded several key performance metrics. The framework achieved an Area Under the Curve (AUC) of 0.761, indicating its ability to discriminate between congestion-escalation and non-escalation events. Average Precision (AP) was measured at 0.344, representing the precision of positive predictions across different recall levels. Recall, or the ability to correctly identify all actual congestion-escalation events, was determined to be 0.504. These metrics collectively assess the framework’s performance in accurately and reliably predicting daily congestion events.

Beyond Prediction: Trustworthy and Transparent AI

To rigorously evaluate the trustworthiness of explanations produced by the framework, researchers employed Directional Consistency as a key metric. This measure goes beyond simply checking if an explanation mentions relevant factors; it assesses whether the direction of the explanation aligns with the direction of the evidence within the underlying model. Essentially, it verifies that if a feature is identified as positively influencing a prediction, the model’s internal data also supports that positive correlation, and vice versa. This detailed alignment process ensures explanations aren’t merely plausible-sounding statements, but genuinely reflect the reasoning process of the artificial intelligence system, fostering confidence in the insights generated and facilitating responsible application of the technology.

A rigorous evaluation of the AIS-TGNN Framework revealed an exceptionally high Directional Consistency of 99.6%. This metric demonstrates a strong correlation between the explanations generated by the model and the actual factors influencing its predictions. Essentially, the system doesn’t just state what drove a decision; it accurately reflects the underlying reasoning process. This level of fidelity is crucial for building trust in artificial intelligence systems, allowing stakeholders to confidently interpret insights and utilize them for informed decision-making in critical applications. The result suggests a robust and transparent framework where explanations are not merely post-hoc justifications, but faithful representations of the model’s internal logic.

The AIS-TGNN Framework’s reliability fosters confidence among those utilizing its insights for critical decision-making. A system capable of consistently aligning its explanations with the data driving its predictions minimizes the risk of acting on misleading or fabricated justifications. This transparency is paramount in fields where accountability and understanding are crucial; stakeholders can not only observe what a prediction is, but also confidently grasp why it was made. Ultimately, the framework’s faithfulness empowers users to integrate its outputs into complex workflows, enabling data-driven strategies with assurance and informed judgment, rather than relying on opaque algorithmic outputs.

A Vision for Resilient and Intelligent Logistics

The escalating pressures on global supply chains demand proactive strategies for managing port congestion, and the AIS-TGNN framework offers a significant advancement in this area. By integrating Automatic Identification System (AIS) data – which tracks vessel movements – with a Temporal Graph Neural Network, the system can forecast potential bottlenecks with unprecedented accuracy. Crucially, this predictive capability is amplified by leveraging robust geospatial data from sources like NOAA’s Marine Cadastre, providing detailed information on channel depths, mooring locations, and critical infrastructure. This combination allows for the identification of congestion risks before they materialize, enabling port authorities and logistics providers to optimize vessel scheduling, reroute traffic, and proactively allocate resources – ultimately fostering smoother, more reliable maritime operations and reducing costly delays.

The developed AIS-TGNN framework is not limited to maritime logistics; researchers envision its adaptation to bolster resilience across a spectrum of critical infrastructure systems, including energy grids, transportation networks, and supply chains. This expansion necessitates integrating diverse, real-time data streams – encompassing weather patterns, traffic flow, and even social media feeds – to provide a holistic and dynamic understanding of potential disruptions. By continuously updating the framework with live information, predictive capabilities can be significantly enhanced, allowing for proactive interventions and optimized resource allocation during crises. This move towards a unified, data-driven approach promises to create interconnected and adaptable infrastructure networks capable of withstanding increasingly complex challenges and ensuring operational continuity.

The development of predictive frameworks for maritime logistics represents a significant step towards establishing networks capable of withstanding disruption and operating with enhanced efficiency. By integrating advanced data analytics with comprehensive datasets, this research facilitates a proactive approach to identifying and mitigating potential bottlenecks, thereby bolstering the resilience of global supply chains. Furthermore, increased transparency – enabled by real-time monitoring and predictive modeling – allows for more informed decision-making across all stakeholders, reducing delays and optimizing resource allocation. This ultimately contributes to a more sustainable system, minimizing environmental impact through reduced fuel consumption and optimized vessel routing, and fostering long-term economic viability within the maritime sector.

The pursuit of predictive accuracy, as demonstrated by AIS-TGNN, isn’t merely about forecasting port congestion; it’s about deconstructing the system governing maritime traffic. This framework actively challenges the limitations of traditional methods by integrating large language models – a calculated risk in the name of improved spatiotemporal analysis. As Bertrand Russell observed, “The difficulty lies not so much in developing new ideas as in escaping from old ones.” The researchers didn’t simply accept existing predictive models; they interrogated their foundations, seeking to break free from conventional approaches and forge a more nuanced, interpretable understanding of port dynamics. This echoes a fundamental principle: true understanding comes from dismantling assumptions and rebuilding them with demonstrable evidence.

What Lies Ahead?

The integration of large language models with graph neural networks, as demonstrated by AIS-TGNN, offers a superficially elegant solution to the problem of port congestion prediction. However, true progress demands a rigorous examination of the assumptions baked into these models. The current architecture treats language as a descriptive addendum to quantifiable spatiotemporal data. A more disruptive approach would be to explore whether the structure of language itself encodes predictive information about logistical bottlenecks – to reverse-engineer the inherent constraints within global trade networks, rather than simply mapping their effects.

Furthermore, the pursuit of ‘explainability’ often devolves into post-hoc rationalization. Providing interpretable risk reports is useful, certainly, but it sidesteps the fundamental question of whether these models genuinely understand the causal mechanisms at play. A more honest assessment would acknowledge the inherent limitations of pattern recognition – that correlation does not equal causation, and that even the most sophisticated algorithm remains a black box, albeit one with a nicely formatted output.

Ultimately, the field must move beyond incremental improvements in predictive accuracy. The true test lies in building models that can anticipate unforeseen disruptions – the ‘black swan’ events that expose the fragility of complex systems. This requires not just better data and algorithms, but a willingness to challenge the underlying assumptions about predictability itself. The goal is not to predict the future, but to understand the limits of prediction.

Original article: https://arxiv.org/pdf/2603.04818.pdf

Contact the author: https://www.linkedin.com/in/avetisyan/