AI-POWERED PORT CONGESTION INSIGHTS
LLM-Grounded Explainability for Port Congestion Prediction via Temporal Graph Attention Networks
This paper introduces AIS-TGNN, a novel evidence-grounded framework designed for port congestion prediction and natural-language explanation. By combining a Temporal Graph Attention Network (TGAT) with a Large Language Model (LLM) reasoning module, AIS-TGNN provides interpretable and auditable risk reports without sacrificing predictive performance. Using daily spatial graphs from AIS data, the TGAT predicts congestion escalation, while the LLM transforms model-internal evidence (feature z-scores, attention weights) into coherent, domain-specific explanations. Experiments on Port of Los Angeles/Long Beach data demonstrate superior performance (AUC 0.761, AP 0.344, Recall 0.504) and high explanation reliability (99.6% directional consistency). This framework offers a practical pathway for AI-assisted maritime decision support and supply-chain risk management.
Key Performance Indicators
Our AIS-TGNN framework sets new benchmarks for predictive accuracy and explanatory reliability in maritime logistics.
0.761
Test AUC
0.344
Average Precision (AP)
0.504
Recall
99.6%
Explanation Consistency
Deep Analysis & Enterprise Applications
Select a topic to dive deeper, then explore the specific findings from the research, rebuilt as interactive, enterprise-focused modules.
The Temporal Graph Attention Network (TGAT) is the core predictive engine, building daily spatial graphs from Automatic Identification System (AIS) data. Each grid cell represents vessel activity, and inter-cell interactions are modeled via attention-based message passing. TGAT effectively captures complex spatiotemporal congestion dynamics, adapting neighbor influence based on learnable attention coefficients.
This allows for more nuanced understanding of localized conditions and 'flow-back' effects, leading to robust congestion escalation predictions. The model leverages current and prior-state features from sequential snapshots to make informed predictions.
Our framework bridges the explainability gap by integrating a structured Large Language Model (LLM) reasoning module. The LLM synthesizes model-internal evidence—such as feature z-scores and attention-derived neighbor influence—into coherent, domain-specific natural-language risk reports.
This grounding ensures that explanations are verifiable and faithful to the model's outputs, preventing the generation of plausible-sounding but unsupported narratives. The LLM generates multi-section reports, including feature drivers, neighbor influence, and counterfactual suggestions, making AI-assisted decision support transparent and auditable.
The AIS-TGNN pipeline processes raw NOAA AIS broadcasts, transforming them into a chronological sequence of 89 daily spatiotemporal graph snapshots. Each node in the graph represents a 0.1° x 0.1° grid cell with at least one AIS observation, and connections are formed using a k-nearest-neighbor (k=8) topology based on Euclidean distance.
Nodes are enriched with ten kinematic and traffic-density features (e.g., mean SOG, slow-vessel ratio, vessel count), which are z-score normalized. The binary congestion-escalation label identifies if a cell's slow ratio increases day-over-day, signaling a transition to a more congested state.
| Model | AUC | AP | F1 | Recall |
|---|---|---|---|---|
| LR (no graph) | 0.713 | 0.300 | 0.375 | 0.480 |
| GCN (static graph) | 0.759 | 0.326 | 0.383 | 0.445 |
| TGAT (graph + attention) | 0.761 | 0.344 | 0.398 | 0.504 |
The TGAT model significantly outperforms LR and GCN baselines, especially in terms of Average Precision and Recall, demonstrating its superior ability to identify true congestion escalation events in an imbalanced dataset.
Explanation Reliability
99.6% Directional Consistency RateThe generated natural-language explanations exhibit a 99.6% directional consistency, ensuring that the stated risk directions (increase/decrease risk) align faithfully with the model's underlying statistical evidence (feature z-scores and point-biserial correlations). This validates the factual reliability of the LLM-generated reports.
Enterprise Process Flow
AIS Data Acquisition
→
Graph Construction
→
TGAT Model Training
→
Evidence Extraction
→
LLM Risk Explanation
Case Study: LLM-Generated Risk Report
Node 321_-1195, 2023-06-02
A representative case illustrates how the LLM generates a detailed risk report for a specific grid cell with a predicted congestion-escalation probability of 0.659.
- Primary Drivers: Highlights features like Mean Draft Depth (z=+1.001, r=-0.130), showing a negative association with escalation, and COG Variance (z=+0.441, r=+0.175), showing a positive association.
- Neighbor Influence: Identifies strong spatial neighbors (e.g., 320_-1197) and their key features that contribute to the prediction via attention weights.
- Counterfactual Suggestions: Provides actionable insights, such as decreasing COG variance or speed standard deviation to potentially reduce escalation risk.
- Directional Consistency: This report, like others, adhered to the 99.6% directional consistency, ensuring fidelity to model evidence.
This detailed, model-grounded report allows port planners to understand not just 'what' is predicted, but 'why', facilitating proactive decision-making.
Estimate Your Potential AI Impact
Input your operational metrics to see the potential annual hours reclaimed and cost savings with an explainable AI solution like AIS-TGNN.
Annual Savings
Hours Reclaimed Annually
Your Explainable AI Implementation Roadmap
A clear path to integrating AIS-TGNN into your operations.
Phase 1: Data Integration & Baseline Model Development
Establish data pipelines for AIS and relevant covariates. Develop initial baseline models and define performance metrics, focusing on data quality and feature engineering.
Phase 2: TGAT Model Training & Optimization
Implement and train the Temporal Graph Attention Network (TGAT) using historical data. Optimize hyperparameters and validate predictive performance on unseen data, ensuring robustness.
Phase 3: LLM Integration & Explanation Generation
Integrate the Large Language Model (LLM) module. Develop structured prompts to extract and synthesize model-internal evidence into natural-language explanations. Establish directional consistency validation.
Phase 4: Pilot Deployment & Validation
Conduct a pilot deployment within a controlled operational environment. Gather feedback from port planners and logistics operators. Iterate on the explanation quality and model accuracy based on real-world usage.
Phase 5: Continuous Monitoring & Refinement
Deploy the AIS-TGNN framework for continuous monitoring. Implement feedback loops for ongoing model retraining and explanation refinement, adapting to evolving maritime conditions and user needs.
Ready to Transform Your Maritime Operations?
Connect with our AI experts to discuss how AIS-TGNN can enhance your port congestion prediction and supply chain risk management.
Ready to Get Started?
Book Your Free Consultation.
Let's Discuss Your AI Strategy!
Lets Discuss Your Needs
Select a Date
Sun
Mon
Tue
Wed
Thu
Fri
Sat