Integrating graph neural networks and LSTM for path optimization in smart port multi-modal systems.

Area of Science:

• Maritime Logistics
• Artificial Intelligence
• Operations Research

Background:

• Smart ports face challenges in optimizing multimodal transport due to dynamic environments and complex data fusion.
• Existing methods struggle with integrating diverse data sources and adapting to real-time network changes.

Purpose of the Study:

• To propose an advanced model for multimodal transport path optimization in smart ports.
• To enhance path planning efficiency by effectively fusing multimodal data and leveraging network topology.

Main Methods:

• Developed the GL-SSL Model, integrating Graph Neural Networks (GCN), Long Short-Term Memory (LSTM), and Self-Supervised Learning (SSL).
• Utilized AIS data, global shipping data, and port revenue data for model training and validation.
• Exploited graph-structured port networks and temporal variations for optimized path planning.

Main Results:

• Achieved an optimized path length of 80 km and reduced transport costs to 200 cost-units.
• Maintained a low delay rate of 0.05 (5%), outperforming traditional and other deep learning models.
• Demonstrated stable performance and robustness in complex scenarios like peak traffic and adverse weather.

Conclusions:

• The GL-SSL Model offers significant improvements in multimodal transport path optimization for smart ports.
• The model shows strong adaptability, robustness, and practical application potential in dynamic port environments.
• This research provides effective technical support and holds theoretical significance for smart port logistics.