Semantic-driven multi-view periodic-temporal network for predicting vessel traffic flow

Predicting Vessel Traffic Flow (VTF) is a critical component of intelligent maritime transportation management, using Automatic Identification System (AIS) data analysis. Traditional deep learning methods often fail to fully capture the complex and dynamic nature of VTF data. To overcome these limitations, this study introduces a novel approach: the Multi-view PeriodicTemporal Network with Semantic Representation (MPTNSR). This model integrates three perspectivesperiodic, temporal, and semantic-to provide a comprehensive solution for VTF prediction. The periodic and temporal features of VTF, often hidden in its evolution, are effectively extracted using a hybrid Convolutional Neural Network and Bidirectional Long Short-Term Memory framework. Meanwhile, in real-world applications where multiple target regions are involved, the semantic view leverages a Graph Convolutional Network (GCN) to capture correlations in VTF f luctuations across regions based on geographical and statistical similarities. By fusing insights from all three views, the model achieves accurate prediction. Additionally, an optimised loss function incorporating both local and global metrics ensures robust performance. Extensive comparative experiment results demonstrate that the MPTNSR model significantly outperforms fourteen methods in VTF prediction tasks. Its ability to deliver precise prediction makes it a valuable tool for port planning, enhancing safety and sustainability in maritime