Spatio-temporal graph neural network fused with maritime knowledge for predicting traffic flows in ports

Abstract

The COVID-19 pandemic highlighted significant challenges in maritime industries, such as port congestion and supply chain disruptions, necessitating accurate port traffic prediction. Traditional models often fail to capture the complex and dynamic interactions within maritime networks, particularly the spatial correlations between ports. To address these limitations, this study proposes a spatio-temporal graph neural network (MSTGNN) tailored for maritime traffic prediction. MSTGNN integrates temporal and spatial features using a sample convolution and interaction network (SCINet) and a graph attention network (GAT), enhanced by a k-core mechanism for prioritizing key ports. Extensive experiments with global and regional datasets demonstrate MSTGNN's superior performance in accuracy, stability, and adaptability compared to existing models. The model excels in predicting traffic flows across different ship types, making it highly applicable for optimizing port operations, route planning, and mitigating supply chain risks. The findings imply that MSTGNN can significantly enhance decision-making in maritime logistics by offering more precise traffic predictions, thereby supporting more efficient, resilient, and sustainable global supply chains.