Semi-dynamic spatial–temporal graph neural network for traffic state prediction in waterways

Abstract

In the field of maritime traffic, the assessment of waterway passage conditions constitutes a significant research topic, particularly in the context of traffic flow analysis and prediction for busy waterway segments. This is crucial for the regulation of waterways and the enhancement of navigational potential. This paper presents a new semi-dynamic spatio-temporal graph neural network (SDSTGNN) for processing traffic state prediction of waterways. A semi-dynamic spatial graph network (SDSG) module is firstly provided for capture the spatial features, in which a semi-dynamic adjacency matrix based on a predefined adjacency matrix and adaptive matrix is designed for constructing the graph network. Furthermore, to enhances the ability to capture temporal features from input features, a new structure of temporal feature extraction module combining LSTM and Transformer is designed. The comparison of experiments on three datasets illustrate the proposed model SDSTGNN has superior forecasting ability to the compared models. In a comprehensive analysis, MAE, RMSE, MAPE, and WAPE exhibit remarkable reduction averagely on 1–6 period with 6.9%, 5.5%, 18.1%, 9.5%, respectively. By implementation of proposed model, it could optimize ship operations and regulate ship traffic, which benefit in the stakeholders, such as maritime regulatory authorities, ship operators, freight agents, and port management departments.