Advanced prediction of maritime vessel traffic flow using spatio-temporal multi-graph convolutional networks within inland waterways

Abstract Accurate vessel traffic prediction is significant for efficient waterway management and lock scheduling. This paper presents a deep-learning framework that integrates a multi-graph convolutional network with a gated recurrent unit network, considering spatio-temporal patterns appropriately, for vessel traffic flow prediction. Three unstructured graphs are constructed to represent spatio-temporal relationships among traffic flows at different locations. Subsequently, multi-graph convolution is employed to quantitatively extract such patterns among adjacent nodes in the graphs. Those extracted patterns are then passed to a gated recurrent unit layer for further temporal features extraction in sequential data. The model is believed to improve prediction accuracy and reliability. To prove this, extensive experiments on regional and station-based predictions are conducted using two real-world datasets to evaluate the model’s capability. The jointly trained model demonstrates superior performance and outperforms conventional methods. The strong forecasting ability enables managers to adjust schedules promptly, enhancing efficiency and intelligence of waterway operations.