A Bayesian-network-guided spatiotemporal model for ship navigation-status prediction in the Three Gorges Reservoir Area

Abstract

The Three Gorges Reservoir Area is a critical bottleneck in China's inland waterway system. Rapid hydrological variation and dense vessel traffic can quickly change whether cargo ships are permitted to navigate in the regulated reach between the Three Gorges Dam and Gezhouba Dam. Reliable prediction of future navigation status is therefore important for vessel admission, traffic organization, and dispatch preparation. This study develops a Bayesian-network-guided spatiotemporal framework for predicting ship navigation status 2∼8 h ahead. The model uses a fixed 18-node Bayesian network to complete historical intermediate states, provide the topology for graph-based dependency propagation, and supply auxiliary posterior guidance during training. A graph neural network and Transformer learn variable dependencies and temporal evolution, while final predictions remain directly supervised by recorded future labels and inference uses historical observations only. Experiments use 25216 event-level samples from 200 cargo vessels with relatively complete records for 2024, combining automatic identification system records with hydrological, traffic-management, and meteorological data. At 2 h, the harmonic mean of precision and recall averaged across the three hydrological conditions was 0.948, compared with 0.814 for direct Bayesian-network prediction and 0.826 after temporal smoothing. The corresponding mean area under the receiver operating characteristic curve was 0.984. Its advantage over the strongest competing model widened at longer horizons. Ablation and perturbation analyses showed that label supervision was essential, Bayesian structural and posterior guidance were complementary, and missing core dependencies reduced performance. Within the study area, the model can support vessel screening, traffic organization, and rolling coordination of waterway and ship-lock operations.