To analyze the changing characteristics of ship traffic flow in wind farms water area, and to improve the accuracy of ship traffic flow prediction, a Gated Recurrent Unit (GRU) of a Recurrent Neural Network (RNN) was established to analyze multiple traffic flow sections in complex waters based on their traffic flow structure. Herein, we construct a spatiotemporal dependence feature matrix to predict ship traffic flow instead of the traditional ship traffic flow time series as the input of the neural network. The model was used to predict the ship traffic flow in the water area of wind farms in Yancheng city, Jiangsu Province. Autoregressive Integrated Moving Average (ARIMA), Support-Vector Machine (SVM) and Long Short-Term Memory (LSTM) were chosen as the control tests. The GRU method based on the spatiotemporal dependence is more accurate than the current mainstream ship traffic flow prediction methods. The results verify the reliability and validity of the GRU method.

Accurate vessel trajectory prediction from Automatic Identification System (AIS) records is important for maritime traffic monitoring, route planning, and intelligent vessel traffic management. However, reliable prediction remains challenging for long forecasting horizons and turning maneuvers. To address this problem, this study proposes the Geometry-Consistent Trajectory Transformer (GCT-Former), a progressive and refinement-based framework for AIS-based vessel trajectory prediction. The proposed model integrates multi-scale historical trajectory encoding, progressive residual future-position generation, and global–local trajectory refinement to improve the stability and continuity of long-horizon trajectory prediction. The predicted trajectories are evaluated as geometric future-position estimates and can provide trajectory-level information for downstream maritime traffic monitoring and decision-support applications. Experiments are conducted on three real-world Danish maritime regions: Aarhus Bay, Great Belt, and Skagen. Compared with representative conventional and deep-learning trajectory prediction models, the proposed model shows its most consistent advantage in long-horizon prediction, particularly in terms of ADE and FDE. In the long-term setting, it achieves average displacement errors of 0.344, 0.546, and 0.218 km and final displacement errors of 0.774, 1.368, and 0.525 km on Aarhus Bay, Great Belt, and Skagen, respectively. The ablation analysis further shows that removing the multi-scale encoding module increases the long-term average displacement error by 7%, 4%, and 3%, while removing the progressive residual decoder leads to larger increases of 15%, 9%, and 8% on the three datasets. The turning-maneuver analysis also shows lower geometric prediction errors under mild-turning and sharp-turning scenarios. These results indicate that GCT-Former improves AIS-based vessel trajectory prediction, especially for long-horizon and maneuvering cases.

Accurate short-horizon forecasting of ship CO₂ emissions is important for maritime carbon monitoring, operational assessment, and the transition toward sustainable maritime transportation. However, reliable forecasting remains challenging because ship emissions are affected by temporally varying operating states, vessel-specific characteristics, and nonlinear emission relationships. This study proposes a Carbon-oriented Spatio-Temporal Graph Transformer (C-STGT) for forecasting STEAM-estimated ship CO₂ emissions from Automatic Identification System (AIS) trajectories. The proposed model integrates temporal representation learning, single-vessel graph modeling, vessel-specific attributes, carbon-related attention bias, and emission-oriented decoding to capture vessel operating and emission dynamics. C-STGT is evaluated on five containerships operating along the Shanghai–Los Angeles corridor and compared with ten baseline models. Experimental results show that C-STGT achieves a root mean squared error (RMSE) of $0.89$, a mean absolute error (MAE) of $0.51$, a weighted absolute percentage error (WAPE) of $3.11\%$, and an $R^2$ of $0.96$, outperforming the compared models. Ablation, emission-formulation sensitivity, and statistical analyses further support the effectiveness and robustness of the proposed framework. By improving short-horizon vessel-level carbon-emission forecasting, the proposed approach can support more timely emission monitoring and operational assessment, providing quantitative information for maritime operations and maritime decarbonization. In summary, C-STGT provides an effective forecasting framework that contributes to the development of more sustainable maritime transportation.

As vessel traffic becomes increasingly dense in inland waterways, accurate prediction and interpretation of multi-ship collision risk are important for navigational safety and intelligent maritime supervision. However, existing methods mainly focus on pairwise risk assessment or scenario-level prediction, while providing limited insight into which vessels and interactions contribute to predicted future risk. To address these limitations, this study proposes a Multi-Ship Risk Attribution Network (MSRA-Net) for interpretable short-term collision-risk prediction from historical Automatic Identification System (AIS) data. The proposed framework integrates three key components: a Dynamic Multi-Ship Interaction Graph (DMIG) to represent time-varying vessel states and encounter relationships, a Spatio-Temporal Risk Prediction Network (STRP-Net) to capture temporal vessel dynamics and inter-vessel interactions, and a Multi-Ship Risk Attribution (MSRA) module to quantify vessel-level and interaction-level contributions to the predicted risk. The performance of MSRA-Net is evaluated using 6,433,925 AIS observations collected from a representative section of the Yangtze River in Wuhan, China. Experimental results show that MSRA-Net achieves the lowest average MAE of 0.0493 across 1, 3, and 5 min prediction horizons and obtains an MAE of 0.0632, RMSE of 0.0740, and $R^2$ of 0.862 at the 5 min horizon. Progressive vessel-deletion experiments further demonstrate attribution fidelity, with high-to-low deletion achieving a DelAUC of 0.1730, compared with 0.3016 for random deletion and 0.4384 for low-to-high deletion. Controlled AIS perturbation experiments also show that the learned attribution patterns remain relatively stable under small variations in vessel states. Overall, MSRA-Net provides a unified framework for dynamic multi-ship interaction modeling, future collision-risk prediction, and model-level risk attribution, offering an interpretable approach for maritime traffic safety analysis.