Predicting ship main engine fuel oil consumption is an important task for improving fuel efficiency, reducing operational cost, and minimizing harmful emissions in maritime transportation. Conventional shallow machine learning models often fail to capture complex operational patterns, while deep neural networks with fully connected layers may increase computational complexity and reduce practical deployment efficiency. To address these limitations, this project proposes a Hybrid Temporal-Spatial Attention Network (HTSAN) for accurate prediction of ship main engine fuel oil consumption. The proposed system uses historical ship operational and emission-related data from the Kaggle ship fuel consumption and CO₂ emissions dataset. HTSAN captures temporal dependencies from sequential operating conditions and spatial relationships among engine, fuel, and environmental parameters using attention mechanisms. In addition, a Graph Neural Network (GNN) with Engine-Environment Graph is used to represent the relationship between ship engine behaviour and external environmental factors, while a Transformer-based Regression Model is applied to learn long-range dependencies in the data. The experimental results show that the proposed HTSAN model achieves a high R² score of 0.99, outperforming the Transformer-based Regression Model, which achieves an R² score of 0.93. The results indicate that the hybrid attention-based approach provides better prediction accuracy, model robustness, and decision support for energy-efficient ship operation.

Keywords: Ship fuel consumption, Main engine, HTSAN, Graph Neural Network, Transformer regression, Fuel efficiency, CO₂ emissions, Maritime transportation.