Parallel TCN-BiGRU architecture with dynamic attention for ship energy consumption prediction under variable navigation conditions

Abstract

Accurate prediction of ship energy consumption is essential for improving operational efficiency and reducing emissions. However, existing models often fail to capture complex spatiotemporal dependencies inherent in dynamic maritime environments. This study presents a parallel hybrid deep learning framework (TSBG-Para), which integrates Temporal Convolutional Networks (TCN), Bidirectional Gated Recurrent Units (BiGRU), and two attention mechanisms: Squeeze-and-Excitation (SE) and Global Attention (GA). Unlike conventional serial models, TSBG-Para adopts dual parallel branches for spatial and temporal feature extraction, followed by attention-based feature fusion. Experiments on real-word voyage data show that TSBG-Para outperforms 20 benchmark models, achieving up to 46.3 % reduction in MSE under stable operating conditions. It also maintains robustness under dynamic conditions, with a MSE of 0.0719. Compared to serial counterparts, the parallel architecture reduces MSE and RMSE by 28.3 % and 15.1 %, respectively. Ablation studies further demonstrate that the SE and GA modules jointly enhance feature discrimination and improve prediction stability. These results underscore the effectiveness of parallel, attention-enhanced architectures for ship energy prediction and provide a scalable foundation for intelligent maritime energy management.