An adaptive-weighted knowledge-embedded neural network for enhancing ship energy consumption prediction performance

Artikel i vetenskaplig tidskrift, 2026

Accurate prediction of ship energy consumption is essential for improving operational energy efficiency and supporting emission reduction in maritime transport; however, traditional black-box models often achieve strong data-fitting performance but exhibit limited extrapolation capability and reliability. To address this limitation, this study proposes an adaptive-weighted knowledge-embedded neural network (AKENN) for ship energy consumption prediction. First, prior knowledge of ship energy consumption is automatically discovered from operational data using polynomial regression and embedded into the neural network loss function as a constraint. Then, the weights of knowledge loss and data loss are dynamically optimized based on logarithmic uncertainty theory to balance knowledge constraints and data-driven learning. Finally, experiments using operational data from a large container ship show that AKENN improves the comprehensive performance indicator (CPI) by 10.5%-66.5% compared with mainstream benchmark models, demonstrating improved robustness and extrapolation capability for ship energy consumption prediction.