A pioneering approach for improving ship operational energy efficiency: The quantitative application of deep learning interpretable results

Although various artificial intelligence methods have been widely applied in fuel consumption prediction and operational optimization, no clear transformation path has been formed in the specific application of interpretable analysis, which directly restricts further improvement of ship energy efficiency. This paper pioneers the classification of interpretable result applications into qualitative and quantitative levels, while innovatively provides specific measures for the quantitative application. Firstly, based on actual navigation data, an AO-Gandalf fuel consumption prediction model that can well coordinate accuracy and interpretability is constructed. The interpretable qualitative results are then obtained, which are used to define the types of operational parameters to be optimized. Secondly, a route segmentation method integrating time-aware windows and time-effect aggregation voting is proposed to form an operational parameters optimization framework with reasonable spatiotemporal granularity. Finally, leveraging the interpretable quantitative weights of each operational parameter, a weight-stratified Gaussian mutation strategy is formulated. Combined with a cosine-annealed recombination crossover strategy and an enhanced elite mechanism, a non-parametric IAEGA with dynamic response matching capability is obtained to optimize the operational parameters during navigation and provide direct energy efficiency improvement guidance for two ships. Results demonstrate that applying this pioneering approach enhances operational energy efficiency by 6.41 % and 7.05 %, respectively, effectively reducing fuel consumption and CO₂ emissions. This study provides crucial references for ship energy conservation and emission reduction as well as the energy efficiency improvement of similar complex systems, while offering illuminating methodologies and knowledge evolution pathways for the classical “prediction-optimization” solution paradigm.