Ship fuel consumption prediction based on transfer learning: Models and applications

Abstract

Data-driven fuel consumption rate (FCR) prediction models largely depend on the amount of training data, which can be scarce for new ships with limited operating time. To tackle this issue, we implement three transfer learning strategies to leverage knowledge from another seven container ships to construct artificial neural network (ANN)-based FCR prediction models for a target ship with limited data. Numerical experiments reveal that the ANN models incorporating the three transfer strategies outperform the model trained solely on the target ship data, reducing mean absolute percentage error by 12.57%, 6.44%, and 16.03%, respectively. This study also investigates the impacts of target dataset size on the performance of transfer strategies using ship FCR prediction as an example, revealing that the smaller amount of available data, the greater improvement in prediction accuracy using the transfer strategy. These insights contribute to the development of effective operational solutions for enhancing ship energy efficiency and promoting sustainable shipping practices.