Enhancing transfer learning in vessel shaft power prediction: Addressing catastrophic forgetting and feature integration

To meet global emission reduction targets, the international shipping industry must enhance its energy efficiency. Predictive models are essential for optimizing strategies such as weather routing, hull cleaning scheduling, and just-in-time arrival, all contributing to reduced fuel consumption and emissions. This paper builds on a previously introduced transfer learning framework that integrates physics-based and data-driven models for vessel shaft power prediction under varying operating conditions. The present study extends this framework by addressing two key challenges that limit its applicability in real-world settings: (i) catastrophic forgetting during fine-tuning and (ii) mismatch between source and target feature spaces. To address these challenges, the study evaluates parameter-efficient and regularization-based fine-tuning strategies (LoRA and EWC) and introduces two architectural modifications (TL-Ext and TL-Concat) to enable the integration of additional input features. The contribution lies in the integration and systematic evaluation of these methods within a regression-based, physics-informed transfer learning framework. Three use cases are examined, including controlled scenarios with synthetic ground truth and a real-world case study. The results show that the proposed enhancements mitigate catastrophic forgetting and improve extrapolation performance, particularly under limited training data conditions. The method remains effective when the target domain includes additional input features not present in the source domain. Validation on real-world data demonstrates competitive or improved performance compared to a purely data-driven model, while exhibiting increased robustness and reduced data requirements. These findings highlight the applicability of the framework in real-world operational settings.