Enhancing transfer learning strategies for ship fuel consumption prediction under data scarcity

Reliable fuel consumption (FC) prediction is crucial for enhancing the energy efficiency of ships and achieving low-carbon shipping. However, the scarcity of individual ship data due to limited operation time or sensor failures remains a major obstacle to developing accurate data-driven models. This study proposes a transfer learning framework to address this challenge, which includes two model structures: bidirectional long short-term memory network (BiLSTM) and random forest (RF). By using the operation data of similar ships with sufficient historical records as the source domain, it supports FC prediction for target ships with limited data. Experimental results show that the performance of both transfer learning models is superior to that of the baseline model and the mixed data model. Compared with the baseline model, the MAE of the TL-BiLSTM and TL-RF models is reduced by 42 % and 36 %, respectively. The paper also innovatively and systematically analyzes the influence mechanism of the freezing strategy and the source-target sample ratio on the transfer performance. The proposed method provides an effective solution for FC prediction in data-scarce situations, can provide practical guidance for ship energy efficiency management. • Network-based BiLSTM transfer and mapping-based RF transfer learning are proposed. • Two control groups (baseline and mixed data) are designed to validate transfer effectiveness. • The impact of freezing strategies on BiLSTM transfer learning is quantitatively analyzed. • The source-to-target sample ratio is shown to affect transfer learning performance.