Accurate prediction of fuel consumption is critical for achieving efficient and low-carbon ship operations. However, the variability of the marine environment introduces significant challenges, as it leads to dynamic changes in monitoring data, complicating real-time and precise fuel consumption prediction. To address this issue, the authors proposed an incremental learning-based prediction framework to enhance adaptability to temporal dependencies in fuel consumption data. The framework dynamically adjusts a dual adaption mechanism for input features and target labels while incorporating rolling retraining to enable continuous model updates. The effectiveness of the proposed approach was validated using a real-world dataset from an LPG carrier, where it was benchmarked against conventional machine learning models, including Random Forest (RF), Linear Regression (LR), Support Vector Machine (SVM), and Multi-Layer Perceptron (MLP). Experimental results demonstrate that the proposed approach could significantly improve prediction accuracy in both offline and online scenarios. In offline mode, the proposed framework improves the R² of various machine learning models by at least 21.97%. In online mode, the proposed method increases R² by at least 17.97%. This work provides a new solution for real-time fuel consumption prediction in dynamic marine environments.