Adaptive fuel consumption prediction for wing-assisted vessels based on incremental learning

Precise fuel consumption prediction is key to reducing energy consumption and carbon emissions in maritime transportation. However, the performance of conventional prediction models tends to deteriorate over time due to evolving operational conditions, such as hull fouling and propeller wear. To solve this problem, this study proposes an adaptive prediction framework for fuel consumption of wing-assisted vessels, based on incremental learning. First, a hybrid kernel-function-based support vector regression model tree is developed, incorporating an online data-update mechanism to enable continuous model evolution. On this basis, the Robust Selective Instance Regression (RSIR) algorithm is introduced for ship fuel consumption modeling. A joint prediction method combining RSIR and Stochastic Gradient Descent (SGD) is then established, integrated with a Statistical Process Control (SPC) mechanism to proactively monitor model performance and trigger updates when deviations occur. The proposed RSIR-SGD algorithm is evaluated against several benchmark methods using operational data from wing-assisted vessels. The results demonstrate that the RSIR-SGD algorithm achieves the lowest prediction errors, with an RMSE of 0.0554 and an MAE of 0.0421. The average deviation between predicted and actual values is consistently smaller than that of comparative algorithms, indicating superior accuracy and adaptability. This study provides an effective and practical solution for real-time fuel consumption prediction under dynamically changing navigational conditions.