An integrated method of advanced optimisation and adaptive ensemble learning for ship fuel consumption prediction

• Proposes an adaptive cluster-based multi-ensemble (ACME) model for robust ship fuel consumption prediction. • Develops a SHAP-weighted multi-model feature selection (SWFS) algorithm for dimensionality reduction. • Introduces a hierarchical adaptive parameter space exploration (HAPSE) method for efficient tuning. • Establishes a dual-layer SHAP framework for global and local model interpretability. • Integrates multi-source data fusion to enhance prediction accuracy and operational relevance. Accurate prediction and interpretable analysis of Ship Fuel Consumption (SFC) are critical for optimising maritime operations and supporting decarbonisation efforts in maritime transport, yet existing approaches face significant challenges including limited model generalisation, redundant features due to multi-source data integration, and a lack of transparency in model outputs. These limitations stem from fragmented modelling pipelines that fail to holistically address the challenges of data heterogeneity, feature relevance, parameter tuning, and model adaptability in complex maritime environments. To address these challenges, this study develops an integrated framework that comprises advanced optimisation techniques with adaptive ensemble learning, structured through a synergised pipeline of four technical components. Firstly, multi-source data fusion employs spatio-temporal alignment to integrate ship noon reports, Automatic Identification System data, ECMWF Reanalysis v5, and Global Ocean Physics Analysis and Forecast data, to construct a comprehensive feature space. Secondly, a SHAP-based Weighted Feature Selection algorithm leverages multi-model SHapley Additive exPlanations (SHAP) value assessment with recursive feature elimination to identify and eliminate redundant features, enhancing model generalisation and prediction efficiency. Thirdly, a Hierarchical Adaptive Parameter Space Exploration algorithm integrates global random search and local grid search for efficient hyperparameter optimisation. Finally, an Adaptive Cluster-based Multi-Ensemble model incorporates data clustering and model fusion to capture operational heterogeneity and adaptively assign optimal weights across different data clusters. Comparative experiments demonstrate that the proposed model significantly outperforms six mainstream machine learning models and three classical ensemble methods across multiple evaluation metrics. Moreover, SHAP-based interpretability analysis quantifies feature contributions and reveals specific effects of value changes, enhancing model transparency for decision support. This framework provides a robust technical solution for SFC prediction, offering reliable data-driven tools for energy efficiency management and sustainable maritime operations. The source code is publicly available at: https: //github. com/AdvMarTech/shipfuelconsumₚredic.