Confidence-Weighted Explainable Learning for Vessel Fuel-Consumption Analysis: A Multi-Model Comparative Framework
Abstract Improving vessel energy efficiency is essential for reducing fuel consumption, greenhouse-gas emissions, and operating costs while supporting maritime decarbonisation objectives. Although advances in maritime analytics and machine learning have improved vessel monitoring and fuel-consumption prediction, existing studies focus mainly on predictive accuracy and offer limited insight into the engineering, operational, and reliability factors that are associated with vessel energy demand. In addition, confidence-aware learning, uncertainty quantification, explainable artificial intelligence (XAI), and comparative evaluation of machine-learning models remain underexplored within maritime energy-intelligence research. This paper presents a confidence-weighted explainable learning framework for vessel fuel-consumption analysis and conducts a systematic comparison of five model families, namely Random Forest (RF), Extreme Gradient Boosting (XGBoost), Gaussian Process Regression (GPR), Bayesian Ridge Regression (BRR), and Elastic Net, within a unified confidence-aware, explainable, and uncertainty-aware analytical architecture. Confidence integration is formulated as a reliability-weighted signal-fusion problem in which vessel-level confidence scores derived from heterogeneous data sources are incorporated directly into model learning. The Vessel Energy Efficiency Intelligence (VEEI) combines statistical driver analysis, duty-cycle intelligence, explainable machine learning, and uncertainty quantification to generate interpretable vessel-level energy intelligence. Experiments use a tidal River Thames inland-waterway dataset comprising 26 representative vessels and 131 engineering, operational, propulsion, fuel-consumption, and confidence-related attributes. Spearman rank correlation analysis identifies operating hours ($OH$) and installed engine power ($EP$) as the only statistically significant direct fuel-consumption drivers ($p<0.05$), while medium-cruise operation exhibits the strongest significant duty-cycle association with fuel demand. Among the evaluated models, Bayesian Ridge Regression achieves the highest predictive performance with an MAE of 41.23~L, an RMSE of 59.87~L, and an $R^{2}$ of 0.881, followed by XGBoost ($R^{2}=0.876$) and Random Forest ($R^{2}=0.873$). Confidence-weighted learning consistently improves predictive performance across all model families relative to their unweighted counterparts. Cross-model SHAP analysis identifies vessel draft and duty-cycle behaviour as the dominant explanatory factors, while uncertainty analysis consistently highlights vessels with elevated predictive uncertainty. The results show that operational utilisation, propulsion characteristics, vessel geometry, duty-cycle behaviour, and data reliability jointly influence vessel fuel consumption. By integrating confidence-aware learning, multi-model explainability, and uncertainty quantification in a unified framework, the proposed approach provides transparent, trustworthy vessel-level energy intelligence to support operational optimisation, fleet benchmarking, and evidence-based maritime decarbonisation planning.
Authors
- Syed Mohsen Naqvi (ORCID: https://orcid.org/0000-0002-1547-0908)
- Douglas Amobi Amoke (ORCID: https://orcid.org/0009-0007-7304-6172)
- Paul Simavari (ORCID: https://orcid.org/0009-0005-8817-4877)
Institutions
- Newcastle University (GB)
Publication Details
- Journal
- Research Square
- Published
- 2026-11-13
- DOI
- https://doi.org/10.21203/rs.3.rs-11388094/v1
- Primary Topic
- Maritime Transport Emissions and Efficiency
- Type
- preprint