Generalisation, Physical Plausibility, and CII-Proxy Sensitivity in Ship Fuel-Consumption Prediction: A Three-Layer Interpretable Machine-Learning Framework for Maritime ITS

Maritime decarbonisation requires fuel-consumption models that are accurate, physically plausible, and suitable for operational decision support. We develop a three-layer interpretable machine-learning framework combining predictive validation, physical-plausibility auditing, and operational aggregation. Using 489,620 speed-stable cruising observations from 21 vessels, the predictive layer separates record-level interpolation from temporal and cross-vessel transfer; RMSE increases by about threefold under stricter validation, showing that random splits overstate transferability. The explanatory layer audits SHAP-derived responses against pre-specified physical expectations and empirical support. Speed responses are strongly monotonic and nonlinear, the primary wave perturbation is not confirmed under the locked C4 rule, and the rudder response is empirically unidentifiable. The operational layer evaluates speed-reduction sensitivity using an observation-period main-engine CII proxy. Fixed-time responses are heterogeneous across vessel groups and reproduced across XGBoost, LightGBM, and Random Forest models; deeper perturbations progressively leave multivariable training support. Paired ERA5 controls retain the principal transfer and wave-audit conclusions but do not remove pathway–vessel-type confounding. These findings are local, model-based sensitivities in the studied 21-vessel sample and do not estimate a statutory annual CII or establish causal vessel-type mechanisms. The framework provides an auditable validation protocol for maritime ITS; applicability beyond shipping requires independent validation.