Physics-informed residual learning for ship speed prediction across wind–wave–current scenarios

Accurate ship-speed prediction under heterogeneous wind, wave and current forcing is important for voyage optimisation, energy-efficiency assessment and operational decision support. This study is framed as a single-vessel, single-voyage AIS–metocean diagnostic study rather than as an operational model with demonstrated cross-voyage or cross-vessel generalisation. The study compares a non-learning Physics model, a data-driven model (DDM), and a physics-informed hybrid-driven model (PHDM) in which a simplified physics-based speed-loss baseline is corrected by a data-driven residual learner. Self-organising map (SOM) regimes are used as diagnostic environmental strata rather than as a general-purpose sea-state classification. Using the predefined C1–C5 scenarios of the SHANGHAI EXPRESS voyage dataset, the exploratory post-hoc best-learner comparison gives an average MAPE of 3.0168% for PHDM, compared with 7.5794% for the Physics model and 6.6594% for DDM. In C4, MAPE decreases from 13.4291% for the Physics model and 12.9513% for DDM to 5.4178% for PHDM. These results provide the primary scenario-wise description of model performance within the analysed voyage. In C1, the comparison is metric-dependent: the Physics model has lower RMSE and MSE, whereas PHDM has lower MAE and MAPE. A fixed-reference stress test uses chronological partitioning, matched learner identity, and non-position feature sets to examine the stability of the scenario-wise pattern. Positive PHDM gains are observed for all fixed learners in C4 and for most fixed learners in C5, whereas the pattern differs across C1–C3. These findings apply to the analysed vessel, voyage period, validation design, and SOM-derived environmental regimes.