Abstract
Reliable collision-risk warning during low-speed ship docking requires accurate and efficient hydrodynamic prediction. This study develops a physics-SVR (support vector regression) framework combining a three-degree-of-freedom maneuvering model with three SVR models that learn residuals in longitudinal force, lateral force, and yaw moment from computational fluid dynamics (CFD) data. The corrected loads support 120 s trajectory prediction and hull-envelope reconstruction. Minimum lateral and longitudinal clearances and their times to safety-threshold crossing distinguish normal, warning-alert, and emergency-alarm states. The database comprises 20 model-development conditions and three condition-level holdout tests representing unseen loading/draft, heading, and lateral-offset conditions. Across the holdout tests, overall relative errors decrease from 22.00–30.20% for the baseline model to 14.91–19.93%. In two hazardous cases, warning alerts precede contact by 154.9 and 172.09 s, while the non-hazardous control case triggers no alarm. The complete prediction-and-warning update requires 0.32 s on average, demonstrating potential for shore-based docking assistance under calm-water conditions.