Physics-Structured Residual Learning for Ship Maneuvering Prediction: Multi-Source Disturbance Decomposition and Compensation

Ship maneuvering models based on MMG or Abkowitz formulations often suffer from systematic mismatches under real operating conditions, where shallow water, hull fouling, rudder degradation, and wind loads may coexist. This study proposes a physics-structured residual learning framework for multi-source disturbance decomposition and compensation. Disturbance-specific expert networks are introduced to map different disturbance sources into separate residual channels. A CNN-SE-BiLSTM encoder is further designed to estimate the slowly varying latent disturbance states from residual sequences, whereas wind is treated through an external pathway owing to its directly measurable and higher-frequency nature. Simulations on the KVLCC2 benchmark vessel under single-source, triple-source, and wind-inclusive disturbance scenarios demonstrate stable long-horizon closed-loop autoregressive prediction, with position-RMSE reductions of 74.7–91.7% relative to the corresponding nominal-MMG and wind-ablation baselines. These results indicate that the proposed physics-structured residual learning framework improves long-horizon prediction accuracy while retaining interpretable and modular disturbance-specific correction channels under complex operating conditions.