Online prediction method of ship maneuvering motion under wave influence based on improved LSTM neural network

To address the insufficient real-time capability and long-horizon accuracy degradation of ship maneuvering motion prediction under environmental disturbances such as waves, an online prediction method based on an improved Long Short-Term Memory (LSTM) neural network is proposed. A multi-layer LSTM is adopted as the core predictor, and an embedded sliding-window structure is introduced to compute the error metrics within the window in real time. When the window-averaged error exceeds a preset threshold, model retraining and updating are triggered, thereby achieving timely online prediction. The results indicate that, compared with offline prediction, the proposed online method maintains stable prediction accuracy under long-horizon conditions with continuously switching wave states. With the same window length, the online method with a stricter threshold achieves a maximum RMSE improvement of 56.85%, while the cumulative update time is only 3.82 s. The proposed online prediction method delivers satisfactory long-horizon prediction performance for ship maneuvering motion and shows practical value for accurate long-horizon prediction under complex sea conditions.