An extremely short-term ship motion prediction method based on an IDE-optimized dynamic coupling filtering graph neural network

Extremely short-term ship motion prediction can provide feedforward information for wave-compensation systems, but accurate multi-step forecasting remains challenging because of nonlinear, time-varying, and dynamically coupled responses. This study proposes an improved differential evolution-optimized dynamic coupling filtering graph neural network (IDE-DCF-GNN). Fine-grained patch nodes and Temporal, Spatial, and Spatial-Temporal candidate edges characterize temporal continuity, synchronous cross-DOF coupling, and lagged interactions, while region-wise gating, Top-k sparsification, and improved differential evolution adaptively retain informative connections and optimize key hyperparameters. On ANSYS/AQWA data, IDE-DCF-GNN achieves the best or second-best results for most DOFs and metrics; all RMSE/ROM values are below 2.25% for 3s forecasting, with average values of 2.2655% and 3.3278% at 5s and 7s. Cross-condition tests on unseen simulated conditions and measured offshore-platform data further verify robustness. On public ACE real-ship data, six measured states—Heading, Roll, Pitch, North Speed, East Speed, and Vertical Speed—are jointly predicted for 3–10s horizons from a high-speed training condition to an unseen low-speed condition. At 10s, all six RMSE/ROM values remain below 10%, demonstrating multi-horizon real-ship forecasting and cross-speed generalization.