A temporal prediction model for ship maneuvering motion based on multi-head attention mechanism

In real-world maritime conditions, accurately predicting ship maneuvering motion over ultrashort periods can enhance the development of more precise vessel control algorithms. This study proposes a prediction model for ship maneuvering that utilizes a multi-head attention mechanism. To form a basis for this data-driven approach, we gathered training data from turning and zigzag tests of a ship conducted in natural sea environments. The model's predictive capability is assessed by comparing two strategies over a fixed 30-time-step forecast horizon: iterative multi-step forecasting (IMS) and direct multi-step forecasting (DMS). Notably, the model employs distinct attention layers to examine the temporal relationships among various motion variables across the three degrees of freedom in the output. Moreover, it uses separate attention layers to explore the distinct time response traits of the input motion states and control signals. The results demonstrate that different motion state variables display unique temporal correlations under different signal sequences, and the proposed network architecture effectively captures these temporal response traits.