Aiming at the low prediction precision and poor adaptability of ship models based on the data-driven modeling strategy, an enhanced bi-directional long short-term memory (Bi-LSTM) model is proposed for the high-precision non-parametric modeling of ships.

First, the feature extraction of the bi-directional time dimension is realized using bi-directional long short-term memory (Bi-LSTM) neural networks. On this basis, the spatial dimension features of the one-dimensional convolutional neural network （1D-CNN） extraction sequence are designed. Then, a multi-head self-attention (MHSA) mechanism is used to deal with the sequence from multiple angles. Finally, using the navigation data of KLVCC2 ships, the prediction effects of the enhanced Bi-LSTM model are compared with those of the Support Vector Machine (SVM), Gate Recurrent Unit (GRU), and long short-term memory (LSTM) models.

The Root Mean Square Error (RMSE) and Mean Absolute Error (MAE) performance indicators of the enhanced Bi-LSTM model in the test set are lower than 0.015 and 0.011 respectively, and the coefficient of determination（R²）is higher than 0.99913, demonstrating prediction accuracy significantly higher than that of the SVM, GRU, and LSTM models.

The proposed enhanced Bi-model has excellent generalization performance and excellent prediction stability and precision, and effectively realizes ship motion identification.