A hybrid model for ultra-short-term vessel motion prediction based on dynamic time series decomposition and recurrent neural networks

Ultra-short-term vessel motion prediction is essential for offshore decision-making and risk mitigation. With advances in artificial intelligence, hybrid modeling that integrates time series decomposition and neural networks has become an effective approach for developing such prediction models. This study proposes a novel dynamic time series decomposition algorithm (DTSDA) designed to overcome the limitations of existing decomposition techniques, including high complexity, multiple parameters, and decomposition time dependence on data length. Using a single hyperparameter (decomposition step), DTSDA achieves near real-time decomposition ( ∼ 0.1 ms) with low computational cost and ease of implementation. Validation with model test data demonstrates that hybrid models combining DTSDA with long short-term memory (LSTM) or gated recurrent unit (GRU) (i.e., DTSDA-LSTM and DTSDA-GRU) accurately predict heave, roll, and pitch motions. Across various prediction horizons (9.6 – 15.36 s), sea states (wave steepness: 0.0232 – 0.0349), and wave directions (180°and 210°), both models outperform optimized LSTM and GRU, while DTSDA-GRU exhibits strong generalization across sea states. Moreover, compared with several benchmark models, DTSDA-GRU maintains high accuracy without compromising computational efficiency. Overall, DTSDA simplifies hybrid modeling and significantly enhances ultra-short-term vessel motion prediction. • Dynamic time series decomposition (DTSDA) is proposed for vessel motion prediction. • DTSDA significantly reduces model learnable parameters and simplifies model tuning. • DTSDA-based hybrid model predicts heave, roll, and pitch up to 15.36 s. • DTSDA exhibits strong applicability and generalization in high sea states.