Session: 15-01-02 Seakeeping and Operability

Submission Number: 183625

Quantifying Uncertainty in Recurrent Neural Network-Based Models for Ship Motion Time Series Prediction 

Accurate short-term prediction of ship motions is essential for ensuring safe maritime operations, dynamic positioning control, and decision-support systems operating under stochastic sea environments. Recurrent neural networks (RNNs), such as the long short-term memory (LSTM) and gated recurrent unit (GRU) architectures, have demonstrated strong capabilities in capturing the complex temporal dependencies of nonlinear ship dynamics. However, the quantification and comparative evaluation of their predictive uncertainties under varying input-space configurations have not been sufficiently explored.

This study conducts a systematic comparison of uncertainty estimation methods applied to bidirectional LSTM (Bi-LSTM) and bidirectional GRU (Bi-GRU) models for predicting six-degree-of-freedom (6-DOF) ship motions in irregular waves. A heteroscedastic Gaussian negative log-likelihood (NLL) formulation is adopted to jointly learn the mean and log-variance of the predictive distribution, allowing for the decomposition of total uncertainty into aleatoric (data-related) and epistemic (model-related) components.

The models are trained using 64 distinct sea states generated from impulse-response-function (IRF)-based ship–wave interaction simulations that span a wide range of significant wave heights and peak periods. Three input configurations are considered to assess the influence of information richness on prediction performance: a single-motion time series representing sensor-limited or minimal data conditions, full 6-DOF motion inputs exploiting coupled kinematic information, and a combined configuration incorporating 6-DOF motions with concurrent wave-elevation records to capture exogenous excitation effects. Optimal hyperparameters are identified through Optuna-based Bayesian optimization, and network weights trained for mean prediction are extended to produce both mean and log-variance outputs without complete retraining. This zero-shot expansion allows immediate estimation of aleatoric uncertainty, while epistemic uncertainty is approximated using Monte Carlo dropout.

The comparative investigation examines how different input configurations impact the sharpness and reliability of predictive distributions, as well as how uncertainty characteristics vary across different sea-state conditions. Aleatoric and epistemic uncertainties are analyzed and contrasted under both calm and extreme wave scenarios to reveal their relative contributions and potential implications for model generalization. In addition, both Bi-LSTM and Bi-GRU models are evaluated in terms of their capability to provide well-calibrated uncertainty estimates, with attention to differences arising from their respective memory and gating mechanisms.

Finally, gradient-norm analysis is conducted across all motion components to identify how the learning process distributes sensitivity among the degrees of freedom, offering insights for potential improvements in sensor placement and feature representation. The proposed framework provides a flexible extension to existing deterministic RNN-based marine motion predictors, enabling the incorporation of uncertainty quantification into real-time ship-motion forecasting systems. By facilitating efficient zero-shot heteroscedastic expansion, the approach supports uncertainty-aware deployment on computationally constrained onboard systems, enhancing the reliability and safety of future autonomous surface operations.

Presenting Author: Yangjun Ahn Sungshin Women's University

Presenting Author Biography: Dr. Yangjun Ahn is an Assistant Professor in the Department of Artificial Intelligence at Sungshin Women’s University. His research focuses on ship motion dynamics, artificial intelligence, and the interpretable analysis of time-series models. He investigates how complex, nonlinear hydrodynamic behaviors of ships can be accurately predicted and understood through advanced AI methodologies. His recent work includes the development of deep learning models for predicting ship motions and sloshing loads, as well as collaborative research on uncertainty quantification and explainable neural networks for maritime applications. He has contributed AI-based software tools to Hanwha Ocean and HD Hyundai Heavy Industries and has provided expert consultation on marine safety and prediction systems. Dr. Ahn actively serves as a reviewer for leading international journals, including Marine Structures and Ocean Engineering, as well as other major publications in the fields of ship hydrodynamics and marine AI applications. He is a recipient of the 2021 Songam Award from the Society of Naval Architects of Korea, recognizing his contributions to the advancement of maritime science and artificial intelligence integration. His research has been published in high-impact journals, including Marine Structures, Ocean Engineering, Journal of Fluid Mechanics, Proc IMechE Part M: Journal of Engineering for the Maritime Environment, and the International Journal of Offshore and Polar Engineering.

Authors:

Yujin Seo Sungshin Women's University
Yangjun Ahn Sungshin Women's University