Deterministic real-time prediction of ship roll motion with quantified uncertainty based on machine learning

Abstract

Real-time prediction of the ship motion in advance can effectively enhance the safety and efficiency of maritime operations. However, the current prediction methods mainly focus on the motion time series forecasting without considering the uncertainty existing in measured motions. In this paper, a novel prediction method combined with the confidence interval forecasting of the motion is proposed. The method integrates the probability prediction module into the time-series prediction model. The normal distribution and student’s T-distribution are considered and the long short-term memory (LSTM) neural network is selected as the time-series prediction model. A set of measured full-scale ship roll motion of Yukun Ship is used to verify the prediction performance. The results demonstrate that the proposed method can effectively predict the confidence intervals of future ship motions, especially for extreme motions. This approach circumvents the issue of reduced accuracy over longer prediction periods, which is commonly existed in traditional time-series prediction models due to the influence of non-stationary characteristics of the data. Particularly, at the confidence level of 99 %, the prediction results could cover >90 % of the motion time series for future 12 s, which can significantly ensure the safety of offshore operations.