Extreme Short-Term Prediction of Ship Motions Based on LSTM Recurrent Neural Network
Researchers are trying to apply long short-term memory(LSTM)recurrent neural network in extreme short-term prediction of ship motions owing to its inherent advantages in handling nonlinear time series.This study is dedicated to examining prediction accuracy and effective duration employing the field data of the rolling and pitching motions of an icebreaker navigating at two typical sea conditions in the Arctic Ocean.It has demonstrated that the predictions adopting LSTM model are satisfying when the ship navigates in favorable sea conditions and its motions are charactered by periodicity with dominant frequencies.As expected,the prediction accuracy decreases with the increase of prediction time ascribed to the error accumulations.An improved direct multi-step prediction model is proposed by use of a multilayer perceptron based on the classical encoder-decoder iterative framework.Compared to traditional LSTM model,the new model can effectively reduce error accumulations,improve prediction accuracy and increase the effective prediction time,especially when the ship navigates at rough sea conditions.The study can be a workbench for developing an effective model based on neural network for extreme short-term prediction of ship motions before practical applications.
NETL
NSTL
万方数据
