基于LSTM循环神经网络的船舶运动极短期预报  

Extreme Short-Term Prediction of Ship Motions Based on LSTM Recurrent Neural Network

作　　者：张怡 孟帅[1] 刘震 封培元[2] ZHANG Yi;MENG Shuai;LIU Zhen;FENG Peiyuan(State Key Laboratory of Marine Engineering,Shanghai Jiao Tong University,Shanghai 200240,China;Marine Design and Research Institute of China,Shanghai 20001l,China)

机构地区：[1]上海交通大学海洋工程国家重点实验室,上海200240 [2]中国船舶及海洋工程设计研究院,上海200011

出　　处：《船舶工程》2024年第5期33-40,55,共9页Ship Engineering

基　　金：国家自然科学基金面上项目(51879161,52271335)。

摘　　要：长短期记忆(LSTM)循环神经网络对于预报非线性时间序列有优势,尝试将LSTM网络应用于船舶运动极短期预报。利用某破冰船在北冰洋航行时两段典型海况下的横摇和纵摇运动实测数据,探究LSTM神经网络模型的预报精度和有效时长。研究发现,LSTM神经网络模型在海况良好、船舶的运动周期性强且主导频率突出时可以取得满意预报效果。但随着时长的增加,误差会不断累积,精确度逐步降低。在编码器-解码器逐步迭代框架基础上利用多层感知机提出直接多步预报改进模型。研究发现,改进模型可以有效减少误差积累、提高预报精度和延长有效时间,尤其在恶劣海况下预报结果改善更为显著。研究成果可以为基于神经网络开发高效准确的船舶运动极短期预报方法提供参考。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.

关 键 词：极短期预报 长短期记忆循环神经网络 直接多步预报 误差积累 

分 类 号：U664.82[交通运输工程—船舶及航道工程]