基于注意力机制的堆叠LSTM短时船舶交通流预测模型  

Stacked LSTMs short⁃term ship traffic flow prediction model based on attention mechanism

作　　者：廉清云[1] 孙伟[2] 李润生 LIAN Qingyun;SUN Wei;LI Runsheng(Merchant Marine College,Shanghai Maritime University,Shanghai 201306;College of Information Engineering,Shanghai Maritime University,Shanghai 201306)

机构地区：[1]上海海事大学商船学院,上海201306 [2]上海海事大学信息工程学院,上海201306

出　　处：《大连海事大学学报》2024年第1期57-65,共9页Journal of Dalian Maritime University

基　　金：交通运输部2021年度交通运输行业重点科技项目(2021⁃ZD6⁃095)。

摘　　要：针对由短时船舶交通流数据的非线性与非平稳性特征导致的预测精度低的问题,提出基于注意力机制的堆叠LSTM船舶交通流预测模型。采用堆叠式LSTM神经网络捕捉短时船舶交通流数据的时序特征,并通过引入注意力机制更好地学习全局性特征,以提高船舶交通流预测精度。提取长江下游三个航段的船舶AIS数据构建船舶交通流数据集,并将其用于本文模型的训练和测试。结果表明,相较HA、ARIMA、GPR、LSTM和Seq2Seq等基线模型,在交通流量宏观参数预测中,本文模型的均方根误差和平均绝对误差两个评价指标均有所降低;与最优基线模型相比,本文模型在船舶交通流预测中表现出更高的精度,其均方根误差降低4.05%,平均绝对误差降低4.04%。A stacked LSTM ship traffic flow prediction model based on attention mechanism was proposed to address the problem of low prediction accuracy caused by the nonlinear and non⁃stationary characteristics of short⁃term ship traffic flow da⁃ta.Stacked LSTM neural network was used to capture temporal features of short⁃term ship traffic flow data,and the attention mechanism was introduced to better learn global features,and improved accuracy of ship traffic flow prediction.Ship AIS data were extracted from three sections of the lower reaches of the Yangtze river to construct the ship traffic flow dataset,and used for training and testing of the proposed model.Results show that compared to baseline models such as HA,ARIMA,GPR,LSTM,and Seq2Seq,the model proposed in this paper,both the root mean square error and mean absolute error reduce in predicting macro traffic flow parameters.Compared with the optimal baseline model,the proposed model exhibits higher ac⁃curacy in predicting ship traffic flow,with a root mean square error reduction of 4.05%and a mean absolute error reduction of 4.04%.

关 键 词：内河交通 船舶交通流预测 长短时记忆网络(LSTM) 堆叠LSTM 注意力机制 

分 类 号：U692[交通运输工程—港口、海岸及近海工程] TP391[交通运输工程—船舶与海洋工程]