A Bi-directional LSTM Ship Trajectory Prediction Method based on Attention Mechanism

• Main Authors: Zhang, Sheng, Wang, Long, Zhu, Mingdong, Chen, Siwen, Zhang, Haisu, Zeng, Zhaowen
• Format: Conference Proceeding
• Language: English
• Subjects: attention mechanism; auto-encode; bi-directional LSTM; Bidirectional control; Data mining; Data models; Feature extraction; Marine vehicles; Predictive models; Trajectory; trajectory prediction
• ISSN: 2689-6621
• DOI: 10.1109/IAEAC50856.2021.9391059

Ship trajectory prediction can effectively predict navigation trends and realize orderly management of ships, which are of great significance to maritime traffic safety. This paper proposes a novel ship trajectory prediction method based on the integrated model of LSTM Auto-encode, Attention Mechanism, and Bi-directional LSTM (AABiL) structure. Our research in this paper mainly includes two parts, the first part is the extraction of ship trajectory features, and the second part is the trajectory sequence prediction. The LSTM auto-encoder used in this paper is to extract the features of the trajectory data, and trajectory data come from Automatic Identification System (AIS) of the ship. After preprocessing the trajectory data, we combined the extracted features with the trajectory longitude and latitude data to represent the current sailing state of the ship as the model input. The Bi-directional LSTM (Bi-LSTM) neural network model with attention mechanism is used to train and learn the regular pattern of ship motion implied in the trajectory data to predict the location of the ship at the next moment. Experiments demonstrate that use the integrated model of AABiL has decreased the Mean Square Error (MSE) of ship trajectory prediction to 0.00047, which can effectively improve the accuracy of trajectory prediction.