Ship Trajectory Prediction Method Based on Multi‐Layer Recurrent Neural Network Structure and AIS Data Driven

ABSTRACT In the current era, improving the intelligence level of vessels and ensuring the construction of a safer and more reliable maritime traffic environment has become an extremely crucial task. And intelligent vessel trajectory prediction undoubtedly impacts the intelligent navigation and collision avoidance systems of vessels. However, unfortunately, in the past few decades, the analysis work on massive trajectory data has been relatively scarce. At the same time, whether the current research focus on vessel trajectory prediction is short‐term or long‐term, it has led to the situation that the accuracy of trajectory prediction is far from satisfactory. In view of this, this study innovatively introduces a data‐driven Long Short‐Term Memory (LSTM) approach for the Automatic Identification System (AIS). This method realizes the accurate prediction of the entire vessel trajectory through the fusion of forward and reverse sub‐networks (named FRA‐LSTM here). Specifically, the forward sub‐network in our proposed method cleverly combines LSTM with an attention mechanism to accurately extract key factors from the forward past trajectory data. Correspondingly, the reverse sub‐network organically integrates the attention mechanism with a Bidirectional LSTM (BiLSTM) to simultaneously mine the unique characteristics of the backward historical trajectory data. Finally, the features output by the forward and reverse sub‐networks are combined so as to successfully construct the final expected trajectory. After a large number of comprehensive and in‐depth tests, we are delighted to find that compared with BiLSTM and Seq2Seq, the method proposed in this study has achieved an average increase of 96.8% and 86.5% regarding the accuracy of short‐term and mid‐term trajectory prediction respectively. More importantly, in the domain of long‐term trajectory prediction, the average accuracy of our method is as high as 90.1% higher than that of BiLSTM and Seq2Seq, showing excellent performance advantages.