Ship Trajectory Prediction Based on Spatial-temporal Data Using Long Short-Term Memory

Abstract

The frequent exploitation of shipping lines by passengers increased traffic and exposed it to more significant dangers. Precise predictions for ship trajectory conditions at sea must be available to ensure safe navigation across the oceans. This article presents a trajectory prediction approach based on Long Short-Term Memory (LSTM) neural networks applied to time series Automatic Identification System (AIS) position data, expressed in spatial-temporal form. LSTM is highly suitable for ship trajectory predictions as it can capture long-term dependencies and spatial-temporal patterns existing in AIS data, since LSTM is targeted toward sequential data. The proposed model extracts ship trajectories from AIS data and utilizes an LSTM (Long Short-Term Memory) model to predict future ship movements based on historical patterns. The experiments demonstrate that it is effective in predicting where ships to navigate next, providing a valuable tool for enhancing traffic flow and improving navigation safety. The model with LSTM unit 500, tested on 3,478 ship trajectories, showed a median RMSE prediction error ranging from 0.0720 to 0.0841, with prediction M=8 coordinate a head having the highest error (0.0841) and M=2 and M=9 having the lowest (0.0720); the interquartile range (IQR) spanned from 0.0571 to 0.1006, and M=2 had the most outliers (302) while M=8 had the fewest (171), indicating varying prediction stability across different points. Despite these results, challenges remain in maintaining prediction stability across all points. Further optimization could enhance the model's performance and address these limitations by incorporating more complex spatial-temporal features or hybrid techniques.