A Multi-decoder Recurrent Network for Vessel Trajectory Prediction Using Multi-memory LSTMs

Vessel trajectory prediction is the foundation for constructing the marine intelligent traffic management system. Deep learning algorithms driven by massive AIS data provide more possibilities for accurately predicting vessel trajectories. Currently, LSTM models widely used for this task have limited ability to extract temporal features, and prediction errors accumulate rapidly over time. To address the above issues, we design a multi-memory state LSTM (MM-LSTM) unit by introducing global memory states and use a multi-decoder structure to capture features for different prediction periods. Experimental results on the real AIS dataset provided by NOAA show that: first, the MM-LSTM unit can effectively capture the temporal information and reduce the prediction error of vessel trajectory by 1.93km (11.5%) than the standard LSTM; second, the multi-decoder structure can universally improve the trajectory prediction accuracy for each prediction period, which has promising applications in trajectory prediction tasks.