Ship Movement Prediction in Maritime Traffic using Machine Learning Methods

Abstract

Accurate ship trajectory prediction based on Automatic Identification System (AIS) data is essential for modern maritime safety, maritime traffic management, and autonomous vessel navigation. With the rise of Maritime Autonomous Surface Ships (MASS), improving maritime safety has become increasingly critical, with trajectory prediction playing a vital role in collision avoidance systems. Because AIS is widely adopted, most modern prediction approaches rely on forecasting future movements from historical trajectory data.

Various deep learning architectures have been explored for ship trajectory prediction. This thesis focuses on evaluating four specific deep learning models: Bidirectional Long Short-Term Memory (Bi-LSTM), Gated Recurrent Unit (GRU), Temporal Convolutional Network (TCN), and Temporal Fusion Transformer (TFT).

The primary objective of this thesis is to hyperparameter-optimize and compare these four models when predicting ship trajectories in the Baltic Sea. Regarding optimization, the analysis focuses on how unit and dropout values affect model performance. For performance evaluation, Mean Absolute Error (MAE), Mean Squared Error (MSE), Root Mean Squared Error (RMSE), and $R^2$ Score are used to evaluate variations within the same architecture as well as across different architectures.

The experimental results highlight specific scenarios where models failed to predict accurate trajectories. These failures were attributed to several factors, including data selection choices regarding both timeframe and vessel types, as well as the use of unsuitable training loss criteria such as MAE.