Maritime vessel movement prediction: A temporal convolutional network model with optimal look-back window size determination

• Introduced a novel framework to determine the optimal look-back size of AIS data tailored for user-specific future time intervals. • A user-centric, adaptive framework for vessel movement prediction, which integrates real-time AIS data through an API. • Emphasized rigorous data pre-processing for AIS data. • Leveraged a Temporal Convolutional Network (TCN) for enhanced and precise maritime vessel movement prediction. • Validated the proposed method using a comprehensive month-long AIS dataset from a Finnish cargo ship (IMO 468891) in the Baltic Sea. • Demonstrated the efficacy of the approach in two distinct prediction scenarios: achieving optimal results with a 2-day look-back for 1-hour forecasts and a 4-day look-back for 5-hour forecasts. • Findings present significant implications for the maritime industry, emphasizing potential benefits like optimized maritime operations, reduced transit times, and decreased operational costs. Ship movement prediction models are crucial for improving safety and situational awareness in complex maritime shipping networks. Current prediction models that utilize Automatic Identification System (AIS) data to forecast ship movements typically rely on a fixed look-back window size. This approach does not effectively consider the necessary amount of data required to train the models properly. This paper presents a framework that dynamically determines the optimal look-back window size for AIS data, tailored to user-defined prediction intervals. Initially, a DBSCAN clustering method, along with various pre-processing techniques, has been employed to efficiently eliminate non-essential data points and address noise in the raw AIS data. Following this, Temporal Convolutional Networks (TCNs) have been trained using the dynamic characteristics of ship movements based on one month of AIS data (April 2023) collected from the Baltic Sea, evaluating various look-back window sizes to identify the optimal size required for predictions. Subsequently, the framework has been tested using an additional AIS dataset in two scenarios: 1-hour and 5-hour predictions. The experimental results indicate that the proposed framework can effectively select the necessary AIS samples for forecasting a ship’s future movements. This framework has the potential to optimize prediction services by identifying the ideal look-back window size, thereby providing maritime agents with high-quality and accurate predictions to enhance their decision-making processes.