Title:TrAISformer-A generative transformer for AIS trajectory prediction

Abstract:The prediction of vessel positions at a specified point in the future is a fundamental aspect of many maritime applications. While Automatic Identification System (AIS) provides a rich source of information to enable this task, vessel trajectory forecasting using AIS data remains formidably challenging, even for modern machine learning/deep learning, because of the complexity and multimodality inherent in motion data. In this paper, we address these challenges by introducing a novel discrete, high-dimensional representation of AIS data and a new loss function to explicitly account for heterogeneity and multimodality. The proposed model -- referred to as TrAISformer -- is a modified transformer network that extracts long-term correlations of AIS trajectories in the proposed enriched space to forecast the positions of vessels several hours into the future. We report experimental results on publicly available, real AIS data. TrAISformer significantly outperforms state-of-the-art methods and with an average prediction performance below 10 nautical miles up to ~10 hours.