Kinematic-aware AIS preprocessing and bidirectional recurrent modeling for multi-output vessel trajectory prediction

Abstract Precise prediction of vessel movements is crucial for effective maritime traffic monitoring and management. Automatic Identification System (AIS) data offer rich information for trajectory prediction but suffer from irregular sampling, missing values, noise, and transmission errors that complicate modeling. To address these challenges, we propose a two-stage prediction framework. First, a comprehensive preprocessing pipeline performs data cleaning, trajectory extraction, kinematic-aware anomaly correction, and temporal resampling to produce consistent, analysis-ready trajectories. Second, a Bidirectional Gated Recurrent Unit (BiGRU) network jointly predicts future vessel position, Speed Over Ground (SOG), and Course Over Ground (COG), with a specialized cosine distance for the circular COG variable. We benchmark the proposed model against unidirectional Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU), bidirectional LSTM (BiLSTM), and attention-augmented variants of the GRU, BiGRU, and BiLSTM encoders under identical preprocessing and training protocols. A systematic ablation study further confirms that each preprocessing stage contributes measurably to prediction accuracy, validating the importance of rigorous data denoising prior to model training. We evaluate the framework on two real-world AIS datasets from distinct maritime regions: the port of Brest, France, and the Danish Maritime Authority (DMA) dataset, demonstrating generalization across different traffic conditions and vessel populations. These results highlight that combining thorough, kinematic-aware AIS preprocessing with a well-chosen recurrent architecture yields accurate multi-output vessel trajectory prediction across diverse maritime environments.