<p>In busy and complex maritime areas, navigation safety is of paramount importance, and predicting vessel trajectory accurately and identifying potential collision risk in advance is one of the effective means. Existing methods often struggle to capture the intricate spatial interactions and nonlinear temporal dependencies of vessels. To address this challenge, we proposed a novel trajectory prediction model based on an improved spatial-temporal graph multi-attention network, VTP-STGMAN. First, we revise the calculation of distance to the closest point of approach (DCPA), and navigational metrics, including the revised DCPA, time to the closest point of approach (TCPA) and actual distances among vessels, are applied to redesign the calculation method of the attention score in the graph attention network (GAT). The improved GAT model can extract the spatial features of multiple vessel trajectories accurately. Secondly, we integrate an attention mechanism with the temporal convolutional network (TCN) to adaptively capture nonlinear temporal correlation of trajectories. Finally, the probability distribution of the future trajectory rather than a deterministic trajectory of a vessel is predicted. It can reflect the inherent uncertainty of ship motion in the real world more accurately, and allow the collision avoidance decision system to minimize the risk expectation based on multiple high probability trajectories. Numerous experiments are conducted on real-world automatic identification system (AIS) data from Zhoushan (China) and Denmark. Evaluated results demonstrate that VTP-STGMAN can significantly outperform state-of-the-art baselines, reducing average displacement error (ADE) and final displacement error (FDE) by 23.6% and 29.7% in the primary benchmark respectively.</p>