The potential for big data in ship trajectory prediction to revolutionize maritime traffic flow analysis is considerable. By enabling real-time prediction of route conflicts, it can significantly enhance the safety of maritime transportation systems. However, existing methods face challenges in effectively handling sparse data and capturing long-term dependencies. This paper introduces HierFormer, a novel deep learning framework designed to support ship trajectory prediction in port areas using historical AIS data. It addresses the aforementioned challenges. HierFormer enhances data quality and strengthens spatiotemporal correlations through hierarchical encoding and dynamic feature matrices. It also incorporates an attention mechanism that improves the model’s capacity to capture long-term temporal relationships, and it uses a hybrid loss function to maximize prediction performance. Experimental results demonstrate that HierFormer outperforms baseline methods across various datasets, achieving an average error reduction of 40.93%. The approach considerably improves prediction performance in ship trajectory prediction tasks by successfully addressing long-term dependence and data sparsity concerns.