This paper presents a novel multi-task deep learning framework for maritime vessel behavior prediction in port traffic management systems. The proposed approach integrates LSTM and Transformer architectures [1, 2] with feature fusion to simultaneously predict vessel positions, speeds, heading angles, and operational intentions. A systematic evaluation of temporal contexts reveals that 7-day historical windows significantly outperform shorter alternatives, achieving 53.8% MSE reduction compared to 1-day windows. Experimental validation on 2,252,088 AIS trajectory records demonstrates exceptional performance: \(\text {R}^2\) values exceeding 0.91 for regression tasks and 96.41% accuracy for intention classification. The 7-day configuration captures complete operational cycles and extended behavioral patterns that shorter windows miss. Multi-task learning enables effective knowledge sharing across prediction objectives, with the hybrid architecture achieving superior performance through complementary temporal dependency modeling. The framework establishes temporal context extension as a fundamental principle for maritime prediction systems, providing scientific foundation for next-generation intelligent port traffic management.