The increasing volume of Automatic Identification System (AIS) data generated by maritime vessels poses significant challenges in data storage, transmission, and real-time processing, particularly in bandwidth-constrained environments. Traditional trajectory compression methods often fail to preserve safety-critical information, which is essential for collision avoidance and maritime situational awareness. This study proposes a Risk-Aware Deep Learning method that integrates sequence-to-sequence Long Short-Term Memory (LSTM) models with attention mechanisms and a domain-informed risk assessment framework to compress AIS trajectories efficiently. By assigning dynamic risk scores based on proximity to other vessels, traffic density, navigational hazards, and vessel manoeuvres, the model prioritises the preservation of high-risk trajectory segments. Experimental results demonstrate that the proposed method outperforms traditional geometric, spatiotemporal, and autoencoder-based approaches in terms of compression ratio, reconstruction fidelity, and safety feature retention. With a risk preservation score of 95% and a compression ratio of 7.5, this model provides an effective solution for maritime data management and supports real-time monitoring, predictive analytics, and autonomous navigation. Future work will explore real-time deployment, federated learning, and the integration of multi-modal maritime data sources.