Unmanned Surface Vehicles (USVs) are widely used in marine exploration, security, and autonomous navigation, where trajectory prediction plays a critical role in decision-making. This study proposes a prediction framework that combines a high-fidelity ship dynamics model with a Soft Actor-Critic (SAC) algorithm to optimize a Long Short-Term Memory (LSTM) network. The dynamics model includes first-order response delay and stochastic wind-wave disturbances to simulate realistic data. SAC adaptively tunes key LSTM hyperparameters, improving optimization efficiency and enhancing generalization under complex sea conditions. Experimental results show that the SAC-optimized LSTM outperforms manually tuned models, highlighting the potential of integrating deep reinforcement learning with sequence modeling for USV navigation.