The hydrodynamic modeling of a surface vehicle in an aquatic environment is known to be a challenging problem. In particular, it is difficult to calculate the hydrodynamic forces due to the intricate coupling of the vehicle with water. In recent years, deep dynamics models have been utilized to improve the accuracy of dynamics modeling precision. However, it is well known that data-driven approaches do not extrapolate well and are vulnerable to out-of-distribution data. In this paper, we argue that the naive use of neural networks may reduce the accuracy of long-term predictions. In order to address this issue, we employ representation learning techniques that facilitate the learning of the valid data space, so that long-term predictions can be made within the valid data space. In addition, hallucinated replay is incorporated into the prediction network to further improve the accuracy of long-term predictions. We validate the proposed method on experimental data using a robotic surface vehicle and demonstrate its application to path tracking control.