Underwater continuous-variable quantum key distribution (CVQKD) plays an important role in a widespread CVQKD network due to its application scenarios. However, interferences in practical systems, such as equipment imperfections, channel disturbances, and eavesdroppers, may cause invalid communication. In this paper, we consider a risk caused by channel disturbance resulting in communication failure and suggest an ensemble learning (EL) approach, which involves a self-adaptive structure and makes the communication failure predictable, for the purpose of the security improvement of the CVQKD system with discrete modulation. Results show that the EL-based communication failure prediction could lead to the implementation feasibility of CVQKD in practice without the need for adding additional detection equipment at the receiver.