With the rapid development of artificial intelligence, applying machine learning methods in solving fluid-structure interaction problems is becoming an emerging trend. In this study, an integrated model using Computational Fluid Dynamics (CFD) and Deep Reinforcement Learning (DRL) is proposed for optimizing the control strategy of a cylindrical point absorber Wave Energy Converter (WEC). The model employs the Soft Actor-Critic (SAC) algorithm to construct a Markov Decision Process (MDP) between the WEC and the environment, and enables the DRL agent to dynamically adapt to the ocean stochasticity and systematic uncertainty by reconfiguring the state space, agent action and reward function. It is demonstrated that in the open sea environment, the DRL active control strategy is able to increase energy by 92.39% compared with the traditional resistive control. When migrating to a more complex environment with wave reflections, the energy output still can be maximized by an easy retraining, which is a 56.51% improvement over the open sea environment, and is significantly better than the resistive control scheme. The results verify the higher energy capture performance and strong migration capability of the model, and confirm that re-optimizing the parameters in a complex environment improves the energy capture efficiency while ensuring the system stability. Moreover, combining high-fidelity CFD modeling with DRL overcomes the limitations of traditional potential-based methods by simulating more realistic fluid-structure interactions, providing a new solution to the challenge of WEC control.