The high penetration of distributed energy resources has significantly increased the risk of voltage limit violations in distribution networks. Voltage optimization control methods based on deep reinforcement learning (DRL) struggle to ensure constraint satisfaction during online execution and lack adaptability to network topology changes. Additionally, model predictive control (MPC) incurs high computational costs due to full-horizon optimization. To address these challenges, an optimal voltage control method for distribution networks using deep reinforcement learning, considering dynamic correction by a policy safety layer is proposed. First, the spatial coupling features of the distribution network topology extracted by the graph attention network (GAT) are integrated into the state space of the deep deterministic policy gradient (DDPG) algorithm for offline training, thereby enhancing adaptability to frequent topology changes. Then, MPC is embedded into the DDPG framework as a policy safety layer. A value network is introduced to approximate the terminal value term, replacing the long-horizon terminal cost optimization in MPC, while the problem is then transformed into a control sequence sampling distribution using the cross-entropy method (CEM). Finally, the proposed method is validated based on an improved IEEE 123-bus system. Simulation results show that the proposed method not only ensures control performance and constraint feasibility but also exhibits excellent optimization control effects and robustness under uncertainties, topology changes, and measurement inaccuracies.