With the growing integration of distributed renewable energy sources into the distribution network, the participation of distributed grid-connected inverters in volt-VAR control (VVC) is of great significance to improve the safety and economy of power system operation. However, the safety operation constraints are difficult to model in the VVC control model based on reinforcement learning, and it is not possible to ensure that the control strategy satisfies the operation constraints. To address the above problems, this paper proposes a voltage control strategy for distribution networks based on the safe reinforcement learning. Firstly, the voltage control problem with constraints is modeled as a constrained Markov decision process. Then, a primal-dual approach is used to learn the optimization policy to ensure that the control policy satisfies the system operation constraints. Furthermore, an enhanced experience fusion method is introduced to improve the utilization of reinforcement learning experience, so as to improve the sample efficiency of the algorithm. Finally, the effectiveness of the proposed method is verified by the distribution system example.