With the access of distributed power and random load, the problem of voltage fluctuation in distribution network becomes more and more serious. Active distribution network can suppress voltage fluctuation through various voltage and reactive power controllers, but it is usually necessary to solve a complex mixed integer second-order cone programming problem, which is difficult to achieve real-time control. In this paper, a real-time voltage control model of active distribution network is established by using deep reinforcement learning, which can quickly get the control strategy satisfying the power flow constraints. It collects node active power, node reactive power, transformer voltage regulating gear and time step as environmental state variables. It coordinates the three control equipments with the cost related to network loss and equipment operation as return function. It solves the problem through constraint-based reinforcement learning based on long short-term memory network, so as to establish a real-time voltage control model of active distribution network. Based on the 4-node test system and IEEE 33-node test system, the simulation results show that the proposed deep reinforcement learning method can ensure the power flow constraints, and the voltage control model can control the voltage and reactive power controller in real time to ensure the voltage quality of the distribution network.