With a large number of distributed photovoltaic（PV） access to the distribution network, the distribution network faces greater challenges in dealing with network reconstruction and source-load-storage uncertainty, etc. Therefore, a two-stage voltage control strategy for active distribution networks is proposed. In the first stage, the contact switch of the active distribution network is centrally controlled. The network reconstruction is carried out with the goal of minimizing the network loss in an hourly scheduling period, and a mixed integer second-order cone planning model is established to solve the problem. In the second stage, the real-time voltage control of photovoltaic and energy storage systems is carried out, and the real-time voltage control problem is converted into the Markov game process（MGP）. The multi-agent model is implemented and the offline training-online operation method is adopted. Compared with the traditional two-stage mathematical planning approach, the control strategy proposed does not rely on an accurate distribution network power flow model, has low communication requirements and a faster solution speed. Finally, the effectiveness of the proposed control strategy is verified by the improved IEEE 33-node calculation examples.