With the development of power electronics technology, AC/DC hybrid distribution networks have gradually become an important form of distribution networks in the future and have attracted widespread attention. The high proportion access of distributed new energy represented by distributed photovoltaic (PV) brings great challenges to the safe and stable operation of AC/DC hybrid distribution networks, meanwhile the randomness and fluctuation of PVs’ power output put forward higher requirements for the real-time and dynamic response of the AC/DC hybrid distribution network optimal voltage control. Firstly, the Q-V voltage control model for PV in AC network and P-V voltage control model for PV in DC network are established. Then, the dynamic optimal voltage control model of AC/DC hybrid distribution network is constructed. After that, the data-driven and model-driven control strategies are combined, the voltage control controller is trained through the Multi-agent Gradient Descent Strategy (MADDPG) to solve the optimal adjustment action value of the PV droop parameter in the real-time state, so as to realize the online dynamic voltage control of AC/DC hybrid distribution network.