Abstract: The traditional volt/var control（VVC） method in distribution networks is difficult to balance the global optimality and real-time response capability of control decision-making. The decentralization of distributed photovoltaic（DPV） and high proportion of grid-connection results in increasingly prominent contradictions. Combining the optimum searching ability of model optimization and online decision-making efficiency of deep reinforcement learning, a combined model-and data-driven VVC strategy for distribution networks with photovoltaic（PV） clusters strategy is proposed. Firstly, with consideration of DPV cluster division, the framework of the distributed two-stage VVC is established based on the operation characteristics of the day-ahead optimal dispatching and intraday real-time control. Secondly, aiming at minimizing the operation losses of the system, the distributed dayahead VVC model of the distribution network is established, and the distributed solution algorithm based on Nesterov accelerated gradient is proposed. Then, setting the day-ahead decision as input, a real-time VVC model based on partially observable Markov game is established, and an improved multi-agent deep deterministic policy gradient algorithm based on iterative termination penalty function is proposed. Finally, the case analysis is carried out based on MATLAB/PyCharm software platform. The global optimality and real-time response capability of the proposed method is verified, and the economy and security of the operation of the distribution network with high proportion of PV are improved.