High propotional distributed photovoltaic integration changes the operation mode of the distribution networks, and leads to a series of problems such as excessive active power losses, reduced service life of regulating equipment, and exceeding node voltage limits in the distribution networks. Based on this background, firstly the voltage and reactive power optimization problem is modelled as a Markov decision process, which is solved by using a model-free deep reinforcement learning method that captures the intermittence of PV and load fluctuation from historical operating data. A graph convolutional network-proximal policy optimization (GCN-PPO) algorithm is proposed which improves the perception of reinforcement learning agent on graph data of distribution networks by embedding the graph convolutional network. Finally, an arithmetic analysis is carried out with a modified IEEE 33-node test system to verify the effectiveness of the proposed method and its advantages over other methods. The results show that the trained reinforcement learning agent based on graph convolutional networks exhibits better performance when the topology of the distribution network changes and the measurement data are lost.