Abstract: In view of the problem that the high proportion of distributed power supply access leads to the intensification of voltage fluctuation and the decline of power quality in distribution network, the research on the control of output voltage fluctuation of distribution network based on deep reinforcement learning is carried out. A multi-objective function integrating static voltage deviation and time-series voltage fluctuation is established, and the flexible multi-state switch and system operation constraints are fully considered to construct the voltage control model of distribution network. A deep reinforcement learning framework is designed to decouple the sequential decision-making problem into single-point decision-making. By training Actor network and Critic network in sequence, the end-to-end rapid mapping from state to FMSS optimal control instruction is realized. The experimental results show that this method can effectively suppress the voltage fluctuation of distribution network, improve the stability of system operation and power quality, and has high online calculation efficiency and strong engineering practical value.