With the rapid development of distributed photovoltaic (PV) technology in China, the issues of voltage limit violation and fluctuation caused by the large-scale penetration of distributed PVs into the distribution network have become increasingly prominent. Traditional voltage control methods are limited by solution speed and quality. While conventional deep reinforcement learning methods lack the necessary constraints during action exploration to ensure safety. Therefore, an active voltage control method of distributed PV based on safe deep reinforcement learning is proposed. Firstly, the voltage control problem is modeled as a distributed partially observable Markov decision process. Secondly, a safety layer based on the integration of “mechanism (mathematical computation) and knowledge (gird dispatch rules, etc.)” is added for agent design. Test results on the modified IEEE 33-bus case show that the proposed method can control the voltage within safe limits, significantly reduce voltage deviation and network losses, and improve the voltage safety of the power system.