The current deep reinforcement learning (DRL) method has some issues with voltage optimization, such as challenging credit allocation and low exploration efficiency. These all lead to poor performance in model training speed and optimization effect. Considering regionalization and imitation learning, a voltage optimization strategy based on the guidance signal-based multi-agent deep deterministic policy gradient (GS-MADDPG) is proposed. First, electric vehicle (EV) clusters, distributed generation (DG) and reactive power regulators are taken as decision agents to build the reinforcement learning optimization model. Secondly, the external reward is decoupled through regionalization of the distribution network, and combined with imitation learning, an internal reward is introduced through the guidance signal to help agents search for optimization quickly. Finally, an example test is conducted on the improved IEEE 33-node distribution network. The results indicate that the proposed voltage optimization strategy has higher sample utilization, more stable convergence, and higher model training efficiency than the traditional DRL method, and improves the voltage optimization effect.