Abstract：The risk of voltage violations and fluctuations has surged due to the high penetration of distributed

renewable energy and the large-scale application of new power electronic devices. Therefore，it is essential to

optimize the control system of reactive voltage through intelligent methods for the reliability of smart distribution

networks. A multi-agent deep reinforcement learning algorithm based on prioritized experience replay（PER）and

multi-agent twin delayed deep deterministic policy gradient（MATD3）was proposed. First，the voltage control

problem based on photovoltaic inverters was modeled as a partitioned Markov decision process. Then，the PER

sampling method was used to optimize the reinforcement learning sampling operation based on TD error. Finally，

the MATD3 policy was used to solve the model. The reactive voltage was optimized by means of photovoltaic

inverter. Optimization validation was conducted using open-source photovoltaic generation and load data. The

voltage optimization experiments were carried out using the data from typical days. Compared with the traditional

MATD3 and MADDPG algorithms，the average voltage deviation decreases by 4.7% and 15.5%，respectively.

Network losses are reduced by 1.6% compared to the traditional MATD3 model. These results validate the

effectiveness and economic benefits of the proposed method.

        Key words：smart distribution grid；PV inverter；reactive voltage optimization；reinforcement learning

     Format Citation:周建华，张睿凝，陶锴,等. 基于强化学习PER⁃MATD3模型的智能配电网无功电压优化控制[J].电气传动,2026,56(07):47-55. ZHOU Jianhua，ZHANG Ruining，TAO Kai,et al. Reactive voltage optimization and control for smart distribution networks based on reinforcement learning PER-MATD3 model [J].Electric Drive, 2026,56(07):47-55