Intelligent Reactive Voltage Optimization Method for Distribution Networks Based on Value Function Decomposition and Graph Attention

With the large-scale integration of photovoltaic, the issue of frequent voltage fluctuations in distribution networks due to its variability and intermittency has become increasingly prominent. Traditional model-based reactive voltage optimization methods rely heavily on precise grid modeling, making it challenging to achieve the required solution accuracy and computation speed in distribution networks that include photovoltaic energy. However, existing deep reinforcement learning methods exhibit limited feature extraction capabilities and suboptimal control performance when dealing with complex distribution network scenarios. To address this issue, this paper proposes a multi-agent deep reinforcement learning control strategy that incorporates value function decomposition and graph attention mechanisms. First, the reactive voltage control problem is modeled as a distributed partially observable Markov decision process. Then, a graph attention network is introduced into the actor network of MADDPG. For credit assignment in the critic network, a mixing network is designed based on value function decomposition. Finally, case studies are conducted on an improved IEEE 33-bus distribution system. Experimental results demonstrate that the proposed control strategy effectively extracts state features and memorizes network topology information, exhibiting excellent voltage stabilization and loss reduction performance, as well as robust robustness and training stability.