Virtual Power Plant Reactive Power Voltage Support Strategy Based on Deep Reinforcement Learning

After the large-scale access of distributed power sources to the distribution network, significant high/low voltage problems have emerged. Using a virtual power plant to provide reactive power voltage regulation as an ancillary service effectively addresses voltage issues. However, since a third party manages the virtual power plant and contains both discrete and continuous regulation devices internally, there is a need to consider privacy protection. To address this, a training method that requires minimal boundary information and reward–penalty information for interaction between discrete and continuous action agents is proposed. This method uses distributed two-layer multi-agent deep reinforcement learning for the virtual power plant’s reactive power voltage support strategy. By utilizing actual engineering data and comparing it with the “centralized training” framework algorithm, this study proves the effectiveness of the deep reinforcement learning training method and reactive power voltage control strategy. It demonstrates advantages such as protecting the privacy of the virtual power plant and low training difficulty.