Abstract
To protect power systems against different severe disturbances, the ways to effectively control voltage have become the important issue in power systems. In light of that, a Multi-Agent System (MAS) structure has been proposed to deal with the issue of voltage regulation. With our system, once the voltage violations occur, the agents would detect the abnormality and try to eliminate these voltage violations by injecting the reactive power. Besides, to make a good decision under abnormal conditions, a reinforcement learning scheme has been proposed to provide better and faster regulation. Based on the concept of distributed control, there are two parts in our reinforcement learning scheme, self-regulated learning and cooperative learning. On the other hand, an altruistic rate has been proposed to consider the impact of other neighbor agents. The performance of the proposed multi-agent reinforcement learning are demonstrated using various conditions in a benchmark of power networks.