Discrete reactive power optimization considering safety margin by dimensional Q-learning

This paper focuses on implementing a dimensional Q-learning (DQL) for solving reactive power optimization with discrete control variables. The proposed algorithm applies the traditional Q-learning to search the feasible region dimensionally, so that the memory amount of each agent can be largely reduced. Meanwhile, the safety margin of voltage amplitude and reactive power output of generators are also considered in the optimization, and the objective function merely includes the power loss without penalty terms. According to the experiment studies in this paper, DQL is able to optimize the reactive power dispatch and safety margin with advantage over other two popular algorithms.