Controller Parameterization for Grid-Connected Power Converters through Reinforcement Learning

Inhalt:
Due to the increasing feed-in of renewable energies (RE) in the power grid, the number of power converters is increasing. This can lead to the destabilization of the system due to various factors, one of which is the control structure of the power converter and its control parameters. Therefore, parameterization and optimization of the controllers in the power converter are necessary and required. This paper addresses how the stability of a converter-dominated network in different states can be improved by parameterization using reinforcement learning (RL). To perform the parameterization, the adapted RL agent and hyperparameters are to be set according to the target. In this work, two parameterization methods are used. In the first method, an RL agent is structured with an artificial neural network so that it can behave like a PI controller. This optimizes the parameters of PI controllers immediately during training. In the second method, the RL agent can adaptively output the parameters for PI controllers according to different operating points of the power converter, which are the input of the RL agent. Root Mean Square Error (RMSE) is used to evaluate the deviation of power injection from the set point. As a result, the proposed parametrization method can improve the power system stability.