Reinforcement Learning-assisted Controller Parameterization based on Improving Oscillatory Stability in Converter-Dominant Power System

Due to the cascade control loop in the single converter and interactions among converters, the converter-based system is more likely to become unstable. The general controller parameterization methods for a converter are based on its parameters and states. Considering the possible oscillation effect from other converters nearby, a new procedure of controller parameterization with the help of reinforcement learning (RL) is then proposed. By setting an oscillatory stability indicator as a reward during the training of an RL-agent, a set of controller parameters are calculated and applied in the converter-dominant power system to improve the system oscillatory stability. To this end, an indicator that can represent the stability margin should be first selected. Afterwards, a mathematical relationship is constructed between the controller parameters, the stability indicator, and the current observed states. Experimental results are presented and compared with the standard analytical parameterization method.