基于事件驱动的强化学习预测控制器设计——用于三相NPC变流器的在线逼近器方法

This article is concerned with a two-step event-driven in reinforcement learning model-free predictive control problem leveraging online approximators for power converter systems, in which the limitations from system uncertainties and unnecessary switching loss are all addressed. To be specific, the key features of this technical note are: 1) a critic neural network to learn the performance function in real-time; 2) an actor neural network to approximate the predictive controller online and to minimize the learned performance function obtained from the critic network; 3) a two-step event-driven control protocol to attenuate the switching frequency (SF). Also, we further discuss the sensitivity of the proposal to parametric uncertainties and quantify its performance under low SF operation and unknown disturbances conditions. Further, the convergence analysis of the networks' weight estimation errors is manifested. Finally, we evaluate the suggested controller by means of various numerical examples, and the results found are promising and motivate further research in this field.