Deep reinforcement learning-based algorithm applied to drive control

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Alternating current machines are widely employed in industrial applications, and their efficiency is highly dependent on applied control schemes. Traditional control methods, such as the linear proportional–integral (PI) controller, are widely adopted in industry because of their simplicity. Their effectiveness hinges on the accuracy of the plant model, and can decrease when system parameters change (a phenomenon known as parametric perturbations). In this study, to overcome the dependency of the plant model on the control performance, a deep reinforcement learning (DRL) method is proposed to achieve a dynamic speed control performance. DRL is a unique technique that enables an agent to learn the drive control from scratch by only interacting with appropriate control state variables. The main objective of this study is to design an intelligent speed controller that overcomes the effect of external disturbances or parametric perturbations, i.e., external load torques or electrical or mechanical parameters of the motor drive. An actor–critic (AC) network is implemented, which assists in achieving the combined advantages of value-based and direct policy mapping. The policy-based mapping (actor) gives the probability of the action to be taken based on state observations. In contrast, the value-based mapping (critic) gives the prediction of the higher action based on the function value. Both networks tune themselves to more effectively evaluate and select actions in the future. The DRL agent is trained using a deep deterministic policy gradient (DDPG) to address the speed control of a surface-mounted permanent magnet synchronous motor (SPMSM) drive. The performance of the optimized PI speed controller is compared to that of the DRL method, and the test results are presented to demonstrate the effectiveness of the proposed method.