Improved performance of a PMSM control system based on a high-order SMC, MPC, and reinforcement learning agent for rapid control prototyping

Because controlling a permanent magnet synchronous motor (PMSM) is difficult due to its nonlinear nature and the occurrence of electrical and mechanical disturbances, this article proposes a global control system that uses the classic field-oriented control (FOC) structure. Thus, a Model Predictive Control (MPC) type controller is used for the inner control loop, while a High-Order Sliding Mode Control (HO-SMC) structure is used for the outer control loop. Implementing the proposed control structure improved quality indicators, steady-state error, and speed ripple. These were all affected by chattering in the SMC controller structure. HO-SMC controllers are inherently structured with Levant-type differentiators. To reduce execution time without compromising control performance, these controllers use feed-forward neural networks (NNs) to calculate the derivatives of the sliding surfaces. Observers are implemented for electrical disturbances of the Sliding Mode Observer Electrical Disturbance (SMO-ED) type and mechanical disturbances of the Sliding Mode Observer Mechanical Disturbance (SMO-MD) type. These observers eliminate the disturbances through feed-forward loops. To improve control performance, a reinforcement learning (RL) twin-delayed deep deterministic (TD3) agent is used. This agent works in tandem with the controllers in the outer control loop to provide correction signals for the inner control loop reference. Due to the large number of tuning parameters for controllers and observers, the Remora Optimization Algorithm (ROA), a computational intelligence meta-heuristic algorithm, is used. The following indicators are also used for the comparative presentation of control system performance: response time, overshooting, Fractal Dimension (FD) of the rotor speed signal, and Total Harmonics Distortion (THD) of the PMSM stator currents in one phase. We presented the parametric robustness of the controllers using a Monte Carlo procedure. Numerical simulations are performed in Matlab/Simulink to validate the controller design during the Software-In-the-Loop (SIL) stage. One representative task is the real-time comparative presentation of controller performance in the Hardware-In-the-Loop (HIL) Rapid Control Prototyping (RCP) stage. This stage provides information on the equivalent implementation of execution speed in low/medium performance standalone Digital Signal Processors (DSPs) with a sampling time of T s = 5×10 -5 s and high performance DSPs with a sampling time of T s = 10 -5 s, respectively. The SpeedGoat Performance Real-Time Target Machine is the hardware testing platform in the HIL-RCP stage. The performances of the proposed global control system are compared with the performances obtained by other control structures.