Deep Reinforcement Learning Agent Based Speed Controller for DTC-SVM of PMSM Drive

Abstract

High-performance applications extensively use permanent magnet synchronous motor (PMSM) drives because of their high torque density and efficiency. However, conventional PI controllers employed in the outer speed control loop of direct torque control with space vector modulation (DTC-SVM) are limited by parameter sensitivity, poor adaptability under dynamic conditions, and the need for extensive manual tuning. To overcome these challenges, a Twin Delayed Deep Deterministic Policy Gradient (TD3) agent is introduced, incorporating a customised reward function to ensure precise torque reference generation. The TD3 agent is trained in MATLAB/Simulink using random speed and load profiles and deployed on a TMS320F28379D digital signal processor. Real-Time validation is carried out using an OPAL-RT 4512 simulator under a hardware-in-the-loop (HIL) framework. The inner-loop DTC operates at 20 kHz for torque and flux control, while the TD3 agent regulates speed at 2 kHz. Experimental results on 4.5 kW and 7.5 kW PMSMs show a 50% reduction in settling time, elimination of overshoot, and stable current responses without requiring controller retuning. The proposed method demonstrates robust and adaptive performance, confirming its effectiveness for embedded motor drive applications.