Deep Reinforcement Learning-Based Direct Torque Control of PMSM

28 de julho de 2026

This thesis investigates finite-control-set torque control for permanent magnet synchronous motors under parameter mismatch and multi-objective constraints. An end-to-end controller based on a deep Q-network is proposed to reduce dependence on accurate motor parameters and manual tuning. The torque control task is formulated as a Markov decision process. Normalized motor states and torque reference are used as observations, and the eight inverter switching states are used as actions. A composite reward function is designed for torque tracking, current safety, current utilization, torque ripple reduction, and switching smoothness. Sine and cosine electrical-angle features are introduced to represent periodic states. The controller is trained in a Python reinforcement learning environment and compared with conventional direct torque control in MATLAB/Simulink. Results show that the proposed controller achieves average torque tracking under low-, medium-, and high-torque steps and adapts current output to torque demand. Under nominal parameters, conventional direct torque control gives better torque tracking, lower torque ripple, and higher current utilization. Under stator resistance mismatch, conventional direct torque control degrades obviously and may cause overcurrent, whereas the proposed controller keeps stable operation without retraining, indicating zero-shot robustness. At the 0.8 p.u. high-torque boundary, the controller is still sensitive to the triggering state and lacks long-term stability. Torque reference smoothing extends high-torque holding time but does not replace further robustness optimization.

Autores da comunidade :

• 

  D

  Dongrui Cui

  ist1115681

Orientadores desta instituição:

• 

  A

  António José Castelo Branco Rodrigues

  ist12288

electrical-engineering-electronic-engineering-information-engineering - Engenharia Eletrotécnica, Eletrónica e Informática

por - Português

Instituto Superior Técnico