Permanent Magnet Synchronous Magnet Motor Intelligent Parameter Estimation Using Machine Learning Based Nonlinear Control

Modern industry has been witnessing the wide adoption of Permanent Magnet Synchronous Motors (PMSMs) for their highpower density and efficiency in applications such as transportation. However, it is well-known that nonlinearity and thesensitivity to variation of parameters in PMSM dynamics pose challenges to developing a highly efficient controller. Moreover,parameters’ variation caused by temperature rise, magnetic saturation, and aging of components is a problem in many controlproblems. In this paper, three different intelligent parameter estimation approaches that can be implemented alongside thecontrol strategies are suggested and compared: Recurrent Neural Network based Online Parameter Estimation for ModelPredictive Control (RNN-MPC), Reinforcement Learning-based Online Parameter Estimation for Model Predictive Control(RL-MPC), and Reinforcement Learning-based Online Parameter Estimation for Sliding Mode Control (RL-SMC). All of theabove frameworks suggest estimating the following PMSM parameters: stator resistance (Rs), d-axis inductance (Ld), q-axisinductance (Lq), permanent magnet flux linkage (ψf ), and rotor inertia (J) under various speed and load conditions. Specifically,the RNN estimator uses Long Short-Term Memory architecture trained offline using operational data, while reinforcementlearning estimators implement the Twin Delayed Deep Deterministic Policy Gradient algorithm to adapt parameters online. Thesimulation results show that all methods provide successful parameter estimation with an error level less than 3.5% for electricalparameters and 0.5% for mechanical ones. RNN based online parameter estimation for MPC demonstrates the smoothestparameter trajectories and minimum torque ripple (1.0 N·m peak-to-peak), whereas RL based online parameter estimation forMPC gives the highest transient performance with rise times equal to 0.12–0.18s and low overshoot. Boundedness of theestimation error and closed-loop system’s stability are proven using the Lyapunov-based approach. Based on the findings, it isclear that intelligent parameter estimation improves the performance of PMSM controllers. The RNN based online parameterestimation for MPC shows the best smoothness, RL-MPC shows the fastest adaptation and RL based online parameterestimation for SMC shows robustness but with oscillatory behavior.