Gradient Descent Algorithm Based on Extended State Observer for Loss Minimization of Ultralocal Model Predictive Controlled PMSM Drives

Hybrid loss minimization algorithms in electrical drives combine the benefits of search-based and model-based approaches to deliver fast and robust dynamic responses. This article presents a novel hybrid loss minimization method that replaces the conventional physical model with a mechanical ultralocal model integrated with a gradient descent (GD) algorithm. By substituting explicit computations of individual power losses with a straightforward extended state observer (ESO) for load torque estimation, the proposed approach simplifies and enhances the precision of efficiency calculations. The estimated efficiency is then incorporated into the GD algorithm to determine the optimum angle for the reference current vector. Moreover, the method offers an additional advantage by employing the same ESO for robust speed control. The obtained ESO-based GD algorithm is implemented in an ultralocal model predictive control (UMPC) approach for a permanent magnet synchronous motor (PMSM). Therefore, the reliance of the overall control system on conventional mathematical models is eliminated. Experimental evaluations, compared to a UMPC with a maximum torque per ampere (MTPA) control approach, demonstrate significant improvements in efficiency and dynamic response for the proposed UMPC.