Differentiable Predictive Current Control of Permanent Magnet Synchronous Motors

Control systems with non-linear characteristics, such as highly-utilized permanent magnet synchronous motors (PMSMs), presents significant challenges. Extensive system knowledge is necessary to design strategies that handle these complexities effectively. The differentiable predictive control (DPC) framework offers a promising solution by leveraging advances in machine learning to learn a neural network-based controller in a data-driven manner. This approach directly considers the nonlinear dynamics of the system by defining an end-to-end differentiable optimal control problem. In this work, DPC is applied to the current control of a PMSM within a numerical simulation. The performance of this learned controller is compared with both an optimal open-loop controller with unlimited foresight and a classical field-oriented controller to evaluate the feasibility of DPC in the electrical drive domain.