Optimal Control of Interior Permanent Magnet Synchronous Motor Leveraging Differentiable Predictive Control Based on Deep Learning

This paper proposes a novel approach to substitute the outer-loop proportional-integral controller of an interior permanent magnet synchronous motor (IPMSM) with a predictive neural network-based alternative. The proposed methodology involves developing a Simulink IPMSM control scheme with a maximum torque per ampere (MTPA) algorithm and PI controllers. It employs measured data from the angular velocity control system to train neural ordinary differential equations and constrained differentiable predictive control models with the goal of substituting the classical controller using proportional inputs and outperforming classical current reference generation, thus optimizing the integrated MTPA algorithm. The neural network is then deployed into the Simulink model with an additional Kalman-filter-based compensator term to produce a control signal that prioritizes tracking accuracy with minimal overshoot or violation of the machine's operational constraints.