Differentiable Predictive Control of Permanent Magnet Synchronous Motors

Recently, there has been an exploration of neural network-based approximate model predictive control (AMPC) for achieving real-time optimal control of permanent magnet synchronous motors (PMSMs). However, these methods suffer from extensive offline sampling of the state and parameter spaces, which limits their use for such system. To address this challenge, this paper presents a differentiable predictive control (DPC) approach for the torque control of PMSMs and shows its reduced sampling effort while maintaining accuracy and constraint adherence in comparison to an AMPC. In particular, a neural network policy is trained using gradient descent by minimization of a loss function and differentiating through closed-loop system trajectories. The paper discusses the efficient integration of the voltage and current constraints into the DPC closed-loop system and the policy training process. The comparison between the policy learned using DPC and a state- of-the-art AMPC in test bench evaluations shows that DPC achieves comparable performance without demonstrations from a supervisory controller, eliminating the need of a separate datageneration step.