Reinforcement Learning based Weight-Tuning Model Predictive Control of Permanent Magnet Synchronous Motor

Compared to traditional PI control methods, finite control set model predictive control exhibits enhanced applicability in managing complex and multi-variable systems, particularly when addressing constraints and nonlinear dynamics. However, a significant challenge lies in the determination of the weights for the cost function, typically relying on empirical approaches due to the absence of a robust theoretical foundation. To address this issue, this paper proposes a novel approach for both offline and online self-tuning of the cost function weights based on reinforcement learning. In the offline phase, the optimal weight is derived through controlled trial-and-error learning of permanent magnet synchronous motor. In the online phase, the weight of the model predictive control cost function is refined using the Q-learning algorithm. By undertaking two pivotal steps involving integer optimization and fractional data selection, which provide precise fractional information, the number of learning iterations is significantly reduced. Ultimately, this process culminates in the determination of the requisite optimal weights.