An Improved Reinforcement Learning Approach for Cost Function Optimization in Multiobjective FCS-MPC of PMSM Drives

Model predictive control has gained increased popularity in widespread applications using electrical drive systems, due to its merits of the customized design of the cost function and removal of the modulation stage. However, the weighting parameters in the cost function are hard to fine-tune for modification of various control objectives. More specifically, as more conflicting objectives are involved in the reformulated cost function, optimization of a combination of weighting parameters has become a time and resource consuming task. To cope with this issue, an improved reinforcement learning approach is proposed for cost function optimization in the multiobjective finite control set model predictive control (FCS-MPC) scheme. First, a customized cost function is designed to involve multiple conflicting targets as well as the weighting parameters to be optimized. Based on the above, the improved reinforcement learning framework is employed for weighting parameters tuning without prior knowledge, to obtain a balanced performance metric of torque, flux, and meanwhile optimize switching sequences. Furthermore, a twin delayed deep deterministic policy gradient (TD3) algorithm is proposed to mitigate the instability and performance degradation inherent in the conventional deep deterministic policy gradient (DDPG) algorithm. By incorporating delayed target updates and a more robust exploration strategy, TD3 effectively addresses key issues such as overestimation bias, thereby improving learning stability. The proposed method is experimentally conducted and comprehensively compared with the existing FCS-MPC methods on a lab-built 4.8 kW permanent magnet synchronous machines drive, which effectively addresses the challenge of parameters optimization in FCS-MPC schemes.