Optimal Weighting Factors Design for Model Predictive Current Controller for Enhanced Dynamic Performance of PMSM Employing Deep Reinforcement Learning

This paper presents a novel control strategy employing a deep reinforcement learning (DRL) scheme for online selection of optimal weighting factors in cost functions of the finite control set model predictive current controller of a permanent magnet synchronous motor (PMSM). Indeed, when designing predictive controllers for PMSMs’ phase currents, competing objectives appear, such as managing current convergence and switching transitions. These objectives result in an asymmetric cost function where they have to be balanced through weighting factors in order to enhance the inverter and motor performance. Leveraging the twin delayed deep deterministic policy gradient algorithm, the optimal weighting factor selection policy is obtained for online balancing of the choice between current deviation in the dq frame and inverter commutations. For comparison, a metaheuristic-based artificial neural network is trained on static data obtained through a multi-objective genetic algorithm to predict the weights. The key performance markers, such as torque ripple, total harmonic distortion, switching frequency, steady-state, and dynamic performance, are provided through numerical simulations to verify the effectiveness of the proposed tuning scheme. The results of these simulations confirm that the proposed dynamic control scheme effectively resolves the challenges of weighting factor choice, meeting inverter performance requirements, and delivering better dynamic and steady-state performance.