Q-Learning-based Finite Control Set Model Predictive Control for LCL-Coupled Inverters with Deviated Parameters

Abstract

Finite Control Set (FCS) Model Predictive Control (MPC), as an efficient method used for current tracking of LCL-Coupled three-phase inverters, runs into high computational complexity while finding its optimal version with a long predictive interval. For such a difficult problem we take a value function with discounted factors as an indicator to measure the pros and cons of control and propose a novel alternative method based on Q-learning algorithm. In the control scheme, the value function is approximated by reinforcement learning(RL) algorithm and furthermore, the long horizons prediction is transformed into an iterative multi-step matrix calculation. At the same time, the optimal switching position is directly obtained without a modulation link, which greatly reduces the computational complexity. Accordingly, a data-driven Q-learning algorithm is designed with a proof of convergence. Last, the proposed algorithm's performance in the case of complete deviation from the (unknown) system parameters is verified by simulations.