A novel online reinforcement learning-based linear quadratic regulator for three-level neutral-point clamped DC/AC inverter

This article proposes a novel online reinforcement learning-based linear quadratic regulator for the three-level neutral-point clamped DC/AC voltage source inverter. The proposed controller employs online updated fixed-weight recurrent neural network (NN) and policy iteration to dynamically adjust the optimal control gains based on real-time measurements without any knowledge of the system model or offline pre-training. Moreover, it produces a constant switching frequency with low current harmonics. Compared to the existing control methods, it provides superior control performance, guaranteed control stability, and simplified NN design. Experimental results are presented to verify the effectiveness of the proposed control method.