A New Model-Free Space Vector Modulation Technique for Multilevel Inverters Based On Deep Reinforcement Learning

Multilevel inverters provide various advantages over conventional inverters. However, control and switching these types of inverters is a challenge for an increased number of levels that necessitate a higher number of semiconductor devices. Space Vector Modulation (SVM) switching method is a popular one that provides several advantages over conventional Sinusoidal Pulse Width Modulation techniques. Advantages such as higher output voltage, lower switching loss, lower dv/dt, etc. Despite its advantages, the SVM switching method requires complex time-interval equations. This complexity increases exponentially by increasing the number of levels. In this paper, a new SVM method based on Deep Reinforcement Learning (DRL) is proposed. The latter is based on a Reinforcement Learning Agent (RL-Agent) that tries all the "allowed" switching states of an inverter regarding the reference signal in each time step. Then based on the accuracy of the output, the agent receives a reward or punishment (negative reward). Over several episodes of trial and error, the agent learns how to perform the switching to receive the maximum long-term reward. Since the agent is also rewarded for keeping the voltage of the DC-link capacitors in a given range, this method is capable of voltage balancing. This method is implemented on a three-phase three-level Neutral Point Clamped (NPC) inverter for performance evaluation. The RL-agent is trained in MATLAB while the inverter model is simulated in the Simulink environment. The simulation results demonstrate the decent performance of this method.