Assessment of Deep Reinforcement Learning Algorithms for Three-Phase Inverter Control

• Main Authors: Menendez, Oswaldo, Lepez-Caiza, Diana, Tarisciotti, Luca, Ruiz, Felipe, Auat-Cheein, Fernando, Rodriguez, Jose
• Format: Conference Proceeding
• Language: English
• Subjects: artificial intelligence; Computational efficiency; Deep learning; deep reinforcement learning; Frequency control; Heuristic algorithms; Inverters; Neural network; non-linear control; power converter; Reinforcement learning; Uncertainty; voltage source inverter
• ISSN: 2832-2983
• DOI: 10.1109/SPEC56436.2023.10407331

Deep reinforcement learning (DRL) offers outstanding algorithms to develop optimal controllers for power converters with uncertainties and non-linear dynamics. This work comprehensively analyses a model-free control algorithm for three-phase inverters using DRL agents. To this end, different deep deterministic policy gradient (DDPG) agents with variable hyperparameters were conceptualized, designed, and tested. On average, DDPG agents were shown to have excellent performance in the control of power inverters. Indeed, DDPG agents reduce the impact of model uncertainties and non-linear dynamics. To validate the proposed control policy, the two-level voltage source power inverter is simulated. Also, the main features of the control strategy are analyzed in terms of computational cost, root medium square error (RMSE), and total harmonic distortion (THD). Simulated results reveal that the proposed control strategy exhibits strong performance in the current control task, achieving a maximum RMSE of 0.78 A and a THD of 3.17% for a 6 kHz sampling frequency.