Regresar

Lightweight Neural Network Architectures for Robust Data-driven Control System of Three-Phase Voltage Source Inverters

Abstract:

Data-driven control systems based on Deep Reinforcement Learning (DRL) agents are emerging as a promising alternative to traditional control approaches in power converter applications because of the enhanced robustness under uncertainties, external disturbances, and system nonlinearities. However, the hyperparameters of required neural networks- such as the number of layers, neurons, and activation functions- are still selected through empirical tuning, resulting in suboptimal performance and limited generalization. This work presents a data-driven control system for current tracking problem in three-phase voltage source inverter (VSI) using a simple perceptron as actor policy within a Reinforcement Learning framework. The control policy is trained using Soft Actor Critic (SAC) algorithm in order to generate continuous duty cycle using a simple state vector. Two neural network (NN) architectures are evaluated: a simple perceptron, and a simple perceptron with extended state inputs, including real-time current measurements. In addition, the performance of the data-driven control system is analyzed in terms of root mean square error (RMSE) and total harmonic distortion (THD). Results disclose that a simple perceptron is enough to control the VSI for the specific current tracking problem. Under this configuration, the proposed controller achieves a maximum RMSE of 1.20 A and a THD of 6.28% at a 5 kHz sampling frequency. The results validate the applicability of compact, data-driven control architectures for power electronic converters, offering a balance between computational efficiency and control performance.