Robust Voltage-Reactive Power Support using GAN-DRL Architecture for Grids with IBRs

This paper proposes the integration of Deep Reinforcement Learning (DRL) and Generative Adversarial Neural Network (GAN) architecture for robust voltage-reactive power scheduling support for transmission grids with aggregated Inverter Based Resources (IBRs). Grid operators endeavor to keep the voltage profile consistent during normal system operations and in the event of emergencies. In many cases, they use modelfree approaches in the form of machine learning-based models to provide network situational awareness and decision support systems such as voltage-reactive power (Q-V) grid service. Nevertheless, these model-free methods rely on the input system state. As a result, they are susceptible to becoming inaccurate and ineffective when faced with perturbations in their input state, such as incorrect measurements through phasor measurement devices or improper data transmission not noticeable. Therefore, a generative adversarial architecture for detecting and responding to perturbations is proposed as a robust approach for supporting Q-V scheduling decisions made by system operators. The architecture is built with an additional layer for robustness, containing two artificial neural networks - the Generator and the Discriminator- and is trained and verified on the modified IEEE 30-bus transmission system with aggregated IBRs. The results show that the proposed model provides robust support against perturbed system states for voltage stability during grid operation.