Utilizing Adversarial Training for Robust Voltage Control: An Adaptive Deep Reinforcement Learning Method

Researchers Sungjoo Chung and Ying Zhang have published a novel framework that applies adversarial training to deep reinforcement learning (DRL) for securing voltage control in modern power grids. The core innovation addresses a critical vulnerability: conventional control methods are susceptible to strategic cyber-attacks because they typically only account for random or 'black-box' disturbances. This new approach proactively formulates 'white-box' adversarial attacks using the Projected Gradient Descent (PGD) method and uses these optimized, worst-case perturbations to train the DRL agent. The result is a control policy that learns to be robust against highly intelligent, targeted attacks designed to destabilize the grid.

The method is specifically designed for distribution networks with a high penetration of distributed energy resources (DERs) like solar and wind, which introduce complex volatility. By training the AI agent against these strategically crafted adversarial examples, the system learns to adapt in real-time to maintain stability. Simulations on DER-rich network models demonstrate that the adversarially trained DRL agent successfully maintains voltage within safe limits and preserves operational efficiency even under simulated, realistic attack scenarios. This work, presented at the Texpas Power and Energy Conference 2026, highlights a significant step forward in using advanced AI security techniques to protect essential, real-world infrastructure from emerging digital threats.