This paper presents a novel framework to maintain the stability of smart grids. While conventional smart grid stability assessment faces the difficulty of securing real unstable data, this study generates Out-Of-Distribution (OOD) samples representing unstable states using only stable data through Generative Adversarial Network (GAN). Using the generated OOD samples, we learn robust decision boundaries that distinguish between stable and unstable states, and additionally enhance resilience against attacks through adversarial training. The evaluation results using real datasets show that we achieve up to 98.1% stability prediction accuracy and 98.9% adversarial attack detection accuracy, and that real-time decision making is possible with an average response time of less than 7 ms on a single-board computer.
