Transient voltage stability emergency control of power systems based on an integrated SAC-GAIL framework

The growing integration of renewable energy elevates the risk of transient voltage instability in power systems. Meanwhile, conventional pre-planned control lacks adaptability, and deep reinforcement learning (DRL) approaches suffer from low sample efficiency, high exploration risks, and insufficient utilization of expert knowledge. To address these challenges, this paper proposes a transient voltage stability emergency control method by integrating the soft Actor-Critic (SAC) algorithm with generative adversarial imitation learning (GAIL). First, an integrated SAC-GAIL framework is established, which incorporates a hybrid reward mechanism and a dual-buffer sampling strategy. This integration effectively balances environmental feedback with expert guidance, thus enhancing training stability and sample efficiency. Second, behavior cloning pre-training is adopted to safely initialize the policy network. Meanwhile, a dynamic expert data augmentation mechanism is introduced to continuously improve generalization through bootstrapped learning. Furthermore, an interactive platform is developed to integrate DRL with the PSD-BPA simulator, forming a SAC-GAIL-based decision-making architecture for voltage emergency control. Finally, tests on the CSEE-VS benchmark and a provincial power grid show that the proposed method achieves millisecond-level online decision speed and strong environmental adaptability while ensuring high control success rates.