Active power security correction strategy for power grids based on spatiotemporal graph deep reinforcement learning

With the growing penetration of renewable energy sources, modern power systems are increasingly challenged in maintaining active power security. Conventional optimization-based correction methods often lack real-time adaptability and robustness, particularly under the uncertainty and volatility of renewable generation. Deep reinforcement learning (DRL) has emerged as a promising alternative, however, most existing DRL approaches overlook the graph-structured characteristics of power networks and the temporal dynamics induced by renewable fluctuations, which limits their correction performance. To address these challenges, this paper proposes a spatiotemporal graph-based DRL framework that integrates graph convolutional networks (GCNs) for topological feature representation and temporal convolutional networks (TCNs) for capturing dynamic system evolution. The correction problem is formulated as a Markov decision process (MDP), and the agent is trained using the soft actor-critic (SAC) algorithm, which enhances robustness through entropy-regularized exploration. By jointly learning spatial and temporal patterns, the proposed method improves the agent’s situational awareness and decision-making capability under uncertainty. Experiments on a modified IEEE 118-bus system demonstrate that the proposed approach achieves more effective corrective actions, faster convergence, and higher renewable utilization compared to state-of-the-art baselines. • Formulates active power security correction as an MDP with spatiotemporal states. • Proposes a spatiotemporal DRL agent using GCN for topology and TCN for dynamics. • Develops a GCN–TCN-enhanced SAC framework with improved robustness and convergence. • Demonstrates superior performance on the IEEE 118-bus system over DRL baselines.