Data augmented offline deep reinforcement learning for stochastic dynamic power dispatch

• An offline deep reinforcement learning framework is proposed for stochastic dynamic power dispatch. • A CGAN-based method augments rare scenarios to enhance data diversity and learning robustness. • A COSAC algorithm integrates conservative Q-learning with SAC for safe offline policy learning. • The proposed method avoids unsafe online interactions while ensuring high dispatch performance. • Extensive experiments show superior reliability and generalization over existing DRL approaches. Operating a power system under uncertainty while ensuring both economic efficiency and system security can be formulated as a stochastic dynamic economic dispatch (DED) problem. Deep reinforcement learning (DRL) offers a promising solution by learning dispatch policies through extensive system interaction and trial-and-error. However, the effectiveness of DRL is constrained by two key limitations: the high cost of real-time system interactions and the limited diversity of historical scenarios. To address these challenges, this paper proposes an offline deep reinforcement learning (ODRL) framework tailored for power system dispatch. First, a conditional generative adversarial network (CGAN) is employed to augment historical scenarios, thereby improving data diversity. The resulting training dataset combines both real and synthetically generated scenarios. Second, a conservative offline soft actor-critic (COSAC) algorithm is developed to learn dispatch policies directly from this hybrid offline dataset, eliminating the need for online interaction. Experimental results demonstrate that the proposed approach significantly outperforms both conventional DRL and existing offline learning methods in terms of reliability and economic performance.