关键词: 深度强化学习, 电力系统调度, 动态状态表征, 分层奖励机制, 安全约束嵌入

Abstract:

The high proportion of renewable energy integration brings significant challenges of randomness, multi-objective coupling, and security constraints to power systems. Traditional model-driven methods have limitations in modeling accuracy and adaptability. To address these issues, this paper proposed a safety-constrained PPO algorithm (SC-PPO). The method included three improvements. A temporal convolutional network was utilized to construct a dynamic state encoder that integrated historical operation, real-time monitoring, and prediction data to form a causal state representation. A hierarchical reward structure was designed, and an adaptive weighting mechanism based on constraint satisfaction degree was introduced to coordinate multi-objective optimization. Physical constraint projection operators were embedded in the policy output layer, transforming constraints such as unit ramp rates, energy storage state of charge, and voltage magnitudes into feasible region mappings in the action space. Simulation results show that SC-PPO reduces voltage limit violations by 75% while improving wind power accommodation rate to 95.6% and reducing carbon emissions to 15 220 t, the research provides a new paradigm of intelligent decision-making that combines adaptability and security for high renewable energy penetration power systems.

Key words: deep reinforcement learning, power system dispatch, dynamic state representation, hierarchical reward mechanism, security constraint embedding