A Reducible-Loss-Based Deep Reinforcement Learning for Automatic Generation Control in Renewable-Integrated Power Systems

The large-scale integration of renewable energy sources, such as wind and solar power, introduces strong randomness and intermittency into modern power systems, leading to deteriorated frequency stability and degraded control performance in automatic generation control (AGC). To address these challenges, this paper proposes a reducible-loss-based deep reinforcement learning (RL) approach for AGC in renewable-integrated power systems. The proposed method improves the training efficiency and stability of the traditional RL framework by incorporating a reducible-loss experience replay strategy. This mechanism evaluates the learnability of training samples by calculating the difference between current and target network losses, enabling the algorithm to prioritize samples with higher learning potential while reducing the influence of noisy or low-quality samples. A two-area load frequency control model and a multi-area renewable-integrated AGC system are established to evaluate the effectiveness of the proposed method under step, random, and white-noise load disturbances. Simulation results demonstrate that the proposed approach achieves faster convergence, smaller frequency deviations, and improved control performance compared with conventional reinforcement learning algorithms, showing strong robustness against renewable fluctuations and stochastic disturbances.