Load frequency control in isolated island city microgrids using deep graph reinforcement learning considering extensive scenarios

To address the challenges of handling the dynamic load variations caused by the unpredictable nature and energy asymmetry of renewable energy sources in isolated microgrids, this study introduces a novel approach known as Learning-Enhanced Load Frequency Control (LE-LFC). This method conceptualizes controllers as autonomous entities capable of making independent decisions. It employs a sophisticated High Scene Generalization Soft Actor-Critic algorithm, augmented with transfer learning, to enhance decision-making speed, generalization, robustness, and efficiency. This algorithm leverages environmental data for interaction, aiming for optimal frequency management and economic operation of isolated urban microgrids. By incorporating a maximum entropy approach, it enhances the robustness of conventional deep reinforcement learning and integrates dominance learning to refine Soft Actor-Critic’s Q-value function update, mitigating overestimation issues and boosting algorithmic performance. In addition, transfer learning is utilized to bolster the agents’ learning efficacy and adaptability to new conditions. Demonstrated effectively in China Southern Grid’s island microgrid setup, LE-LFC emerges as an advanced solution for modern grid variability, offering superior robustness, adaptability, and learning speed, thus enabling flexible and efficient energy system management.