基于多智能体强化学习的多能互补能源系统优化运行OA

To address the dynamic coordinated optimization challenges of multi-energy complementary systems under high renewable energy integration and the limitations of traditional centralized methods in multi-agent interest coordination and real-time response,dynamic optimization modeling research was conducted.A three-layer multi-agent reinforcement learning(MARL)framework—consisting of a physical layer,decision layer,and coordination layer—was developed,with energy producers,consumers,and system schedulers classified as independent agents.Based on the improved proximal policy optimization algorithm,a dynamic reward function integrating economic efficiency,environmental friendliness,and stability was designed,and distributed decision-making with global coordination was achieved through a centralized training-decentralized execution mechanism.A typical park-level multi-energy complementary system was used as a case study.The results showed that the proposed MARL model increased the renewable energy consumption rate to 92.3%,reducing the unit electricity cost by 28.9%compared to the traditional mixed integer programming(MIP)method.Under a 50%load abrupt change scenario,the system recovery time was shortened to 90 s,which was 900%faster than the MIP method.Even with±20%wind and solar forecasting errors,the load satisfaction rate remained at 98.7%.This dynamic optimization model effectively addressed the multi-agent coordination and uncertainty adaptation challenges in multi-energy complementary systems,providing technical support for the real-time optimization and scheduling of high-penetration renewable energy systems.