电子学报 ›› 2006, Vol. 34 ›› Issue (10): 1847-1851.

• 论文 • 上一篇    下一篇

基于多智能体强化学习的焦炉集气管压力多级协调控制

秦 斌1,2, 吴 敏1, 王 欣1,2, 阳春华1   

  1. 1. 中南大学信息科学与工程学院,湖南长沙 410083;2. 湖南工业大学电气工程学院,湖南株洲 412008
  • 收稿日期:2005-12-20 修回日期:2006-05-16 出版日期:2006-10-25 发布日期:2006-10-25

Multi-Level Coordination Control Based on Multi-Agent Reinforcement Learning for the Pressure of Gas Collectors of Coke Ovens

QIN Bin1,2, WU Min1, WANG Xin1,2, YANG Cun-hua1   

  1. 1. School of Information Science & Engineering,Central South University,Changsha,Hunan 410083,China;2. School of Electrical Engineering of Hunan University of Technology,Zhuzhou,Hunan 412008,China
  • Received:2005-12-20 Revised:2006-05-16 Online:2006-10-25 Published:2006-10-25

摘要: 针对焦炉集气管压力这类多变量强扰动非线性耦合系统,提出了一种基于Multi-Agent system(MAS)的焦炉集气管压力智能多级协调控制系统方案.采用基于Agent单元系统梯级协调体系和基于任务分解的实时Agent的组织与演化机制,通过Agent模态变迁进行模式切换,以适应快速突变环境.在控制Agent中采用Actor-critic强化学习方法,运用TS回归模糊神经网络实现行动和评判模块,使用分布式学习算法对多个Agent协调优化.工程应用表明,提出的控制策略有效地解决了高压氨水大干扰对集气管压力的冲击控制问题.

关键词: 焦炉集气管, 梯级协调, 多智能体, 强化学习, TS回归模糊神经网络

Abstract: For the multi-variable nonlinear coupled system with strong disturbance such as the gas pressure of collectors of coke ovens,this paper proposes an intelligent multi-level coordinated control strategy based on multi-agent system.It adopts the multi-level coordination architecture with agent agency and the organization and evolution mechanism based on task decomposition.The system can be switched to different modes using the state change of agents in order to operate in rapidly time-varying environments.The reinforcement learning method is used in Agent learning,the TS type recurrent fuzzy neural network(TSRFNN) is employed to realize the actor-critic elements.The agents in system are optimized coordinately by using the distributed learning algorithm.The real-world application shows that the proposed control strategy has successfully solved the process coordination control problem of the gas pressure of collectors of coke ovens with the strong disturbance produced by high pressure ammonia.

Key words: gas collectors of coke ovens, cascade coordination, multi-Agent system, reinforcement learning, TS recurrent fuzzy neural network

中图分类号: