一种具有记忆自学习能力的快速动态寻优算法及其无功优化求解  

A fast dynamic optimization algorithm with memory and self-learning and its application on reactive power optimization

作　　者：余涛[1] 张孝顺 

机构地区：[1]华南理工大学电力学院,广州510641

出　　处：《中国科学：技术科学》2016年第3期256-267,共12页Scientia Sinica(Technologica)

基　　金：国家重点基础研究发展计划(批准号:2013CB228205);国家自然科学基金(批准号:5177051;51477055)资助项目

摘　　要：针对传统人工智能在随机复杂环境的适应及交互能力较低问题,有机地将经典强化学习Q（？）算法与多主体协同行为进行高度融合,提出了一种具有记忆自学习能力的快速动态寻优算法.该算法通过与外部环境反复的交互来进行自学习改进,并利用值函数矩阵储存状态-动作对记忆,提出了联系记忆方式,有效地对传统Q（？）算法的动作空间进行降维处理,减小了记忆矩阵的规模;基于多主体协同合作的概念,采用多个主体同时对记忆矩阵进行迭代更新,明显提高了更新速度;在预学习形成良好的记忆后,能快速地进行在线动态优化.最后,文章利用电力系统经典无功优化模型进行了算法测试,IEEE 118节点和IEEE 300节点标准算例仿真表明：本文所提算法在保证较高收敛性的同时,寻优速度能提高到遗传算法、蚁群算法、粒子群等传统人工智能方法的5~40倍,非常适用于大规模复杂电网的在线滚动无功优化.Based on the low adaptability and interaction of the conventional artificial intelligence（AI） in a complex stochastic environment, this paper proposes a novel fast dynamic optimization algorithm with memory and self-learning（MSL） by combining the classical Q（？） learning with the collaboration mechanism of multiagent. The proposed algorithm can learn new knowledge through the self-learning by interaction with the external environment, and can storage the memory of state-action pairs by updating the value function matrices. The association memory is introdued in the action space to reduce the size of the memory matrices, thus the matrices can be updated by multiagent simultaneously because of the combination of multiagent collaboration mechanism, and the convergence speed is obviously increased. After obtaining the good memory matrices in the pre-learning process, the MSL can be used for fast dynamic optimization. The performance of MSL has been fully tested for reactive power optimization on the benchmarked IEEE 118-bus and IEEE 300-bus systems. Comparative studies have not only demonstrated the high convergence stability of the proposed algorithm but also confirmed its fast convergence speed which can be approximate 5 to 40 times faster than that of conventional AI algorithms such as genetic algorithm, ant colony system, particle swarm optimization, and so on, and it provides a powerful tool for reactive power optimization of large-scale power systems.

关 键 词：自学习 记忆矩阵 快速动态寻优 多主体协同 无功优化 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]