基于单纯形法的神经元PID控制器学习参数优化  

Optimization of Learning Factors for Neural PID Controller Based on Simplex Method

作　　者：朱学贵[1] 王毅[1] 昝建明 

机构地区：[1]北京交通大学电气工程学院,北京100044 [2]长安汽车工程研究院试验测试研究所,重庆400023

出　　处：《系统仿真学报》2006年第11期3030-3033,3037,共5页Journal of System Simulation

摘　　要：利用单纯形法优化神经元PID控制器的学习速率和神经元比例系数。一种是动态优化法,每次学习使用的学习参数由单纯形优化器在之前的学习过程中进行预测。为此,定义了一种新的目标函数,实现了学习参数的良好预测。另一种是离线优化法,这是一种全局优化法,使用单纯形优化器只需一步优化,得到最优学习参数。针对不同的初始学习参数和被控对象,经过大量的仿真实验,表明单纯形神经元PID控制器具有很好的动态性能和稳态性能,增强了系统的适应性和鲁棒性,降低了学习参数选择的盲目性和对经验的高度依赖性。Learning rate and neuron＇s scale factor of the neural PID controller are optimized by the simplex method. One is the dynamical optimization method that the learning factors at every step are previously predicted by the simplex optimizer. For the purpose, a new objective function is defined for the better prediction of optimal learning factors. Another is offiine global optimization method that one-off use of the simplex optimizer is equal to the optimal learning factors. Considering miscellaneous initial learning factors and controlled objects, quite a number of simulation experiments ate well done on computer to prove superiority of the simplex neural PID controller over the traditional one, which shows better dynamical and steady behavior, strengthens fitness and robustness of the system against outside situation, and weakens blindness to selection of learning factors and high dependency on experience.

关 键 词：PID控制器 神经网络 单纯形法 学习参数 优化 预测 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程] TP273[自动化与计算机技术—控制科学与工程]