基于神经元自增长消减的双神经网络预测控制  

Dual Neural Network Predictive Control Based on Neuron Self-growth and Subtraction

作　　者：杨青[1] 张雨 葛亮[2] 周建兴 YANG Qing;ZHANG Yu;GE Liang;ZHOU Jian-xing(School of Electrical Information,Southwest Petroleum University,Chengdu 610500,China;School of Mechanical and Electrical Engineering,Southwest Petroleum University,Chengdu 610500,China)

机构地区：[1]西南石油大学电气信息学院,四川成都610500 [2]西南石油大学机电工程学院,四川成都610500

出　　处：《控制工程》2023年第3期401-411,共11页Control Engineering of China

基　　金：国家自然科学基金面上项目(51974273);四川省国际科技合作与交流研究项目(18GJHZ0195)。

摘　　要：针对模型预测控制算法进行改进,提出了一种基于神经元自增长消减的双神经网络模型预测控制方法,利用可变结构的径向基神经网络精确逼近被控系统模型。变结构神经网络根据实际情况进行隐层神经元的自增长消减,可解决神经元个数难以确定的问题,在保证逼近精度的同时能够简化神经网络结构、减小计算量。针对现有滚动优化算法的局限性,在目标函数中的权重因子和初始参数选取方面做出改进,结合自适应权值方法引入逆神经网络结构确定初始值,解决了优化算法易陷入局部最优的问题。利用李雅普诺夫稳定性理论验证了改进算法的稳定性,并通过实验证明了所改进算法的有效性。In order to improve the model prediction control algorithm,a dual neural network model predictive control method based on neuron self-growth and subtraction is proposed.The variable structure RBF neural network is used to accurately approximate the controlled system model.The variable structure neural network performs self-growth and reduction of hidden layer neurons according to the actual situation,solves the problem that the number of neurons is difficult to determine,and simplifies the structure of the neural network and reduces the amount of calculation while ensuring the accuracy of the approximation.In view of the limitations of the existing rolling optimization algorithm,improvements are made in the selection of weight factors and initial parameters in the objective function,combined with the adaptive weight method,the inverse neural network structure is introduced to determine the initial value,and the optimization algorithm is easy to fall into the local optimal problem.Lyapunov stability theory is used to verify the stability of the improved algorithm,and the experiment proves the effectiveness of the improved algorithm.

关 键 词：神经网络 神经元自增长消减 权重因子 预测控制 

分 类 号：TP13[自动化与计算机技术—控制理论与控制工程]