改进的神经网络PID火电厂主汽温控制研究  

Study of the Control Over the Main Steam Temperature in a Thermal Power Plant Based on an Improved Neural Network PID(Proportional,Integral and Differential) Control

作　　者：高昆仑[1] 梁宵[1] 王杰[1] 张衡[1] 

机构地区：[1]郑州大学电气工程学院,河南郑州450001

出　　处：《热能动力工程》2012年第6期709-714,740-741,共6页Journal of Engineering for Thermal Energy and Power

摘　　要：针对传统神经网络PID控制系统存在的问题和不足,提出了改进措施。对于网络的结构,通过加入一层单连接的网络层,来干预网络输出所对应的PID控制器的参数。对于网络连接权值的学习策略,选择了一个实时监测系统误差的参数指标,在每个控制周期内,首先根据误差指标决定网络是否需要学习,如果不需要学习,直接采用上一控制周期的PID参数进行控制。通过对火电厂主汽温模型的仿真实验表明,改进后的神经网络PID控制系统,不论是动态性能还是静态性能都明显优于传统神经网络PID,而且网络的训练次数由改进前的7000次减少到1732次,减少了70%以上。此外,改进后的控制系统的鲁棒性也没有受到影响。In the light of problems and shortcomings existing in the traditional neural network PID control systems,presented were measures for improvement.For the structure of the network,by adding a single-connected network layer,the parameters of the PID controller corresponding to the output of the network were intervened.As for the tactics for learning the network linkage weight value,a parameter index was chosen to real time monitor the error of the system.Within each control periods,the error index was based to first determine whether or not it is necessary for the network to learn.If it is not necessary to learn,the PID parameters of the last control period can be used directly for control.The simulation test results obtained by using the model for main steam temperatures of thermal power plants show that the improved neural network PID control system is obviously superior to the traditional one in terms of both dynamic performance and static one.Furthermore,the training frequency of the network decreased from 7000 times before the improvement to 1732 times,lowering by more than 70%.In addition,the robustness of the improved control system has not yet been affected.

关 键 词：神经网络 PID 学习 权值 主汽温 

分 类 号：TM714[电气工程—电力系统及自动化] O242[理学—计算数学]