烟气脱硝喷氨量SA-RBF神经网络最优控制  

Sensitivity Analysis Radial Basis Function Neural Network Control on Spraying Ammonia Flow Denitrification

作　　者：周洪煜[1] 赵乾[1] 张振华 汪正海[1] 

机构地区：[1]重庆大学动力工程学院,重庆400030 [2]华电电力科学研究院,浙江杭州510663

出　　处：《控制工程》2012年第6期947-951,共5页Control Engineering of China

基　　金：重庆市科委重大科技攻关项目(CSTC 2009AB1008)

摘　　要：对于选择性催化还原(SCR)烟气脱硝装置喷氨量的精确控制,传统PID控制器的参数是基于设计负荷预先整定,在变工况下系统呈现出强非线性和滞后性,难以确保最佳控制量。通过引入动态结构的RBF神经网络,利用敏感度法来增加和删除神经元,解决RBF神经网络结构过大或过小的问题,保证预测网络结构的精度。该网络综合学习SCR脱硝装置主要相关参数,以NOx排放量与设定值之间误差最小作为训练信号,实现喷氨量的最优控制。实验结果表明,在变工况下,此方案与传统PID相比,能满足SCR出口NOx排放量,有效减少了氨气逃逸量,具有良好的变工况适应能力。For sparing ammonia of flue gas denitrification device based on the selective catalytic reduction（ SCR）, it is difficult to over- come the system lags and nonlinearity in variable loads to ensure the optimal control volume with the traditional PID controllers, which are based on pre-tuning the design load. This paper introduces the dynamic structure radial basis function（RBF） neural network and uses sensitivity analysis（SA） method to insert or prune neurons, so that the final structure of RBF is not too large or small for the system and the forecasting precision is ensured. This SA-RBF method can comprehensively learn the main relative state parameters of SCR denitrification device, so as to get the optimal spraying ammonia by using least NOx emission. The results show that, when there are variable working conditions as well as flue gas temperature fluctuations, this scheme reduces the NOx emission and ammonia escape in export SCR, and it has a good adaptability, compared with traditional PID.

关 键 词：选择性催化还原 烟气脱硝 径向基神经网络 动态结构 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]