一种基于神经网络的混合气体体积分数预测模型

摘要

Aimed at the problems of too long modeling time and poor generalization in prediction modeling of mixed gas,principal component extraction(PCE) combined with Bayesian regularization neural network method is used. The mixed infrared absorption spectrum of four common polluted gases CH4 ,CO,SO2 and NO2 is analyzed,and each single gas volume fraction is obtained respectively. The network is constructed by programming with Matlab software,and the network parameters are optimized. The result shows that the modeling time of the network reduces from 4 250 s to 8 s, but the prediction goodness of fitting keeps mostly unchangeable, reaching to 95.1％.Compared to conventional back-propagation (BP)neural network, the method has better prediction effect and practical significance in quantitative analysis of mixed gas in air pollution.%针对目前混合气体预测建模中的建模时间过长和泛化能力较差的问题,采用主成分提取(PCE)结合贝叶斯正则化神经网络法进行了改进.通过对4种常见污染气体CH4,CO,SO2,NO2的混合红外吸收光谱进行了分析,得到了各单一气体的体积分数.使用Matlab软件编程构建了网络,并优化了网络参数.结果表明:该方法使网络建模时间从4250 s减少到8 s,但预测拟合度基本不变,达到了95.1%,优于常规的反向传播(BP)神经网络,对于大气污染多气体定量分析具有实际意义.