提出一种基于大数据的神经网络辨识多输入单输出(MultipleInput SingleOutput,MISO)系统的方法，采集现场运行的锅炉燃烧系统总风量、总煤量、炉膛氧量等历史大数据，首先使用数据平滑、去除趋势性、归一化等步骤进行数据预处理，然后利用近邻法删选出表征系统特性的样本数据集，利用神经网络模型进行训练后挖掘出数据之间的关系，最后在升、降30%负荷的情况下分别进行模型预测。结果表明,虽然只将采集到约0.658%数据容量进行训练，但在对整个大数据容量进行测试时，模型误差仍在允许的范围内。

A method of neural network identification MISO system based on large data is presented in this paper.After collecting the data of the total air volume,total coal and oxygen in the boiler combustion system.the data preprocessing is performed as data smoothing,trend removing and normalization.Then,the nearest neighbor method is used to delete the sample data set which can represent the system characteristics.At last,the model is predicted in the cases of increasing and decreasing 30% load.The results show that although only about 0.658% of the collected data capacity is trained,the model error is still within the allowable range when testing the entire large data capacity.

TK227.1

上海市科学技术委员会项目（16111106300，17511109400）；上海市科学技术委员会工程技术研究中心项目（14DZ2251100）