基于神经网络的烧结SCR脱硝系统动态建模研究

Research on dynamic modeling of sintering SCR denitrification system based on neural network

提出一种基于注意力机制的CNN-LSTM神经网络预测模型,通过CNN提取局部特征、LSTM处理时序序列数据、注意力层提取信息,最终输出预测值的方式对脱硝系统进行动态建模。利用最大信息系数法筛选出相关性高的特征,以BP、CNN、LSTM、CNN-LSTM模型作为比较,采用均方根误差(E_(RMS))、平均绝对误差(E_(MA))和平均绝对百分比误差(E_(MAP))作为评价指标,评估NO_(x)浓度预测模型的性能。研究结果表明:基于注意力机制的CNN-LSTM模型预测NO_(x)浓度的准确度最高,E_(RMS)、E_(MA)、E_(MAP)分别0.6731、0.8670、1.5201%。该模型较其他对比模型相比,具有更高的预测精度、更强的泛化能力,可以实现对SCR系统的未来时间出口NO_(x)浓度的精准预测。通过该模型预测喷氨量可实现喷氨量的精准控制和减少NO_(x)排放的要求。

The CNN-LSTM neural network prediction model based on attention mechanism was proposed.The model can dynamically model the denitrification system by extracting local features from CNN,processing time series data from LSTM,extracting information from attention layer,and finally outputting predicted values.The maximum information coefficient method was used to screen out the features with high correlation.BP,CNN,LSTM and CNN-LSTM models were compared,and the root-mean-square error(E_(RMS)),mean absolute error(E_(MA))and mean absolute percentage error(E_(MAP))were used as evaluation indexes to evaluate the performance of the NO_(x)concentration prediction model.The results show that the CNN-LSTM model based on attention mechanism has the highest accuracy in predicting NO_(x)concentration,and E_(RMS),E_(MA)and E_(MAP)are 0.6731,0.8670 and 1.5201%,respectively.Compared with other comparison models,this model has higher prediction accuracy and stronger generalization ability,and can accurately predict the future time outlet NO_(x)concentration of SCR system.The prediction of ammonia injection quantity by this model can realize the precise control of ammonia injection quantity and meet the requirement of reducing NO_(x)emission.