关键词: 火电机组, ＳＣＲ烟气脱硝系统, ＬＳＴＭ神经网络, 注意力机制

Abstract: To address the problem of low accuracy of export nitrogen oxides (NOx) prediction model of selective catalytic reduction (SCR) system in coal-fired units,this paper develops a real-time prediction model of outlet concentration based on the long short-term memory (LSTM) neural network algorithm optimized by attention mechanism (AM).In the LSTM model,the multivariable and multi-scale features are extracted from the SCR denitrification operation data of thermal power plant,which lead to the nonlinear relationship between input time series and outlet NOx concentration time series.The AM is further adopted to optimize the weight value of the LSTM hidden layer output series,from which the SCR denitrification system NOx outlet concentration prediction model is finally obtained.Experimental results of the operation data from a domestic 350 MW thermal power unit denitrification device show that the AM-LSTM,which achieves higher prediction accuracy and stronger generalization ability than the RNN and LSTM algorithms,may be further applied to large-scale denitrification scenarios with more parameters.