Study On NOx Prediction At SCR Denitration Inlet Of Thermal Power Plant Based On LSTM Network

In China’s power market,coal-fired power generation has been dominant,and NOx produced by burning coal-fired boilers in thermal power plants is one of the main sources of air pollutants.With the increasing awareness of environmental protection in China,the optimization of flue gas denitrification has been put on the agenda.At present,most of thermal power plants adopt selective catalytic reduction(SCR)technology to achieve denitrification,so as to reduce NOx emissions.For the nitrogen oxide content in the flue gas,the plants generally use the automatic flue gas monitoring system to measure its concentration in real time,but the system will have a large delay in the measurement,so it can not accurately reflect the real-time change of the NOx concentration at the SCR inlet,so it can not guide the reactor action in time.In this paper,the NOx concentration at the entrance of SCR denitrification system is taken as the research object,and the long-term memory network(LSTM)is used to establish the prediction model and realize the estimation of NOx concentration at the entrance of SCR.In this paper,the mechanism of the main factors that affect the NOx generation in the boiler combustion process is analyzed.After preprocessing the o riginal data samples obtained by DCS on site,the feature variable selection method based on mutual information is adopted to extract the features of the original variables,and the auxiliary variables with the largest correlation and the smallest redundan cy between variables are screened out,so as to reduce the number of variables The coupling degree and data dimension can improve the modeling accuracy and reduce the calculation complexity.On the one hand,the nitrogen oxide prediction model based on LSTM is established and its validity is verified;on the other hand,although the accuracy of the prediction model established for LSTM is up to the expectation,the speed of its training model is relatively slow,and CNN convolution neural network can fully extract data features and reduce data dimensions,so as to improve the training speed of the model.Therefore,this paper further proposes the LSTM prediction model based on one-dimensional convolution,that is cnn-lstm.The input layer in LSTM is replaced by convolution layer and downsampling layer.The experimental results show that,given the same data set,the latter improves the accuracy of the prediction model,but also improves the training speed.Both methods can realize the accurate prediction of NOx,which is of great significance to reduce the NOx emission of thermal power units.