Research On Prediction Of NO_X Emission In A Deep Peak Shaving Coal-fired Boiler Based On Neural Network

Utility coal-fired boilers are the main emission sources of nitrogen oxides(NO_X)which seriously affect the environmental quality.At present,in order to reduce the NO_Xemission concentration of power stations,Low-Nitrogen Combustion System(LNCS)and Selective Catalytic Reduction(SCR)denitrification systems are commonly used in power station boilers to control the generation and removal of NO_Xin the flue gas.However,due to the large amount of system regulation variables and the difficulty of control,it is urgent to establish a comprehensive NO_Xcharacteristic model to support the optimization and control of pollutants in power plant boilers.On the other hand,under the"dual carbon"background,the depth peaking task of thermal power units is heavier,and the operating conditions of boilers are wider,which also puts higher demands on the model.Therefore,this paper takes a 660 MW power station coal-fired boiler as the object,and starts from the three aspects of sample quality,the interpretability of the model,and the multi-condition prediction ability of the model to study the prediction model for the full process NO_Xconcentration of the station boiler based on the neural network method.The main research content includes:(1)In terms of improving sample quality,this study systematically analyzes the formation and removal process of NO_Xduring the operation of a coal-fired boiler and introduces the process flow of a 660MW coal-fired boiler.Furthermore,it addresses the problems encountered when dealing with the operation data of coal-fired boilers,and introduces data preprocessing methods such as outlier detection,data standardization,steady-state detection,and variable importance evaluation.In particular,the steady-state detection of operating data based on the KPCA method is described in detail.By analyzing a case study,this research verifies that using appropriate data preprocessing methods can effectively improve the quality of modeling data and reduce the complexity of subsequent modeling.(2)In terms of model interpretability,firstly,a soft-measurement model of coal quality parameters is established based on experimental data,then the relationship between NO_Xconcentration and coal quality and boiler regulation parameters is calculated using mutual information entropy theory to determine the input features of the model.Finally,a knowledge fusion neural network(KFNN)is proposed to introduce monotonic physical knowledge into the system.Based on this model,a combustion-generated NO_Xconcentration prediction model considering the monotonic relationship of oxygen content was established to obtain the inlet NO_Xconcentration of SCR.On this basis,further consider the monotonic relationship between SCR outlet NO_Xconcentration and urea flow rate,and predict the NO_Xemission concentration of power plant boilers.The results show that compared with support vector regression,random forest,and back propagation neural network,The modeling method proposed in this paper has better generalization and interpretability.(3)In terms of improving the model’s ability to predict multiple working conditions,we propose a coal-fired boiler NO_Xemission prediction model based on the NGMM-DAE-NN algorithm,which combines the NGMM(Nonlinear Gaussian Mixture Model)algorithm and the DAE-NN(Denoising Autoencoder-Neural Network)algorithm.The model is analyzed and validated using actual operational data.The NGMM-DAE-NN model can classify the modal data of coal-fired boiler operation and reflect the complex nonlinear relationship among different measurement parameters.In addition,the NO_Xconcentration prediction model based on the NGMM-DAE-NN algorithm can support coal-fired boiler combustion optimization and ammonia injection adjustment in the SCR system under deep tuning background.