Application Research Of Soft Sensor For NO_x Concentration At SCR Entrance Based On Interval Type 2 Fuzzy Neural Network

Selective catalytic reduction technology(SCR)of coal-fired power plants is currently main means of flue gas denitration.In actual production,the NO_x concentration value at entrance of the SCR system directly affects ammonia supply control system,so timely and accurate measurement of its value has important impacts on improving denitration efficiency,reducing ammonia escape rate,and reducing system energy consumption.Currently,(Continuous Emission Monitoring Systems,CEMS),widely used in SCR inlet NO_x concentration monitoring,have the advantages of mature technology,high measurement accuracy,and easy online calibration.However,in order to ensure the stability and accuracy of the sensor,before the measurement,a long sampling line needs to be set to remove impurities such as fly ash and water vapor in the flue gas,which causes a large lag between true and measured values,and due to the equipment disassembly measurement during system maintenance will lose continuity,which greatly affects the denitration efficiency of the SCR system.In response to this problem,build a soft sensor model based on mathematical methods to achieve real-time accurate measurement of the NO_x concentration at inlet of the SCR.(1)According to actual engineering and related literature research,15 parameters related to NO_x concentration characteristics were selected as initial variables.In order to make the simulation experiment conform to the actual production situation as much as possible,collected the historical operation data of DCS from a power plant,and obtained the experimental samples after pretreatment.(2)In order to simplify the model structure,reduce the amount of calculation,and improve network training speed.Using principal component analysis(PCA)method to reduce model input dimension,choose the final auxiliary variable from the 15 initial variables,furthermore,obtain training samples and test samples,with auxiliary variables as input and SCR inlet NO_x concentration as output.(3)The method of combining fuzzy mean clustering(FCM)and subtractive clustering(SCM)algorithm is used to identify the antecedent layer structure and initial membership function parameters of type 1 TS fuzzy neural network.Simultaneously,the improved SGPSO algorithm is used to optimize the back layer parameters,establish GJTSFNN soft sensor mathematical model.Programming on the MATLAB platform realized the construction of the model,and conducted a comparative experiment with the BP model and TSFNN model.The results show that the measurement accuracy and learning speed of GJTSFNN are better than those of other two models.thus,the effectiveness of algorithm is initially verified.(4)Using the method of combining interval type 2 fuzzy mean clustering(IT2FCM)and subtractive clustering(SCM)algorithm to identify the anterior layer structure of interval type2 fuzzy neural network,determine the membership function parameters,at the same time,the improved SGPSO algorithm is used to optimize the linear parameters of the back layer,improve the overall optimal search ability of the network.Finally,a GJIT2FNN soft sensor mathematical model was established.Similarly,programming on the MATLAB platform realizes the construction of the GJIT2FNN model,the simulation experiment is compared with the IT2FNN model based on gradient descent algorithm to adjust various parameters,which further proves the superiority of the proposed modeling method.