Dynamic soft-sensing method for NOx emission concentration at outlet of SCR system

Abstract

The present invention provides a dynamic soft measurement method for NOx emission concentration at the outlet of an SCR system. The method selects actual operating data that affects changes in NOx emission concentration at the outlet of the SCR system through mechanism analysis; uses an extreme value gradient boosting method to perform importance ranking and select candidate input variables for a subsequent prediction model; estimates the delay time of each relevant variable relative to the NOx emission concentration based on a mutual information coefficient method, a random forest algorithm, and a Pearson algorithm, and reconstructs modeling data; then uses ensemble empirical mode decomposition to decompose the input time series and mine deep frequency domain information in the input time series; again uses a lasso algorithm to perform feature selection on the data information obtained from the ensemble empirical mode decomposition, screens variables with greater correlation as sample data for a prediction model, and uses a deep neural network model to predict the NOx emission concentration at the outlet of the SCR system.

Description

Technical Field

[0001] The invention belongs to the technical field of thermal power generation, and in particular relates to a dynamic soft measurement method for NOx emission concentration at an outlet of an SCR system. Background Art

[0002] Nitrogen oxide (NOx) emissions are one of the main causes of photochemical smog and acid rain, causing environmental pollution and harming human health. Selective catalytic reduction (SCR) devices, with their simple structure, mature technology, and high denitrification efficiency, have been widely used in thermal power plants.

[0003] Data-driven modeling methods are widely used to predict NOx emission concentrations at the outlet of SCR systems. However, these methods suffer from two challenges: First, due to the varying timing of input variables participating in the combustion reaction during the boiler combustion process and sensor measurement delays, the NOx concentration generated by combustion is delayed relative to other operating parameters. Consequently, the relevant parameters in the operating data recorded by the distributed control system (DCS) under the same time tag lack an accurate correspondence, and directly using DCS data for modeling can affect model accuracy. Therefore, a reasonable estimation of time delay plays a crucial role in data-driven modeling. Second, due to the randomness, redundancy, and instability of actual power plant operating data, it is difficult to extract deep information from the data, resulting in significant deviations between the predicted NOx emission concentrations and the measured NOx emission concentrations. To achieve efficient and accurate prediction of NOx emission concentrations at the outlet of SCR systems, it is necessary to consider the time lag and redundancy of the input data, obtain accurate and reliable data based on actual operating conditions, and thus implement soft measurement of NOx emission concentrations at the outlet of SCR systems. Summary of the Invention

[0004] In view of the above problems, the present invention provides a dynamic soft measurement method for NOx emission concentration at the outlet of an SCR system, comprising:

[0005] Step 1: Collect actual operating data including the NOx emission concentration at the outlet of the SCR system and related variables that affect the change of the NOx emission concentration at the outlet of the SCR system from the power plant DCS, and express the collected data set as: DATA = {(X, Y)}, where X is the data set of related variables that affect the change of the NOx emission concentration at the outlet of the SCR system, and Y is the collected data set of the NOx concentration at the outlet of the SCR system. x i (t j ) represents the i-th related variable x i Corresponding to t jThe sampling data at the moment, i = 1, 2, ..., d, j = 1, 2, ..., N, d represents the number of samples collected, N represents the number of values ​​in the sample, Y is the NOx concentration at the outlet of the SCR system, y(t N ) represents t N The sampling data of NOx concentration at the outlet of the SCR system at the moment is used, and the data set DATA is preprocessed. The specific expression is:

[0006] Step 1.1: Based on the collected actual operation data, calculate the standard deviation σ and mean μ of each column of data in the relevant variable data set X;

[0007] Step 1.2: Define the data in each column that does not fall between (μ-3σ, μ+3σ) as outliers. Calculate the average of the 10 previous moments of the moment corresponding to each outlier, and replace the outlier with the average.

[0008] Step 1.3: Repeat steps 1.1 to 1.2 to replace the outliers in each column of the relevant variable dataset X to obtain the preprocessed dataset X';

[0009] The relevant variables affecting the change of NOx emission concentration at the outlet of the SCR system are unit load, ammonia slip, outlet O2 concentration, outlet flue gas temperature, outlet flue gas pressure, ammonia injection amount, inlet O2 concentration, inlet flue gas temperature, inlet flue gas pressure, total air volume, dilution fan current, total coal volume and SCR system inlet NOx concentration;

[0010] Step 2: Use the extreme value gradient lifting method to calculate the importance of each column of data in the preprocessed data set X' and the NOx concentration at the outlet of the SCR system. Arrange the data in X' from the most important to the least important, and select the first m columns of data, that is, select m relevant variables;

[0011] Step 3: Calculate the delay time of the m relevant variables selected in step 2 relative to the NOx concentration at the outlet of the SCR system using the mutual information coefficient method (MI), random forest algorithm (RF), and Pearson algorithm, and reconstruct the model data. The specific expression is:

[0012] Step 3.1: Perform time-delay reconstruction on each column of relevant variable data in X', that is, select the time [t1, t K ] moment as the starting variable, construct the delay time calculation data set, and obtain each related variable x i The corresponding K groups of delay time calculation data sets X' i :

[0013]

[0014] Perform time-delay reconstruction on the data in the SCR system outlet NOx concentration dataset Y to obtain the reconstructed data sequence Y':

[0015]

[0016] Among them, y(t n+K-1 ) represents t n+K-1 Sampling data of NOx concentration at the outlet of SCR system at time t, data reconstruction sequence X' i The length of each column in Y' is n, n≤N;

[0017] Step 3.2: Calculate the delay time using the mutual information coefficient method to calculate the data set X' i =[X' i,1 X' i,2 … X' i,K ] between the K columns of data and the reconstructed data sequence Y', a total of K mutual information entropy values ​​are obtained, expressed as M = (m1, m2, ..., m K );

[0018] Step 3.3: Use random forest method to calculate the delay time calculation data set X' i The Gini index G between the K columns of data and the reconstructed data sequence Y' is obtained, and a total of K Gini index values ​​are obtained, which are expressed as G = (g1, g2, ..., g K ):

[0019] Step 3.4: Use the Pearson algorithm to calculate the delay time of the data set X' i The Pearson coefficient value P between the K columns of data and the reconstructed data sequence Y' is obtained, and a total of K Pearson coefficient values ​​are obtained, which is expressed as P = (p1, p2, ..., p K ):

[0020] Step 3.5: Calculate the delay time data set X' i The mutual information entropy, Gini index, and Pearson coefficient calculated from the K columns of data and the reconstructed data sequence Y' are weighted. The weighted formula is expressed as I = (αM + βG + ηP) = (αm1 + βg1 + ηp1, αm2 + βg2 + ηp2, ..., αm k +βg k +ηp k ), and find the maximum value max(αm after weighting k +βg k +ηp k ) corresponds to the sequence K′, and the time corresponding to K′ is called the maximum delay time T i ;

[0021] Step 3.6: Set time T i The corresponding delay time calculation data set X' i sequence in The reconstructed data sequence is retained and the remaining sequences are deleted;

[0022] Step 3.7: Repeat steps 3.1 to 3.6 to obtain a total of m sequences The sequence matrix composed of m sequences is expressed as:

[0023] in It is represented by the i-th variable in T i The sequence corresponding to the moment;

[0024] Step 4: Use Empirical Mode Decomposition (EMD) to analyze the data sequence Each column in is decomposed into After decomposition, it is expressed as: Among them, IMF is the intrinsic mode function of eigendecomposition, RES i is the residual, c means that the characteristic column vector decomposes c intrinsic mode functions;

[0025] Step 5: Use the Lasso algorithm to determine the importance of each decomposed variable relative to the NOx emission concentration at the SCR system outlet, sort them from most important to least important, and select the top Q variables;

[0026] Step 6: Build a deep neural network (DNN) model as a prediction model for the NOx emission concentration at the SCR system outlet. Use the data sequence of the first Q variables to train the model parameters. Use the trained prediction model to output the predicted value of the NOx emission concentration at the SCR system outlet in real time.

[0027] Step 6.1: Divide the dataset into training and test sets;

[0028] Step 6.2: Determine the number of nodes in the DNN input layer, hidden layer, hidden layer, and output layer.

[0029] Step 6.3: Randomly initialize parameters, including the weights of the network structure and the bias of the hidden layer;

[0030] Step 6.4: Use the backpropagation algorithm to update the network weights w and the bias b of the hidden layer;

[0031] The deep neural network model is represented as:

[0032] in, Represents the predicted value of the i-th moment predicted by the DNN network, x i is the input variable corresponding to the i-th moment after reconstruction and selection in steps 1 to 5.

[0033] The beneficial effects of the present invention are:

[0034] The present invention adopts a deep neural network model to predict the NOx concentration at the outlet of the SCR system at the next moment, and considers the redundancy and time delay characteristics of the data, designs a time delay combination based on mutual information, random forest, and Pearson algorithm, calculates the delay time of each process variable relative to the NOx emission concentration at the outlet of the SCR system, and reconstructs the modeling data. After obtaining the reconstructed data, the data after time delay analysis are decomposed based on the empirical mode, and the data information obtained by the empirical mode decomposition is once again subjected to feature selection to reduce the dimension of the modeling data; based on the data selected by the Lasso algorithm, a prediction model for the NOx emission concentration at the outlet of the SCR system based on the DNN network is established. The original operating data is processed by time delay combination and empirical mode decomposition, and the data information is deeply mined, providing reliable input variables for subsequent prediction modeling, thereby solving the technical problem that the existing model has low prediction accuracy for the NOx concentration at the outlet of the SCR system, and is used for the subsequent prediction of the NOx concentration at the outlet of the SCR system. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 This is a flow chart of a dynamic soft measurement method for NOx emission concentration at the outlet of an SCR system in the present invention;

[0036] Figure 2 Empirical mode decomposition diagram of NOx concentration at the inlet of the SCR system in the present invention;

[0037] Figure 3 It is the predicted value of NOx concentration at the outlet of the SCR system on the test set in the present invention. DETAILED DESCRIPTION

[0038] The invention is further described below with reference to the accompanying figures and specific implementation examples. The method includes data acquisition and preprocessing, designing a time delay combination selection algorithm and reconstructing data based on delay time, data decomposition, and dimensionality reduction of characteristic variables. Ultimately, it establishes an outlet NOx concentration prediction model to address the issue of low prediction accuracy.

[0039] like Figure 1 As shown, a dynamic soft measurement method for NOx emission concentration at the outlet of an SCR system includes:

[0040] Step 1: Collect actual operating data including the NOx emission concentration at the outlet of the SCR system and related variables that affect the change of the NOx emission concentration at the outlet of the SCR system from the power plant DCS, and express the collected data set as: DATA = {(X, Y)}, where X is the data set of related variables that affect the change of the NOx emission concentration at the outlet of the SCR system, and Y is the collected data set of the NOx concentration at the outlet of the SCR system. x i (t j ) represents the i-th related variable x i Corresponding to t j The sampling data at the moment, i = 1, 2, ..., d, j = 1, 2, ..., N, N means each sample contains N values, d means the number of samples collected, Y is the NOx concentration at the outlet of the SCR system, y(t N ) represents t N The sampling data of NOx concentration at the outlet of the SCR system at the moment is used, and the data set DATA is preprocessed. The specific expression is:

[0041] Step 1.1: Based on the collected actual operation data, calculate the standard deviation σ and mean μ of each column of data in the relevant variable data set X (i.e., each relevant variable sampling data);

[0042] Step 1.2: Define the data in each column that does not fall between (μ-3σ, μ+3σ) as outliers. Calculate the average of the 10 previous moments of the moment corresponding to each outlier, and replace the outlier with the average.

[0043] Step 1.3: Repeat steps 1.1 to 1.2 to replace the outliers in each column of the relevant variable dataset X to obtain the preprocessed dataset X';

[0044] The relevant variables affecting the change of NOx emission concentration at the outlet of the SCR system are unit load, ammonia slip, outlet O2 concentration, outlet flue gas temperature, outlet flue gas pressure, ammonia injection amount, inlet O2 concentration, inlet flue gas temperature, inlet flue gas pressure, total air volume, dilution fan current, total coal volume and SCR system inlet NOx concentration;

[0045] Step 2: Use the extreme value gradient lifting method to calculate the importance of each column of data in the preprocessed data set X' and the NOx concentration at the outlet of the SCR system. Arrange the data in X' from the most important to the least important, and select the first m columns of data, that is, select m relevant variables;

[0046] The relevant variables after feature selection in this embodiment are shown in Table 1:

[0047] Table 1 Actual operation data

[0048]

[0049]

[0050] Step 3.1: Perform time-delay reconstruction on each column of relevant variable data in X', that is, select the time [t1, t K ] moment as the starting variable, construct the delay time calculation data set, and obtain each related variable x i The corresponding K groups of delay time calculation data sets X' i :

[0051]

[0052] Perform time-delay reconstruction on the data in the SCR system outlet NOx concentration dataset Y to obtain the reconstructed data sequence Y':

[0053]

[0054] Among them, the data reconstruction sequence X' i The length of each column of data in Y' is n, n≤N, n=1000, K=61;

[0055] Step 3.2: Calculate the delay time using the mutual information coefficient method to calculate the data set X' i =[X' i,1 X' i,2 … X' i,K ] between the K columns of data and the reconstructed data sequence Y', a total of K mutual information entropy values ​​are obtained, expressed as M = (m1, m2, ..., m K );

[0056] Step 3.3: Use random forest method to calculate the delay time calculation data set X' i The Gini index G between the K columns of data and the reconstructed data sequence Y' is obtained, and a total of K Gini index values ​​are obtained, which are expressed as G = (g1, g2, ..., g K ):

[0057] Step 3.4: Use the Pearson algorithm to calculate the delay time of the data set X' i The Pearson coefficient value P between the K columns of data and the reconstructed data sequence Y' is obtained, and a total of K Pearson coefficient values ​​are obtained, which is expressed as P = (p1, p2, ..., p K ):

[0058] Step 3.5: Calculate the delay time data set X' iThe mutual information entropy, Gini index, and Pearson coefficient values ​​calculated from the K column data and the reconstructed data sequence Y' are weighted, and the weighted formula is expressed as

[0059] I=(αM+βG+ηP)=(αm1+βg1+ηp1,αm2+βg2+ηp2,…,αm k +βg k +ηp k ), and find the maximum value max(αm after weighting k +βg k +ηp k ) corresponds to the sequence K′, and the time corresponding to K′ is called the maximum delay time T i , where α=1, β=1, η=1;

[0060] Step 3.6: Set time T i The corresponding delay time calculation data set X' i sequence in The reconstructed data sequence is retained and the remaining sequences are deleted;

[0061] Step 3.7: Repeat steps 3.1 to 3.6 to obtain a total of m sequences The sequence matrix composed of m sequences is expressed as:

[0062] in It is represented by the i-th variable in T i The sequence corresponding to the moment, m = 8;

[0063] Step 4: Use Empirical Mode Decomposition (EMD) to analyze the data sequence Each column in is decomposed into After decomposition, it is expressed as: Among them, IMF is the intrinsic mode function of eigendecomposition, RES i is the residual, c means that the characteristic column vector decomposes c intrinsic mode functions, such as Figure 2 As shown in Figure 2, EMD decomposition is used to decompose the NOx emission concentration at the inlet of the SCR system into 8 intrinsic mode functions and a residual;

[0064] Step 5: Use the Lasso algorithm to determine the importance of each decomposed variable relative to the NOx emission concentration at the SCR system outlet, and sort them from most important to least important, selecting the top Q variables, where Q = 34;

[0065] Step 6: Build a deep neural network (DNN) model as a prediction model for the NOx emission concentration at the SCR system outlet. Use the data sequence of the first Q variables to train the model parameters. Use the trained prediction model to output the predicted value of the NOx emission concentration at the SCR system outlet in real time.

[0066] Step 6.1: Divide the dataset into a training set and a test set. The number of samples in the training set is 700 and the number of samples in the test set is 300.

[0067] Step 6.2: Determine the number of nodes in the DNN input layer, hidden layer, hidden layer, and output layer.

[0068] Step 6.3: Randomly initialize parameters, including the weights of the network structure and the bias of the hidden layer;

[0069] Step 6.4: Use the backpropagation algorithm to update the network weights w and the bias b of the hidden layer;

[0070] The deep neural network model is represented as:

[0071] in, Represents the predicted value at the i-th moment predicted by the DNN network, x i is the input variable corresponding to the i-th moment after reconstruction and selection through steps 1 to 5.

[0072] In this embodiment, a DNN neural network is used to model the selected process variables to predict the subsequent nitrogen oxide concentration. The prediction results are as follows: Figure 3 As shown, it can be seen that the method of the present invention can effectively predict the NOx emission concentration at the SCR outlet.

Claims

1. A dynamic soft measurement method for NOx emission concentration at the outlet of an SCR system, characterized in that: include: Step 1: Collect actual operating data including the NOx emission concentration at the outlet of the SCR system and related variables that affect the change of the NOx emission concentration at the outlet of the SCR system from the power plant DCS, and perform data preprocessing; Step 2: Use the extreme value gradient lifting method to calculate the importance of each column of data in the preprocessed data set X' and the NOx concentration at the outlet of the SCR system. Arrange the data in X' from the most important to the least important, and select the first m columns of data, that is, select m relevant variables; Step 3: Calculate the delay time of m related variables relative to the NOx concentration at the outlet of the SCR system using the mutual information coefficient method, random forest algorithm, and Pearson algorithm, and reconstruct the modeling data. Step 3.1: Select time t1 and [t1,t K ] as the starting variable, and perform delayed reconstruction on each set of related variable data to obtain each related variable x i The corresponding K groups of delay time calculation data sets X' i : Where x i (t n ) represents the i-th related variable x i Corresponding to t n The sampling data at the time, i = 1, 2, ..., d, d represents the number of samples collected; Perform time-delay reconstruction on the data in the SCR system outlet NOx concentration dataset Y to obtain the reconstructed data sequence Y': Among them, y(t n+K-1 ) represents t n+K-1 Sampling data of NOx concentration at the outlet of SCR system at time t, data reconstruction sequence X' i The length of each column in Y' is n, n≤N; Step 3.2: Calculate X' using the mutual information coefficient method i The mutual information entropy between the K columns of data and Y'; Step 3.3: Calculate X' using random forest method i The Gini index between the K column data and Y'; Step 3.4: Calculate X' using the Pearson algorithm i The Pearson coefficient value between the K column data and Y'; Step 3.5: Weight the corresponding information entropy, Gini index, and Pearson coefficient values ​​and find the time T corresponding to the maximum value after weighting. i ; Step 3.6: Set time T i The corresponding sequence The data sequence after reconstruction is retained, and the rest of the sequences are deleted; Step 3.7: Repeat steps 3.1 to 3.6 to obtain m data sequence matrices represented as: Step 4: Decompose the reconstructed data sequence using empirical mode decomposition; Step 5: Use the Lasso algorithm to determine the importance of each decomposed variable relative to the NOx emissions at the SCR outlet, and select the top Q variables with greater importance; Step 6: Build a deep neural network model as a prediction model for NOx emissions, use the data sequence of the first Q variables to train the model parameters, and use the trained prediction model to output the predicted value of SCR outlet NOx concentration in real time.

2. A dynamic soft measurement method for NOx emission concentration at an SCR system outlet according to claim 1, characterized in that: The step 1 comprises: Step 1.1: Calculate the standard deviation σ and mean μ of the corresponding data of each relevant variable based on the actual operating data collected; Step 1.2: Define the data that does not fall between (μ-3σ, μ+3σ) as outliers, calculate the average of the data 10 moments before each outlier, and replace the outlier with the average; Step 1.3: Repeat steps 1.1 to 1.2 to replace outliers and obtain the preprocessed data set.

3. The dynamic soft measurement method for NOx emission concentration at the outlet of an SCR system according to claim 1, characterized in that: The relevant variables affecting the change of NOx emission concentration at the outlet of the SCR system are unit load, ammonia slip, outlet O2 concentration, outlet flue gas temperature, outlet flue gas pressure, ammonia injection amount, inlet O2 concentration, inlet flue gas temperature, inlet flue gas pressure, total air volume, dilution fan current, total coal amount and SCR system inlet NOx concentration.

Images