Power plant boiler NOx concentration prediction method based on data reconstruction

Abstract

The invention relates to a power plant boiler NOx concentration prediction method based on data reconstruction, and the method comprises the following steps: firstly, obtaining a primary characteristic variable set influencing the NOx concentration from a power plant DCS database; secondly, sequentially performing data cleaning, steady state judgment and delay time analysis and correction on the primarily selected feature variable set, and extracting an optimal feature variable set by using an RRCT (Radio Resource Computed Tomography) algorithm; processing the optimal characteristic variable set according to working conditions to obtain an optimal characteristic variable subset corresponding to each working condition; and finally, establishing an LSSVM prediction model under each working condition according to the optimal characteristic variable subset, selecting the corresponding LSSVM prediction model according to the actual working condition, and obtaining a NOx concentration prediction result. Experimental results show that the data set extracted based on the data reconstruction strategy effectively improves the prediction precision and efficiency of the model.

Description

Technical Field

[0001] The present invention relates to the field of neural network technology, and in particular to a power plant boiler NO x Concentration prediction method. Background Art

[0002] As one of the most basic and important energy sources in modern society, electricity is of great significance in all aspects of economy, society and life. Although the power industry has been promoting the development of clean energy in recent years, thermal power generation will still be one of the main sources of electricity for a long time. As we all know, thermal power generation will bring environmental pollution problems, especially nitrogen oxide emissions, which is one of the important sources of air pollution and has become the focus of attention in the field of environmental protection. In view of this, the national environmental protection agency has set up a special supervision system for NOx emissions from thermal power plants. x Emissions regulatory standards are becoming increasingly stringent, and the environmental protection regulatory system is constantly improving, aiming to promote the power industry to develop in a greener and more sustainable direction. At present, the power industry mainly adopts two technical means to reduce NO x Emissions: Combustion process control and flue gas denitrification. Establish a system to accurately predict NO emissions after the combustion process of power plant boilers and before flue gas denitrification. x The concentration prediction model can provide a model basis for setting control parameters in the combustion process and optimizing the amount of ammonia injection in the denitrification process. x Emission has important theoretical significance and practical application value.

[0003] With the continuous advancement of machine learning technology, as well as the vigorous development and application of neural networks and artificial intelligence models, it is possible to build a more accurate prediction model for nitrogen oxides in thermal power plants. However, the quality of power plant data is often affected by many factors, including but not limited to measurement errors, equipment failures, and changes in meteorological conditions. These factors make power plant data highly complex and non-steady-state, which in turn restricts NOx prediction. x The existing technology is used to improve the NO concentration of power plant boilers. x Various attempts have been made to improve the concentration prediction accuracy, but there are always some unsatisfactory aspects. For example, a Chinese patent discloses a coal-fired power plant boiler NO based on LSO-CatBoost. x Emission prediction method (application number: 202310091793 1), and another example is the prediction method of nitrogen oxide content in thermal power plants based on fractional-order grey delay model disclosed in a Chinese patent (application number: 2021113085309). The above technical solutions have improved NO xHowever, the variable data set was only simply processed without deep and effective processing, especially the adverse effects caused by the high dimension, significant time delay and massive characteristics of power plant boiler data were not completely solved, which in turn affected the accuracy of the prediction model. Summary of the invention

[0004] The present invention provides a power plant boiler NO based on data reconstruction x The concentration prediction method solves the problem of NO concentration caused by the high dimension, significant time delay and massive amount of power plant boiler data. x In order to solve the problem of difficult to accurately predict concentration, the prediction model is rebuilt using the data set extracted by the data reconstruction strategy, which effectively improves the prediction accuracy and efficiency of the model.

[0005] A power plant boiler NO based on data reconstruction x The concentration prediction method comprises the following steps:

[0006] S1. Obtain the NO from the power plant DCS database X The set of primary characteristic variables for concentration;

[0007] S2. Perform data cleaning on the preliminary feature variable set to obtain a cleaned feature variable set;

[0008] S3, performing steady-state discrimination on the cleaning feature variable set based on the sliding window method and standard deviation to obtain a steady-state feature variable set;

[0009] S4, analyzing the delay characteristics between variables of the steady-state characteristic variable set, and reconstructing the delay of each variable according to the obtained delay time to generate a delay reconstruction characteristic variable set;

[0010] S5, using a correlation, redundancy and complementarity trade-off algorithm to process the delay reconstruction feature variable set to obtain an optimal feature variable set;

[0011] S6. Using the kernel fuzzy C-means clustering method based on the differential evolution algorithm to divide the working conditions, and dividing the optimal feature variable set according to the working conditions to obtain the optimal feature variable subsets under different working conditions;

[0012] S7. Establish the LSSVM prediction model under each working condition according to the optimal feature variable subset under the working condition, select the corresponding LSSVM prediction model according to the actual working condition, and obtain NO x Concentration prediction results.

[0013] Preferably, in S1, according to NO x The Spearman correlation analysis of production mechanism and power plant DCS database data obtained the factors affecting NO xThe primary characteristic variables of the concentration are selected, and a set of primary characteristic variables is obtained, wherein the primary characteristic variables include unit load, total fuel quantity, oxygen quantity, total air volume, total primary air volume, total secondary air volume, main steam pressure, main steam flow, furnace smoke temperature, AA-FF layer de-cyclone secondary air adjustment damper position feedback, SOFA1-6 layer burnout air adjustment damper position feedback, A-F coal feeder coal feeding rate feedback signal and A-F coal mill inlet primary air volume.

[0014] Preferably, in S2, the data cleaning includes outlier processing and noise reduction processing, and the outlier processing includes checking whether there is a zero value in the preliminary variable data set; if a zero value is detected, it indicates that the group of data may correspond to a unit shutdown state or an abnormal situation, and has no modeling value, so it is deleted; then the box plot method is used to identify outliers, and the data is replaced with the average value; the sliding average method is used to perform noise reduction on the preliminary variable data set.

[0015] Preferably, in S3, it is first determined that the allowable fluctuation range of the main steam pressure under steady-state conditions is ±2%, and then the sliding window method and standard deviation are introduced to continuously monitor the changes in the main steam pressure within the set sliding time window. If the fluctuation amplitude of the main steam pressure within the window remains within the range of ±2%, the data within the window is determined to be valid data under steady-state conditions, and the valid data is extracted to obtain a steady-state characteristic variable set.

[0016] Preferably, in S3, when the sliding window method is used to determine the steady-state operating condition, the window width is determined by comparing the proportion of steady-state data points to non-steady-state data points under different windows.

[0017] Preferably, in S4, the maximum information coefficient method is used to solve the maximum information coefficient between the variable and the target variable at different lag times, wherein the time point corresponding to the maximum value of the maximum information coefficient is the optimal delay time between the variable and the target variable, and each variable is reconstructed with a time delay according to the obtained delay time to eliminate the influence of the delay time on the characteristic relationship between the variables; specifically, the following steps are included:

[0018] S41. Assume that the input variable is X={X1(t),X2(t),...,X M (t)}; furnace outlet NO x The concentration is Y = Y(t), t is the sampling time, and the empty set of delayed reconstruction data is Z = {Z1, Z2, ..., Z M}; Each input variable is equal to NO x The time delay of concentration is different, and each variable X needs to be N (t), N∈[1,M] performs delay reconstruction;

[0019] S42, assuming that the time delay is τ={τ1,τ2,...,τ g}, for each input variable X N (t) The vector after compensating the time delay is expressed as:

[0020] X N ={X N (t-τ i ),i=1,2,...,g};

[0021] S43, calculate the maximum information coefficient of each input variable and output variable after embedding time delay: MIC(X N (t-τ d ); Y), τ d ∈τ, where the maximum information coefficient corresponds to the maximum value d The value is the delay time of each input variable;

[0022] S44, according to the delay time τ of each input variable d , to perform timing matching:

[0023] Z N =X N (t-τ d );

[0024] S45, the final output delay reconstruction feature variable set Z * ={Z1,Z2,...,Z M ,Y}.

[0025] Preferably, in S5, the delay reconstruction feature variable set is processed using a correlation, redundancy and complementarity trade-off algorithm to screen out the variables that can characterize NO x The optimal characteristic variable set with more comprehensive concentration change characteristics specifically includes the following steps:

[0026] S51, reconstructing the characteristic variable set Z for the time delay * ={Z1,Z2,...,Z M ,Y} calculate the characteristic variable Z i With Z j Between and Z i The Spearman correlation coefficient between the target variable Y is denoted as ρ ij and r i , i,j∈[1,M]; according to ρ ij and r i The information theoretical quantity IT is obtained as:

[0027] IT=-0.5·(1-h 2 ), h = r i or ij (1)

[0028] S52, according to formula (1), the correlation r IT (Z j ,Y) and redundancy r IT (Z i ,Z j ), the specific expression is:

[0029]

[0030] The matrix D is generated by formulas (2) and (3). The expression of matrix D is:

[0031]

[0032] In the matrix D, the main diagonal represents the correlation r IT (Z j ,Y), the upper triangle area represents the redundancy r IT (Z i ,Z j );

[0033] S53, add the feature variable with the largest value on the main diagonal of the matrix D as the first optimal feature variable S1 to the optimal feature variable set S; then select the candidate feature Z with the largest correlation with the target variable Y under the given selected feature set from the remaining feature variables Q. j ,j=1,2,...,M-1, and take out r at the same time IT (Z j ,Z s ), s∈S, calculate the candidate feature Z j The average redundancy of all selected features s∈S, that is, Then calculate the new feature Z j In Control Selected Feature Z s The partial correlation coefficient r between the response variable Y p (Z j ,Y|Z s ) and convert it to r p.IT form;

[0034] S54. Define the scoring formula of the complementarity trade-off algorithm to evaluate Z j Performance:

[0035]

[0036] In the formula, sign(·) is used to determine whether r should be increased or decreased in the scoring formula. p.IT If the sign (·) is a positive number, +1 is returned, indicating that r should be increased. p.IT Otherwise, it returns -1, indicating that r should be reduced. p.IT Item; by sign(rp (Z j ,Y|Z s )-r(Z j ,Y)) Determine the candidate feature Z j Under certain conditions, whether its contribution is more significant than that under no conditions; by sign(r p (Z j ,Y|Z s )) Determine the candidate feature Z j whether the additional information contributes positively or negatively to the results;

[0037] S55. Finally, select the candidate feature Z with the highest score according to the scoring formula best , add the feature to the optimal feature variable set S, and update the optimal feature variable set and the remaining feature set variables;

[0038] S56, repeating steps S53 to S55 until the set number of feature subsets n is selected to obtain the optimal feature set S = {S1, S2, ..., S n}.

[0039] Preferably, in S6, the three characteristic variables of unit load, total secondary air volume and total fuel volume are clustered using the kernel fuzzy C-means clustering method based on the differential evolution algorithm to obtain three operating condition ranges; then the optimal characteristic variable set is subdivided according to the characteristics of each operating condition, and the optimal characteristic variable subset is obtained for each operating condition.

[0040] Preferably, in S7, given a data set Where S i is the input feature, Y is the target variable; the goal of the LSSVM prediction model is to find a regression function f(S) to predict the value of Y; the model solution process can be converted into the following optimization problem:

[0041]

[0042] Where: J(ω,e) is the objective function, e=(e1,e2,...,e i ) is the error variable, C is the regularization parameter and C>0, ω is the weight vector, b is the bias constant, is a nonlinear mapping function;

[0043] Use the Lagrangian method to solve the above optimization problem and get the prediction model expression:

[0044]

[0045] Where: K(S,S i ) is the kernel function, λ iis the Lagrangian factor vector. Based on the above principle, the LSSVM prediction model under various working conditions is established.

[0046] The present invention adopts the above method to bring the following beneficial effects:

[0047] The present invention focuses on optimizing the collected data, especially the time-delay reconstruction of the data, effectively eliminating the influence of the delay time on the characteristic relationship between variables; in addition, the correlation, redundancy and complementarity trade-off algorithm is used to process the time-delay reconstruction data, and the data that can characterize NO are screened out. x The optimal characteristic variable set with more comprehensive concentration change characteristics significantly improves the prediction accuracy of the prediction model; the DEKFCM algorithm is used to divide the working conditions and obtain the optimal characteristic variable subset for each working condition, which not only further improves the accuracy of the prediction model, but also helps to improve the prediction efficiency; the LSSVM prediction model is constructed separately on each working condition, so that NO x Concentration prediction is more accurate and efficient, and has higher practical value in actual applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 The power plant boiler NO is reconstructed based on data in the present invention. x Flowchart of the concentration prediction method;

[0049] Figure 2 A flow chart of obtaining a time-delay reconstruction data set for the present invention;

[0050] Figure 3 is a flow chart of the RRCT algorithm of the present invention;

[0051] Figure 4 is a flow chart of the DEKFCM algorithm of the present invention;

[0052] Figure 5 NO before and after data reconstruction of the present invention x Concentration prediction comparison chart. DETAILED DESCRIPTION

[0053] In order to more clearly illustrate the overall concept of the present invention, a detailed description is given below in an exemplary manner in conjunction with the accompanying drawings.

[0054] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Therefore, the protection scope of the present invention is not limited to the specific embodiments disclosed below.

[0055] The various embodiments in this specification are described in a progressive manner, and the same or similar parts between the various embodiments can be referenced to each other, and each embodiment focuses on the differences from other embodiments.

[0056] like Figure 1-5 As shown, the present invention takes a 600MW unit as an example and provides a power plant boiler NO based on data reconstruction. x The concentration prediction method comprises the following steps:

[0057] S1, through NO x The generation mechanism analysis and Spearman correlation analysis of the DCS database data of the 600MW unit were used to obtain the factors affecting NO x The preliminary characteristic variables of concentration are selected, and the preliminary characteristic variable set is obtained. The preliminary characteristic variables include: unit load, total fuel quantity, oxygen quantity, total air volume, total primary air volume, total secondary air volume, main steam pressure, main steam flow, furnace flue gas temperature, AA~FF layer de-cyclone secondary air adjustment damper position feedback, SOFA1~6 layer burnout air adjustment damper position feedback, A~F coal feeder coal feeding rate feedback signal and A~F coal mill inlet primary air volume.

[0058] S2. Perform data cleaning on the preliminary feature variable set screened in S1 to obtain a cleaned feature variable set.

[0059] Specifically, we first check whether there are zero values ​​in the data of all the preliminary feature variable sets. If zero values ​​are detected, it indicates that the data set may correspond to the shutdown state or abnormal situation of the unit. These data have no modeling value and are eliminated. Then, we use the box plot method to identify the abnormal values ​​of the data in the preliminary feature variable set, and replace them with the average value of the data. Finally, we use the sliding average method to reduce the noise of the data set.

[0060] S3. Perform a steady-state discrimination operation on the cleaning feature variable set to obtain a steady-state feature variable set.

[0061] Specifically, the sliding window method is used to determine the steady-state working condition, and the sliding window width is determined for sliding. When the sliding window method is used to determine the steady-state working condition, the optimal window width is determined by comparing the proportion of steady-state data points and non-steady-state data points under different windows. If the window width is too large, some steady-state working condition data will be eliminated. On the contrary, if the window width is too small, non-steady-state working condition data may be included in the result. By calculating the standard deviation of the values ​​in the window, the standard deviation is compared with the allowable fluctuation range of the variable data set of the generator set. Referring to the "Power Plant Boiler Performance Test Procedure", the 600MW unit of this application determines that the allowable fluctuation range of the main steam pressure under steady-state working conditions is ±2%. The change of the main steam pressure is continuously monitored within the set sliding window. If the fluctuation amplitude of the main steam pressure in the window remains within the range of ±2%, the data in the window is determined to be valid data under steady-state working conditions, otherwise it is data under unstable working conditions. In this way, steady-state and non-steady-state data sets can be obtained, and the steady-state data is extracted to obtain a steady-state characteristic variable set.

[0062] S4. Analyze the delay characteristics between variables on the steady-state characteristic variable set extracted in S3, reconstruct each variable data set according to the delay time, and generate a delay reconstruction characteristic variable set (also called delay reconstruction data set). The maximum information coefficient method is used to solve the maximum information coefficient between the variable and the target variable at different lag times. The lag time point corresponding to the maximum value of the maximum information coefficient is the optimal lag time between the variable and the target variable. Delay reconstruction is performed on each variable according to the obtained delay time to eliminate the influence of the delay time on the characteristic relationship between variables.

[0063] like Figure 2 As shown, the specific implementation steps are as follows:

[0064] S41, assuming that the input variable is X = {X1(t), X2(t), ..., X M (t)}; furnace outlet NO x The concentration is Y = Y(t), t is the sampling time, and the empty set of delayed reconstruction data is Z = {Z1, Z2, ..., Z M}. Each input variable is associated with NO x The time delay of concentration is different, and each variable X needs to be N (t), N∈[1,M] performs delay reconstruction.

[0065] S42, assuming that the time delay is τ={τ1,τ2,...,τ g}, for each input variable X N (t) The vector after compensating the time delay is expressed as: X N ={X N (t-τ i ),i=1,2,...,g};

[0066] S43, calculate the maximum information coefficient of each input variable and output variable after embedding time delay: MIC(X N (t-τ d ); Y), τ d ∈τ, where the maximum information coefficient corresponds to the maximum value d The value is the delay time of each input variable;

[0067] S44, according to the delay time τ of each input variable d , perform timing matching, namely: Z N =X N (t-τ d );

[0068] S45, final output delay reconstruction variable set Z * ={Z1,Z2,...,Z M ,Y}.

[0069] S5. Use the correlation, redundancy and complementarity trade-off algorithm (RRCT algorithm for short) to process the delay reconstruction feature variable set and screen out the variables that can characterize NO x The RRCT algorithm quantifies the strength of association between each feature variable and the target variable and between pairs of feature variables through the information theory transformation of the rank correlation coefficient, and uses the partial correlation coefficient to quantify the complementarity between features.

[0070] like Figure 3 As shown, the specific steps are as follows:

[0071] S51, reconstructing the characteristic variable set Z for the time delay * ={Z1,Z2,...,Z M ,Y} calculate the characteristic variable Z i With Z j Between, and Z i The Spearman correlation coefficient between the target variable Y is denoted as ρ ij and r i ,i,j∈[1,M]. The transformation formula of Spearman correlation coefficient and information theory quantity (IT) is:

[0072] IT=-0.5·(1-h 2 ); h = r i or ij (1)

[0073] S52. Use the above transformation formula to convert the correlation coefficient into an information theory quantity, and then we can get the correlation r IT (Z j ,Y) and redundancy r IT (Z i ,Z j ) expression:

[0074]

[0075] The matrix D is generated by formula (2) and formula (3) to intuitively display the data relationship, which can alleviate the need to repeatedly calculate the correlation coefficient in the iterative step. The expression of matrix D is:

[0076]

[0077] In the matrix D, the main diagonal represents the correlation r in the form of information theory. IT (Z j ,Y), the upper triangle area represents the redundancy r IT (Z i ,Z j ).

[0078] S53, select the feature variable with the largest value on the main diagonal of the matrix D as the first optimal feature variable S1 and add it to the optimal feature variable set S. Then, select the candidate feature Z with the largest correlation with the target variable Y under the given selected feature set conditions from the remaining feature set Q. j ,j=1,2,...,M-1, and take out r at the same time IT (Z j ,Z s ), s∈S, calculate the candidate feature Z j The average redundancy of all selected features, that is, Then calculate the new feature Z j In Control Selected Feature Z s The partial correlation coefficient r between the response variable Y p (Z j ,Y|Z s ) and convert it to r p.IT form.

[0079] S54. Define the scoring formula of the RRCT algorithm to evaluate Z j Performance:

[0080]

[0081] In the formula, sign(·) is used to determine whether r should be increased or decreased in the scoring formula. p.IT If the sign (·) is a positive number, +1 is returned, indicating that r should be increased. p.IT Otherwise, it returns -1, indicating that r should be reduced. p.IT Item. By sign(r p (Z j ,Y|Z s )-r(Z j ,Y)) Determine the candidate feature Z j Under certain conditions (i.e., under the condition that there is an optimal feature variable set S), whether its contribution is more significant than that under unconditional conditions; by sign(r p (Z j ,Y|Z s )) Determine the candidate feature Z j Whether the additional information contributes positively or negatively to the results.

[0082] S55. Finally, select the candidate feature Z with the highest score according to the scoring formula best , add the feature to the optimal feature variable set S (also called feature set). Update the optimal feature variable set S = S 上 ∪{Z best}, update the remaining feature set Q, Q = Q 上 -{Z best}.

[0083] S56, repeat steps S53 to S55 until the set number of feature subsets n is selected to obtain the optimal feature variable set S = {S1, S2, ..., S n}.

[0084] S6. The kernel fuzzy c-means clustering algorithm (DEKFCM algorithm for short) based on the differential evolution algorithm is used to classify the operating conditions. The three characteristic variables of unit load, total fuel quantity, and total secondary air volume are clustered using the DEKFCM value to obtain a sample clustering distribution diagram with unit load, total fuel quantity, and total secondary air volume as coordinate variables. The value range of the unit load in each clustering block is stripped out to obtain the corresponding operating condition range. Then, the optimal characteristic variable set is subdivided according to the characteristics of each operating condition to form an optimal characteristic variable subset.

[0085] The KFCM algorithm uses the kernel function to map the data through a nonlinear mapping Φ:X→H:Φ(x)=y, maps the features to a high-dimensional feature space, and divides the data set by the linear function of the high-dimensional feature space.

[0086] The specific steps are as follows:

[0087] S61, for the optimal feature set S = {S1, S2, ..., S n}, assuming the number of clusters c, convergence accuracy ε, randomly set the initial cluster center v0;

[0088] S62,v i (i=1,2,...,c) is the i-th cluster center, μ ij (i=1,2,...,c;j=1,2,...,n) is the membership value of the jth sample belonging to the ith class, and the KFCM objective function is expressed as:

[0089]

[0090] Where: U = {μ ij}; V = {v1, v2, …, v c}; m is the weighted index; |Φ(S j )-Φ(v i )| 2 is Φ(S j ) to Φ(ν i ), K(S j ,v i ) is the Gaussian kernel function; σ is the smoothness parameter.

[0091] S63. The iterative expressions of U and V obtained by Lagrange multiplier method are:

[0092]

[0093] S64, iteratively calculate U and V until the termination condition is met: the difference between the membership matrices of two consecutive generations || U k -U k-1 || is less than the error ε.

[0094] However, the KFCM algorithm is sensitive to the initial cluster center and is prone to fall into local optimality. Differential Evolution (DE) is introduced into the KFCM algorithm. With its fast convergence and powerful global search capabilities, the initial cluster center value of KFCM is optimized to seek better clustering effect.

[0095] For the optimal feature variable set S = {S1, S2, ..., S n}, a cluster center is replaced by a population in the differential evolution algorithm, that is, population N p =V = {v1,v2,…,v c}, assuming that the fitness function f(S j ) is as follows:

[0096]

[0097] From formula (10), we can see that the larger the fitness function value is, the better the clustering effect is.

[0098] like Figure 4 As shown, the specific implementation steps of the DEKFCM clustering algorithm are as follows:

[0099] (1) Determine the number of clusters c = 3 by the silhouette coefficient method, set the error ε, weighting index m, and the number of iterations k = 1;

[0100] (2) Initialize the population N p And set the mutation operator F of the DE algorithm, the crossover probability C R , the maximum number of iterations T max ;

[0101] (3) Starting from t = 1, calculate the kernel matrix and U(t) for each sample;

[0102] (4) Perform differential mutation, crossover, and selection operations, update the population, and update the fitness function value;

[0103] (5) If the fitness function value reaches the maximum or reaches the set maximum number of iterations, stop the iteration and obtain the optimal initial cluster center value, otherwise t = t + 1, go to S3;

[0104] (6) Update the group membership and cluster center according to equations (8) and (9);

[0105] (7) Calculate ||U k -U k-1 ||, if it is less than ε, stop the iteration and output U and V, otherwise k=k+1, go to S6.

[0106] The working range divided according to the above method is:

[0107] Condition 1: 235~351MW;

[0108] Condition 2: 352-462MW;

[0109] Operating condition 3: 463~582MW.

[0110] S7. Establish LSSVM prediction model under various working conditions

[0111] The specific implementation steps are as follows:

[0112] For a given data set Where S i is the input feature, and Y is the target variable. The goal of LSSVM is to find a regression function f(S) to predict the value of Y. The model solving process can be converted into the following optimization problem:

[0113]

[0114] Where: J(ω,e) is the objective function, e=(e1,e2,...,e i ) is the error variable; C is the regularization parameter and C>0. ω is the weight vector; b is the bias constant, is a nonlinear mapping function.

[0115] The Lagrangian method is used to solve the optimization problem of equation (11) and the prediction model expression is obtained:

[0116]

[0117] Where: K(S,S i ) is the kernel function, usually Gaussian radial basis kernel function (RBF), λ i is the Lagrangian factor vector. Based on the above principle, the LSSVM prediction model under various working conditions is established.

[0118] The initial feature variable set (i.e., the initial input variable set), the optimal feature variable set after data reconstruction, and the optimal feature variable subset after data reconstruction are used to construct NO x The concentration prediction model is constructed, and the above variable data sets are tested in the model to generate a prediction effect diagram.

[0119] Depend on Figure 5It can be clearly seen that the model trained based on the optimal feature variable set can track the true value more accurately and better reflect the NO x The concentration trend.

[0120] According to the experimental data in Tables I and II below, compared with the use of the primary feature variable set, the prediction model built based on the optimal feature variable set has significantly lower RMSE and MAE, and the MAPE has also slightly decreased. At the same time, the model running time is shortened from 734 seconds to 245 seconds. After further division into the optimal feature variable subset, the RMSE and MAE of the model are further reduced, and the running time is reduced from 245 seconds to 63 seconds.

[0121] Table I Prediction error before and after data reconstruction

[0122]

[0123] Table II Prediction model running time

[0124]

[0125] The experimental results show that the optimal feature variable set extracted by the RRCT algorithm performs well in improving the prediction accuracy of the model, but there is still room for optimization in terms of running time. This problem is effectively solved by processing the working condition division. Based on the DEKFCM algorithm, three working condition characteristics are divided, and the corresponding optimal feature variable subsets are obtained according to the working condition characteristics. The optimal feature variable subset not only further improves the prediction accuracy of the model, but also significantly shortens the running time of the model, greatly reduces the time cost of model training and prediction, improves efficiency, and makes it have higher practical value in practical applications. It also verifies the feasibility of data reconstruction methods in improving model accuracy and efficiency in complex systems.

[0126] The above specific implementation manner cannot be used as a limitation on the protection scope of the present invention. For those skilled in the art, any substitution, improvement or change made to the implementation manner of the present invention falls within the protection scope of the present invention.

[0127] The matters not described in detail in the present invention are all known technologies to those skilled in the art.

Claims

1. A power plant boiler NO based on data reconstruction x The concentration prediction method is characterized in that The following steps are involved: S1. Obtain the NO from the power plant DCS database X The set of primary characteristic variables of concentration; S2. Perform data cleaning on the preliminary feature variable set to obtain a cleaned feature variable set; S3, performing steady-state discrimination on the cleaning feature variable set based on the sliding window method and standard deviation to obtain a steady-state feature variable set; S4, analyzing the delay characteristics between variables on the steady-state characteristic variable set, and reconstructing the delay of each variable according to the obtained delay time to generate a delay reconstruction characteristic variable set; S5, using a correlation, redundancy and complementarity trade-off algorithm to process the delay reconstruction feature variable set to obtain an optimal feature variable set; S6. Using the kernel fuzzy C-means clustering method based on the differential evolution algorithm to divide the working conditions, and dividing the optimal feature variable set according to the working conditions to obtain the optimal feature variable subsets under different working conditions; S7. Establish LSSVM prediction models under various working conditions based on the optimal feature variable subset, select the corresponding LSSVM prediction model according to the actual working conditions, and obtain NO x Concentration prediction results.

2. According to the preferred data reconstruction method of power plant boiler NO according to claim 1 x The concentration prediction method is characterized in that In S1, according to NO x The Spearman correlation analysis of production mechanism and power plant DCS database data obtained the factors affecting NO x The primary characteristic variables of the concentration are selected, and a set of primary characteristic variables is obtained, wherein the primary characteristic variables include unit load, total fuel quantity, oxygen quantity, total air volume, total primary air volume, total secondary air volume, main steam pressure, main steam flow, furnace smoke temperature, AA-FF layer de-cyclone secondary air adjustment damper position feedback, SOFA1-6 layer burnout air adjustment damper position feedback, A-F coal feeder coal feeding rate feedback signal and A-F coal mill inlet primary air volume.

3. The preferred data-reconstructed power plant boiler NO according to claim 2 x The concentration prediction method is characterized in that In S2, the data cleaning includes outlier processing and noise reduction processing. The outlier processing includes checking whether there is a zero value in the preliminary variable data set; if a zero value is detected, it indicates that the group of data may correspond to the unit shutdown state or abnormal situation, and has no modeling value, so it is deleted; then the box plot method is used to identify outliers, and the average value of the data is used for replacement processing; the sliding average method is used to perform noise reduction processing on the preliminary variable data set.

4. The preferred data-reconstructed power plant boiler NO according to claim 3 x The concentration prediction method is characterized in that In S3, the allowable fluctuation range of the main steam pressure under steady-state conditions is first determined to be ±2%. Then the sliding window method and standard deviation are introduced to continuously monitor the changes in the main steam pressure within the set sliding time window. If the fluctuation amplitude of the main steam pressure within the window remains within the range of ±2%, the data within the window is determined to be valid data under steady-state conditions, and the valid data is extracted to obtain a steady-state characteristic variable set.

5. The preferred data-reconstructed power plant boiler NO according to claim 4 x The concentration prediction method is characterized in that In S3, when the sliding window method is used to determine the steady-state condition, the window width is determined by comparing the proportion of steady-state data points to non-steady-state data points under different windows.

6. The preferred data-reconstructed power plant boiler NO according to claim 5 x The concentration prediction method is characterized in that In S4, the maximum information coefficient method is used to solve the maximum information coefficient between the variable and the target variable at different lag times, where the time point corresponding to the maximum value of the maximum information coefficient is the optimal delay time between the variable and the target variable. According to the obtained delay time, each variable is reconstructed with a time delay to eliminate the influence of the delay time on the characteristic relationship between the variables; The specific steps include: S41. Assume that the input variable is X={X1(t),X2(t),...,X M (t)}; furnace outlet NO x The concentration is Y = Y(t), t is the sampling time, and the empty set of delayed reconstruction data is Z = {Z1, Z2, ..., Z M }; Each input variable is equal to NO x The time delay of concentration is different, and each variable X needs to be N (t), N∈[1,M] performs delay reconstruction; S42, assuming that the time delay is τ={τ1,τ2,...,τ g }, for each input variable X N (t) The vector after compensating the time delay is expressed as: X N ={X N (t-τ i ),i=1,2,...,g}; S43, calculate the maximum information coefficient of each input variable and output variable after embedding time delay: MIC(X N (t-τ d ); Y), τ d ∈τ, where the maximum information coefficient corresponds to the maximum value d The value is the delay time of each input variable; S44, according to the delay time τ of each input variable d , to perform timing matching: Z N =X N (t-τ d ); S45, the final output delay reconstruction feature variable set Z * ={Z1,Z2,...,Z M ,Y}.

7. The preferred data-reconstructed power plant boiler NO according to claim 6 x The concentration prediction method is characterized in that In S5, the correlation, redundancy and complementarity trade-off algorithm is used to process the delay reconstruction feature variable set to screen out the variables that can characterize NO x The optimal characteristic variable set with more comprehensive concentration change characteristics specifically includes the following steps: S51, reconstructing the characteristic variable set Z for the time delay * ={Z1,Z2,...,Z M ,Y} calculate the characteristic variable Z i With Z j Between and Z i The Spearman correlation coefficient between the target variable Y is denoted as ρ ij and r i , i,j∈[1,M]; according to ρ ij and r i The information theoretical quantity IT is obtained as: IT=-0.5·(1-h 2 ) h=r i or ij S52, according to the above formula, we can get the correlation r IT (Z j ,Y) and redundancy r IT (Z i ,Z j ), the specific expression is: The matrix D is generated by the above two equations. The expression of matrix D is: In the matrix D, the main diagonal represents the correlation r IT (Z j ,Y), the upper triangle area represents the redundancy r IT (Z i ,Z j ); S53, add the feature variable with the largest value on the main diagonal of the matrix D as the first optimal feature variable S1 to the optimal feature variable set S; then select the candidate feature Z with the largest correlation with the target variable Y under the given selected feature set from the remaining feature variables Q. j ,j=1,2,...,M-1, and take out r at the same time IT (Z j ,Z s ), s∈S, calculate the candidate feature Z j The average redundancy of all selected features s∈S, that is, Then calculate the new feature Z j In Control Selected Feature Z s The partial correlation coefficient r between the response variable Y p (Z j ,Y|Z s ) and convert it to r p.IT form; S54. Define the scoring formula of the complementarity trade-off algorithm to evaluate Z j Performance: In the formula, sign(·) is used to determine whether r should be increased or decreased in the scoring formula. p.IT If the sign (·) is a positive number, +1 is returned, indicating that r should be increased. p.IT Otherwise, it returns -1, indicating that r should be reduced. p.IT Item; by sign(r p (Z j ,Y|Z s )-r(Z j ,Y)) Determine the candidate feature Z j Under certain conditions, whether its contribution is more significant than that under no conditions; by sign(r p (Z j ,Y|Z s )) Determine the candidate feature Z j whether the additional information contributes positively or negatively to the results; S55. Finally, select the candidate feature Z with the highest score according to the scoring formula best , add the feature to the optimal feature variable set S, and update the optimal feature variable set and the remaining feature variable set; S56, repeat steps S53 to S55 until the set number of feature subsets n is selected to obtain the optimal feature variable set S = {S1, S2, ..., S n }.

8. The preferred data-reconstructed power plant boiler NO according to claim 7 x The concentration prediction method is characterized in that In S6, the kernel fuzzy C-means clustering method based on the differential evolution algorithm is used to cluster the three characteristic variables of unit load, total secondary air volume and total fuel volume to obtain three operating condition ranges; then, the optimal characteristic variable set is subdivided according to the characteristics of each operating condition, and the optimal characteristic variable subset is obtained for each operating condition.

9. The preferred data-reconstructed power plant boiler NO according to claim 8 x The concentration prediction method is characterized in that In S7, given a data set Where S i is the input feature, Y is the target variable; the goal of the LSSVM prediction model is to find a regression function f(S) to predict the value of Y; The model solving process can be converted into the following optimization problem: Where: J(ω,e) is the objective function, e=(e1,e2,...,e i ) is the error variable, C is the regularization parameter and C>0, ω is the weight vector, b is the bias constant, is a nonlinear mapping function; Use the Lagrangian method to solve the above optimization problem and get the prediction model expression: Where: K(S,S i ) is the kernel function, λ i is the Lagrangian factor vector. Based on the above principle, the LSSVM prediction model under various working conditions is established.

Images