A real-time prediction method for oxygen content of flue gas in municipal solid waste incineration process

Abstract

The application relates to the field of data-driven modeling of industrial production process parameters, and discloses a real-time prediction method for oxygen content in flue gas of municipal solid waste incineration process. In view of the problems of high cost, short service life and low precision of the measuring instrument for the oxygen content in flue gas of the municipal solid waste incineration process, the application provides a real-time prediction method for the oxygen content in flue gas based on a deep random configuration network, so that the oxygen content in flue gas can be accurately estimated. The method can realize the prediction model for the oxygen content in flue gas by eliminating samples containing abnormal values and designing a small-scale deep random configuration network, and can realize the prediction algorithm. The method can solve the problem that the measuring instrument does not have a prediction function, can avoid the problem that the deep random configuration network is influenced by the excessively large network scale and the real-time prediction is affected, can reduce the procurement and maintenance costs of the measuring instrument, and is favorable for real-time monitoring of the oxygen content in flue gas.

Description

Technical Field

[0001] This invention relates to the field of data-driven modeling of industrial production process parameters, and more specifically, to a method for real-time prediction of oxygen content in flue gas during urban solid waste incineration. Background Technology

[0002] With the continuous growth of the world's population and the ongoing economic development, the problem of urban solid waste is becoming increasingly prominent. Currently, the mainstream methods for treating solid waste include landfill, composting, and incineration, with landfill accounting for the largest share. However, solid waste that has not undergone harmless treatment can cause irreversible damage to the environment. Compared to other methods, incineration is the optimal way to treat municipal solid waste. Harmful gases can decompose under high-temperature conditions, the heat generated during incineration can be used for power generation, and the stable ash can be directly landfilled. Therefore, incineration is one of the effective means to achieve harmless treatment of urban solid waste. During solid waste incineration, the oxygen content of the flue gas needs to be used to judge the incineration conditions. When the incineration system is not properly controlled, resulting in excessive oxygen content in the flue gas, it will lead to excessive smoke loss; conversely, when the oxygen content in the flue gas is too low, i.e., insufficient airflow, it will cause incomplete combustion of solid waste, reducing thermal efficiency. Therefore, this invention has broad application prospects in the field of solid waste incineration.

[0003] Currently, the oxygen content in flue gas during solid waste incineration is mainly measured using zirconia sensors. However, these sensors are expensive, have a short lifespan, and their accuracy declines over time. Furthermore, they lack predictive capabilities, making it difficult to meet the speed, accuracy, and stability requirements of optimized control in solid waste incineration processes. Neural networks have achieved significant results in parameter prediction in industrial production, particularly the recently developed Deep Stochastic Configuration Network (DSCN). DSCN effectively avoids getting trapped in local optima and can quickly predict parameter changes, thus gaining importance in industrial applications. However, improper settings of the number of hidden layers and hidden nodes in a DSCN can lead to overfitting or insufficient generalization ability. Increasing the number of hidden layers also increases the model's time and structural complexity, affecting real-time performance and speed. Therefore, this invention addresses both the accuracy and complexity of the prediction model. It employs techniques such as removing samples containing outliers and establishing a small-scale DSCN prediction model to achieve accurate and rapid real-time prediction of flue gas oxygen content. This lays the foundation for timely judgment of the operating conditions of solid waste incineration and the realization of optimized operation control. Summary of the Invention

[0004] To address the aforementioned problems, this invention provides a method for real-time prediction of oxygen content in flue gas during urban solid waste incineration. This method provides accurate predicted values ​​for flue gas oxygen content, enabling real-time monitoring of incineration conditions and improving incineration efficiency.

[0005] To achieve the above objectives, the present invention adopts the following technical solution and implementation steps:

[0006] 1. A method for real-time prediction of oxygen content in flue gas during urban solid waste incineration, characterized by the following main steps:

[0007] Step 1: Remove samples containing outliers, specifically including:

[0008] (1) A mechanism analysis was conducted on the waste incineration process, including the following parameters: average primary combustion chamber temperature x1, primary air flow rate x2, secondary air flow rate x3, main steam flow rate x4, furnace negative pressure x5, main steam pressure x6, primary air fan inlet damper opening x7, secondary air fan inlet damper opening x8, economizer outlet flue gas pressure x9, and economizer outlet flue gas temperature x1. 10 As an input variable, the oxygen content in the flue gas is the output variable. The historical dataset is {X,Y}, containing m samples, where X={x1,x2,…,x…}. 10 Let} represent the input variable, and Y represent the oxygen content in the flue gas. Taking input variable X as an example, calculate the mean and standard deviation of each input variable in all samples.

[0009]

[0010] Where, σ j μ is the standard deviation of the j-th input variable; j is the mean of the j-th input variable; n is the total number of input variables, which is 10 here; It is the value of the j-th input variable in the i-th sample; Equation (1) is also applicable to the process of removing outliers from the output variable Y;

[0011] (2) Calculate the absolute error of each input variable value in all samples.

[0012]

[0013] in, It is the absolute error of the j-th input variable value in the i-th sample;

[0014] (3) According to the Raida criterion, for samples that meet the following condition, i.e., input variable values ​​with an absolute error greater than 3 times the standard deviation are outliers and should be removed from the sample.

[0015]

[0016] Step 2: Train a prediction model based on a deep stochastic configuration network (DSCN) using the historical dataset {X,Y} from the waste treatment plant. This includes:

[0017] (1) Considering that increasing the number of hidden nodes and hidden layers will increase the model learning time and may lead to overfitting, and referring to the size of the training dataset, the DSCN parameters are set as follows: maximum number of hidden layers M, 4≤M≤8; maximum number of nodes in the k-th (k=1,2,...,M) hidden layer. Maximum number of configurations T max 50≤T max ≤100; Expected error ε; Node pruning rate threshold a, 0≤a<1; Regularized sequence γ, all elements are greater than 0; The above parameters will be determined by trial and error in subsequent operations; Initialize residual ε0=[y (1) y (2) …y (m) ] T The output matrix of the hidden node Set it to an empty set, and give an empty set Ω to store the constraint value ξ, and an empty set W to store the weight w and bias b of the candidate hidden nodes;

[0018] (2) Configure alternative hidden nodes; in the intervals [-λ,λ] and [-λ,λ] T Internal random configuration T max The weights w and biases b of all candidate hidden nodes in the group, where u = 1, 2, ..., λ is randomly selected from the sequence γ;

[0019] (3) Calculate the output value of each candidate node. The representation method is as follows:

[0020]

[0021] Where g(·) represents the Sigmoid activation function; z (0) = [x1,x2,…,x 10 [] represents the input to the model's input layer, specifically the input variable X from the historical dataset. This represents the input to the remaining hidden layers, where This represents the activation value of the u-th hidden node in the k-th hidden layer. and respectively represent the connection weight matrix and bias matrix between all hidden nodes of the (k-1)-th hidden layer and the candidate node; it should be noted here that, except for the input layer in DSCN, the activation value of the k-th hidden layer is the input variable of the (k+1)-th hidden layer; for convenience of subsequent description, "k-u" is used to represent "the u-th hidden node in the k-th hidden layer";

[0022] (4) Determine candidate nodes: save the parameters w and b of candidate nodes satisfying formula (5) into W. If W is an empty set, that is, there are no candidate nodes satisfying formula (5), return to step (2) until W is not an empty set; then, the constraint value is saved into Ω;

[0023]

[0024] wherein, represents and the inner product of , represents the training residual before configuring k-u, the regularization factor r satisfies 0<r<1, {μ} is a non-negative monotonically decreasing sequence, where μ=(1-r) / (u+1);

[0025] select the maximum value of from Ω, then the candidate node corresponding to this value is used as k-u, and its parameters are recorded as and let the activation matrix update the output weight matrix β * and the residual

[0026]

[0027]

[0028] wherein, represents the Moore-Penrose pseudoinverse of the matrix , represents the training residual after configuring k-u; update

[0029] (5) Determine whether the condition ε0<ε or if the condition is not satisfied, it indicates that the configuration of the k-th hidden layer is not completed, so let u=u+1 to indicate that k-(u+1) will be configured subsequently, and return to step (2);

[0030] (6) Evaluate the importance of each hidden node in the k-th layer: take the mutual information value between the activation value of the hidden node and the output variable as the criterion for evaluating the importance of the hidden node, the importance score s of k-u k,u can be expressed as:

[0031]

[0032] Among the parameters related to ku, This represents the output value of the hidden node, i.e., the activation matrix. H(Y) represents the activation entropy, and H(Y) represents the information entropy of variable Y. Representing variables The joint entropy of variable Y;

[0033] (7) Hidden node importance ranking: Calculate the importance scores of all hidden nodes in the k-th layer and sort them in descending order;

[0034] (8) Delete hidden nodes in the selected k-th hidden layer: Define the pruning rate of the hidden nodes in the k-th hidden layer as pr(k), where 0 ≤ pr(k) < 1. The function pr(k) is defined as:

[0035]

[0036] Where c represents the learning coefficient; from the current network structure, remove the hidden nodes whose scores rank lower than pr(k) in importance score proportion among all hidden nodes in layer k, and retain the hidden nodes ranked first pr(k); specifically, when pr(k) equals 0, it indicates that no pruning operation was performed on the k-th hidden layer; update the relevant network parameters W, β * and make

[0037] (9) Determine whether the condition ε0<ε or k≥M is met; if not, it means that the training termination condition has not been met, and at this time all the hidden nodes of the k hidden layers have been configured. Let k=k+1, u=1 to indicate that the construction of the first hidden node of the (k+1)th hidden layer will be carried out, and return to step (2); if satisfied, the training of the prediction model for flue gas oxygen content is completed.

[0038] (10) Input the test samples into the trained prediction model, record the output value of the prediction model, i.e. the predicted value of the oxygen content in the flue gas, and compare it with the actual value of the measuring instrument to evaluate the performance of the prediction model.

[0039] The inventiveness of this invention is mainly reflected in:

[0040] 1. This invention considers that outliers will reduce the accuracy of flue gas oxygen content prediction. The Raida criterion is used to remove samples containing outliers, which is beneficial for the prediction model to accurately estimate the oxygen content of flue gas.

[0041] 2. This invention employs a mutual information-based layer-by-layer dynamic pruning strategy to reduce the size of the DSCN network, thereby improving the calculation speed of the prediction model and meeting the speed requirements for real-time prediction of flue gas oxygen content.

[0042] 3. The real-time prediction model for flue gas oxygen content established by this invention can solve the problem that measuring instruments do not have prediction functions, reduce the procurement and maintenance costs of measuring instruments, and facilitate the real-time monitoring of flue gas oxygen content. Attached Figure Description

[0043] Figure 1 This is the DSCN topology of the present invention;

[0044] Figure 2 The flowchart illustrates the real-time prediction model for oxygen content in flue gas during urban solid waste incineration, as presented in this invention. Detailed Implementation

[0045] The sample data comes from 1000 data points from a solid waste incineration plant in Beijing. The following is a combination of... Figure 1 and Figure 2 The specific implementation process of the present invention will be further explained.

[0046] A method for real-time prediction of oxygen content in flue gas during urban solid waste incineration, characterized by the following main steps:

[0047] Step 1: Remove samples containing outliers, specifically including:

[0048] (1) A mechanism analysis was conducted on the waste incineration process, including the following parameters: average primary combustion chamber temperature x1, primary air flow rate x2, secondary air flow rate x3, main steam flow rate x4, furnace negative pressure x5, main steam pressure x6, primary air fan inlet damper opening x7, secondary air fan inlet damper opening x8, economizer outlet flue gas pressure x9, and economizer outlet flue gas temperature x1. 10 As an input variable, the oxygen content in the flue gas is the output variable. The historical dataset is {X,Y}, containing m samples, where X={x1,x2,…,x…}. 10 Let} represent the input variable, and Y represent the oxygen content in the flue gas. Taking input variable X as an example, calculate the mean and standard deviation of each input variable in all samples.

[0049]

[0050] Where, σ j μ is the standard deviation of the j-th input variable; j is the mean of the j-th input variable; n is the total number of input variables, which is 10 here; the total number of samples is 1000; It is the value of the j-th input variable in the i-th sample; Equation (1) is also applicable to the process of removing outliers from the output variable Y;

[0051] (2) Calculate the absolute error of each input variable value in all samples.

[0052]

[0053] in, It is the absolute error of the j-th input variable value in the i-th sample;

[0054] (3) According to the Raida criterion, for samples that meet the following condition, i.e., input variable values ​​with an absolute error greater than 3 times the standard deviation are outliers and should be removed from the sample.

[0055]

[0056] Based on the above criteria, a total of 4 samples were found to contain outliers, reducing the sample size from 1000 to 996; 796 of these samples were randomly selected as the training dataset, and 200 samples were selected as the test dataset.

[0057] Step 2: Train a prediction model based on a deep stochastic configuration network (DSCN) using the historical dataset {X,Y} from the waste treatment plant. This includes:

[0058] (1) Considering that increasing the number of hidden nodes and hidden layers will lead to an increase in model learning time and may result in overfitting, and referring to the size of the training dataset, the DSCN parameters are set as follows: maximum number of hidden layers M, which is set to 5 here; maximum number of nodes in the k-th (k=1,2,...,5) hidden layer. Here Set to 100; Maximum number of configurations T max In this T max Set to 100; Expected error ε, here ε is set to 0.001; Node pruning rate threshold a, here a is set to 0.25. The above parameters have been verified multiple times using the validation set; Regularization sequence γ, all elements are greater than 0, here γ = {1, 10, 20, 30, 40, 50, 60, 70, 80, 90, 100}; Initialized residual ε0 = [y (1) y (2) …y (m) ] T The output matrix of the hidden node Set it to an empty set, and give an empty set Ω to store the constraint value ξ, and an empty set W to store the weight w and bias b of the candidate hidden nodes;

[0059] (2) Configure alternative hidden nodes; in the intervals [-λ,λ] and [-λ,λ] TRandomly configure weights of 100 groups of all alternative hidden nodes within and biases where u=1,2,...,100, and λ is randomly selected from sequence Υ;

[0060] (3) Calculate output values of all alternative nodes The representation method is as follows:

[0061]

[0062] wherein, g(·) represents the Sigmoid activation function; z (0) =[x1,x2,…,x 10 is the input of the input layer of the model, that is, the input variable X in the historical dataset, represents the input of the remaining hidden layers, wherein represents the activation value of the u-th hidden node in the k-th hidden layer, and respectively represent the connection weight matrix and bias matrix between all hidden nodes in the k-1-th hidden layer and the alternative node; it should be noted here that in DSCN, except for the input layer, the activation value of the hidden node at the k-th layer is the input variable of the k+1-th hidden layer; for the convenience of subsequent description, "k-u" is used to represent "the u-th hidden node of the k-th hidden layer";

[0063] (4) Determine alternative nodes: save parameters w and b of alternative nodes satisfying formula (5) into W, if W is an empty set, that is, no alternative nodes satisfying formula (5), return to step (2) until W is not an empty set; then, save the constraint value into Ω;

[0064]

[0065] wherein, represents the inner product of and represents the training residual before configuring k-u, the regularization factor r satisfies 0<r<1, herein r is selected from the set {0.9,0.99,0.999,0.9999,0.99999,0.999999}, {μ} is a non-negative monotonically decreasing sequence, wherein μ=(1-r) / (u+1);

[0066] select from Ω the maximum value, then the corresponding alternative node of the value is taken as k-u, and its parameters are recorded as and let update the output weight matrix β * and the residual

[0067]

[0068]

[0069] in, Representation matrix Moore-Penrose pseudo-reverse, Represents the training residual after configuring ku; update

[0070] (5) Determine whether the condition ε0<0.001 or u≥100 is met; if not, it means that the kth hidden layer has not been configured. Let u=u+1 to indicate that k-(u+1) will be configured in the future, and return to step (2).

[0071] (6) Evaluate the importance of each hidden node in layer k: Use the mutual information between the activation value and the output variable of the hidden node as the criterion for judging the importance of the hidden node, and calculate the importance score s of ku. k,u It can be represented as:

[0072]

[0073] Among the parameters related to ku, This represents the output value of the hidden node, i.e., the activation matrix. H(Y) represents the activation entropy, and H(Y) represents the information entropy of variable Y. Representing variables The joint entropy of variable Y;

[0074] (7) Hidden node importance ranking: Calculate the importance scores of all hidden nodes in the k-th layer and sort them in descending order;

[0075] (8) Delete hidden nodes in the selected k-th hidden layer: Define the pruning rate of the hidden nodes in the k-th hidden layer as pr(k), where 0 ≤ pr(k) < 1. The function pr(k) is defined as:

[0076] pr(k) = -0.047(k-3) 2 +0.25 (9)

[0077] Where c represents the learning coefficient, and here c is set to 0.047; Remove the hidden nodes whose scores rank lower than pr(k) in the importance score ratio of all hidden nodes in the k-th layer from the current network structure, and retain the hidden nodes ranked first pr(k); Specifically, when pr(k) equals 0, it indicates that no pruning operation was performed on the k-th hidden layer; Update the relevant network parameters W, β * and make

[0078] (9) Determine whether the condition ε0<0.001 or k≥5 is met; if not, it means that the training termination condition has not been met, and at this time all the hidden nodes of the k hidden layers have been configured. Let k=k+1, u=1 to indicate that the construction of the first hidden node of the (k+1)th hidden layer will be carried out, and return to step (2); if satisfied, the training of the prediction model for flue gas oxygen content is completed.

[0079] (10) Input the test sample into the trained prediction model, record the output value of the prediction model, i.e. the predicted value of the oxygen content in the flue gas, and compare it with the actual value of the measuring instrument to evaluate the performance of the prediction model.

[0080] Test results show that the root mean square error (RMSE) is 0.1865 and the mean absolute error (MAE) is 0.1491, which can achieve accurate estimation of the oxygen content in flue gas. The model training time is 8.3505s, which meets the requirement of rapid prediction of oxygen content in flue gas during solid waste incineration. The implementation of this invention is conducive to real-time monitoring of oxygen content in flue gas, and can lay the foundation for timely judgment of the operating conditions of solid waste incineration process and the realization of optimized operation control.

Claims

1. A method for real-time prediction of oxygen content in flue gas during urban solid waste incineration, characterized in that... Comprising the following steps: Step 1, eliminating samples containing outliers, specifically comprising: (1) A mechanism analysis was conducted on the waste incineration process, including the following parameters: average primary combustion chamber temperature x1, primary air flow rate x2, secondary air flow rate x3, main steam flow rate x4, furnace negative pressure x5, main steam pressure x6, primary air fan inlet damper opening x7, secondary air fan inlet damper opening x8, economizer outlet flue gas pressure x9, and economizer outlet flue gas temperature x1. 10 As an input variable, the oxygen content in the flue gas is the output variable; when the historical dataset is {X,Y}, containing m samples, where X={x1,x2,…,x…} 10 Let} represent the input variable, and Y represent the oxygen content in the flue gas. Taking input variable X as an example, calculate the mean and standard deviation of each input variable in all samples. (1)； Where, σ j μ is the standard deviation of the j-th input variable; j is the mean of the j-th input variable; n is the total number of input variables, where n is 10; x(i)j is the value of the j-th input variable in the i-th sample; Equation (1) is also applicable to the process of removing outliers from the output variable Y; (2) calculating the absolute error of each input variable value in all samples, (2)； wherein e(i) j is the absolute error of the j-th input variable value in the i-th sample; (3) according to the Pauta criterion, for a sample satisfying the following condition, that is, an input variable value whose absolute error is greater than 3 times the standard deviation is an outlier, said sample shall be eliminated, (3)； Step 2, training a prediction model based on deep stochastic configuration networks using the historical data set {X,Y} of a garbage treatment plant, specifically comprising: (1) Considering that increasing the number of hidden nodes and hidden layers will increase the model learning time, and referring to the size of the training dataset, the DSCN parameters are set as follows: maximum number of hidden layers M, 4 ≤ M ≤ 8; maximum number of nodes in the k-th hidden layer L(k) max, where k=1, 2,..., M; 50 ≤ L(k) max ≤ 100; maximum number of configurations T max 50 ≤ T max ≤ 100; Expected error ε; Node pruning rate threshold a, 0 ≤ a < 1; Regularized sequence ϓ, all elements are greater than 0; Initialized residual ε0 = [y (1) y (2) … y (m) ] T The output matrix ℍ of the hidden node is set to an empty set, and an empty set Ω is given to store the constraint value ξ. An empty set W stores the weight w and bias b of the candidate hidden nodes. (2) Configure alternative hidden nodes; in the intervals [-λ, λ] and [-λ, λ] T Internal random configuration T max The weights w and biases b of all candidate hidden nodes in the group, where u = 1, 2, ..., L(k) max, and λ is randomly taken from the sequence ϓ; (3) calculating output values h(k) u of all candidate nodes respectively, and the representation method is as follows: (4)； Where g(·) represents the Sigmoid activation function; z (0) = [x1, x2, …, x 10 [] represents the input to the model's input layer, i.e., the input variables X and z from the historical dataset. (k) = [g(k) 1, g(k) 2, ..., g(k) U] represents the input of the remaining hidden layers, where g(k) U represents the activation value of the u-th hidden node in the k-th hidden layer. and Let represent the connection weight matrix and bias matrix between all hidden nodes and candidate nodes in the (k-1)th hidden layer, respectively. Note that in DSCN, except for the input layer, the activation value of the hidden node in the kth layer is the input variable of the (k+1)th hidden layer. For ease of description later, "ku" will be used to represent "the u-th hidden node in the k-th hidden layer". (4) determining candidate nodes: saving parameters w and b of candidate nodes satisfying formula (5) into W, if W is an empty set, that is, there is no candidate node satisfying formula (5), returning to step (2) until W is not an empty set; then saving constraint values ξ(k) U into Ω; (5)； wherein, <ɛ(k) u-1, h(k) u> represents the inner product of ɛ(k) u-1 and h(k) u, ɛ(k) u-1 represents the training residual before configuration k-u, the regularization factor r satisfies 0<r< 1, {μ} is a non-negative monotonically decreasing sequence, wherein μ = (1 − r) / (u+ 1); The maximum value of ξ(k)U is selected from Ω, and the candidate node corresponding to this value is taken as ku, whose parameter is denoted as . , And h(k)*u; let the activation matrix H := [H,h(k)*u], update the output weight matrix β * and residual ε(k) u: (6)； (7)； Among them, H + Let ε(k) represent the Moore-Penrose pseudoinverse of matrix H, and let ε(k)u represent the training residual after configuring ku; update ε0=ε(k)u; (5) determining whether the condition ε0<ε or u ≥ L(k) max is satisfied; if not, indicating that the configuration of the k-th hidden layer is not completed, then setting u= u+1 to indicate that k-(u+1) will be configured subsequently, and returning to step (2); (6) Evaluate the importance of each hidden node in layer k: Use the mutual information value between the activation value of the hidden node and the true value of the oxygen content in the flue gas as the criterion for judging the importance of the hidden node, and assign an importance score s to ku. k,u It can be represented as: (8)； wherein, among the parameters related to k-u, h(k) u represents the output value of a hidden node, that is, an activation matrix, H(h(k) u) represents activation entropy, H(Y) represents the information entropy of variable Y, and H(h(k) u, Y) represents the joint entropy of variable h(k) u and variable Y; (7) sorting the importance of hidden nodes: calculating the importance scores of all hidden nodes in the k-th layer and sorting them in descending order; (8) deleting the selected hidden nodes in the k-th hidden layer: defining the hidden node pruning rate of the k-th hidden layer as pr(k), and 0≤pr(k) < 1, the function pr(k) is defined as: (9)； Where c represents the learning coefficient; remove the hidden nodes whose scores rank lower than pr(k) in the importance score ratio of all hidden nodes in the k-th layer from the current network structure, and retain the hidden nodes ranked first pr(k); when pr(k) equals 0, it means that no pruning operation was performed in the k-th hidden layer; update the relevant network parameters W, ℍ, and β. * and ε(k) u, let ε0= ε(k) u; (9) determining whether the condition ε0<ε or k ≥ M is satisfied; if not, indicating that the training termination condition is not reached, and all hidden nodes of the k hidden layers have been completely configured at this time, then setting k = k+1 and u=1 to indicate that the first hidden node of the (k+1)-th hidden layer will be constructed subsequently, and returning to step (2); if the condition is satisfied, the training of the prediction model for flue gas oxygen content is completed; (10) inputting a test sample into the trained prediction model, and recording the output value of the prediction model, that is, the predicted value of the flue gas oxygen content.

Images