Thermal power plant boiler NOx emission prediction method based on model fusion

Abstract

The invention relates to the technical field of boiler combustion and pollutant control, in particular to a thermal power plant boiler NOx emission prediction method based on model fusion. According to the method, firstly, a boiler furnace outlet NOx generation mechanism model is established on the basis of a NOx generation mechanism, a De Sote model is adopted, secondly, a thermal power plant boiler furnace outlet NOx generation data driving model is established on the basis of a long-short term memory (LSTM) neural network optimized by a particle swarm algorithm, and then the mechanism model (physical equation) and the data driving model (LSTM) are combined to generate a NOx generation driving model. The black box defect of a data-driven model is made up by utilizing the physical interpretability of the mechanism model, and meanwhile, the static error of the mechanism model is optimized by utilizing the dynamic adaptability of data driving. Compared with a single model, the fusion model has higher robustness and prediction precision under variable working conditions, sparse data or noise interference.

Description

Technical Field

[0001] This invention relates to the field of boiler combustion and pollutant control technology, specifically to a method for predicting NOx emissions from thermal power plant boilers based on model fusion. Background Technology

[0002] At present, a large number of coal-fired power units need to participate in deep peak shaving to ensure the stable operation of the power grid system. However, the rapid increase and decrease of boiler load can easily cause a mismatch between the temperature field and concentration field of the transient combustion process in the furnace and the spatiotemporal process of pollutant removal, resulting in transient exceedance of pollutant emissions. Therefore, it is urgent to develop spatiotemporal dynamic matching technology for pollutant removal.

[0003] The combustion process in power plant boilers is a complex physicochemical process, with some physicochemical changes difficult to express with precise reaction equations. While existing mechanistic models can reflect some chemical kinetic laws, they are primarily applicable to explaining steady-state processes. In actual boiler variable load scenarios, parameters such as furnace temperature field and oxygen concentration distribution fluctuate rapidly, making it impossible for single-mechanistic models to accurately capture transient NOx concentrations under multiple operating conditions. Data-driven methods can directly extract relationships between system parameters from data, offering strong adaptability and high accuracy; their rapid development has provided new solutions for boiler combustion optimization. However, their "black box" nature leads to a lack of physical interpretability, insufficient generalization ability under extreme conditions or sudden changes in coal quality, and an inability to incorporate the physical constraints required for combustion optimization control.

[0004] To address the aforementioned issues, this invention proposes a dynamic NOx prediction method that integrates mechanism and data-driven approaches. By constructing a dual-driven hybrid model of "mechanism-data," the method achieves dynamic decoupling and collaborative prediction of NOx generation and reduction processes under multiple operating conditions in coal-fired boilers. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide a method for predicting NOx emissions from thermal power plant boilers based on model fusion.

[0006] The specific steps of this clearly defined method for predicting NOx emissions from thermal power plant boilers are as follows:

[0007] 1. A NOx formation mechanism model at the boiler furnace outlet is established based on the NOx formation mechanism, using the De Soete model, whose expression is:

[0008]

[0009] In the formula, J （NOx） NOx generation, kg / s; ζ N The percentage of fuel nitrogen conversion is %; N ar The mass fraction of nitrogen received for fuel, %; k vk is the volatile nitrogen release coefficient. c W is the nitrogen emission coefficient of coke. c Coal feed rate, kg / s; V daf The mass fraction of volatile matter on a dry, ash-free basis, is %; R c The value of C is the coke combustion rate, expressed in kg / s. ar The carbon mass fraction received from the fuel is %.

[0010] 2. A data-driven model for NOx generation at the furnace outlet of a thermal power plant boiler was established based on a Long Short-Term Memory (LSTM) neural network optimized by the particle swarm optimization algorithm.

[0011] 3. Based on support vector machines, the NOx generation mechanism model and the NOx generation data-driven model are fused to obtain a NOx emission fusion prediction model. The entire fusion process is as follows:

[0012] Boiler operating parameters are obtained, and the calculated NOx mechanism values ​​at the furnace outlet are obtained based on the constructed NOx formation mechanism model and the NOx formation data-driven model, respectively. Compared with model predictions As a new joint input feature, construct the feature vector X=[ A NOx emission fusion prediction model based on support vector machine (SVM) was established to achieve accurate prediction of NOx emission concentration from boilers.

[0013] The parameters of the support vector machine are as follows: kernel function: radial basis function (RBF); penalty parameter C: 1.0; kernel coefficient γ: 0.01.

[0014] The above technical solution, in step 2, specifically includes the following steps:

[0015] S1: Obtain historical data of the boiler in the thermal power plant under different operating conditions, including the NOx concentration at the SCR inlet and various operating parameters of the boiler. Perform outlier detection and completion on the data, and preprocess the data using wavelet threshold denoising.

[0016] S2: Based on the NOx generation mechanism, the model input parameters are initially screened. Then, the mutual information (MI) statistical method is used to calculate the mutual information value between NOx concentration and each operating parameter. Key influencing factors of NOx generation are comprehensively selected as model input feature variables, and the screened data are normalized.

[0017] The model input feature variables include: load, coal quality parameters (M ad V ad F cad A ad Q net，ad), total coal quantity, total air volume by quality, coal feed rate of each coal mill, primary air flow rate of each coal mill, opening degree of secondary air damper on each floor, opening degree of SOFA damper on each floor, and secondary air temperature at the outlet of air preheater A / B.

[0018] The mutual information (MI) method calculates the nonlinear correlation between variables using the following formula: The larger the mutual information coefficient, the stronger the correlation between variables:

[0019]

[0020] In the formula, I(x,y) is the mutual information coefficient of variables X and Y; p(x,y) is the joint probability density function of X and Y; p(x) is the marginal probability density function of input sample X; and p(y) is the marginal probability density function of input sample Y.

[0021] The normalization process uses the Z-score standardization method, as shown in the following formula:

[0022]

[0023] In the formula, x scaled For the normalized data of variable X; μ x σ is the sample mean; x The standard deviation of the sample.

[0024] S3: Establish a NOx emission prediction model based on a Long Short-Term Memory Neural Network (LSTM), and use the Particle Swarm Optimization algorithm to optimize the number of layers, nodes, learning rate, and batch size of the LSTM neural network to obtain the optimal NOx emission prediction model.

[0025] The entire modeling and optimization process is as follows:

[0026] Initialize the parameters of the particle swarm optimization algorithm, including the inertia weight w, learning factors c1 and c2, and the maximum number of iterations. For each particle, determine the number of layers, nodes, learning rate, and batch size of the corresponding LSTM neural network. Train the model based on historical data from different boiler operating conditions. Calculate the root mean square error (RMSE) as the fitness value, and update the particle state based on the fitness value.

[0027] (1) Adjust the particle velocity according to the velocity update formula:

[0028]

[0029] In the formula, w is the inertia weight; c1 and c2 are learning factors; p besti For the individual's optimal position; g best The position is the global optimum; r1 and r2 are random numbers, x represents the particle position, v represents the particle velocity, i represents the particle index, and k represents the iteration number.

[0030] (2) Adjust the particle position according to the position update formula:

[0031]

[0032] (3) Determine whether the maximum number of iterations has been reached. If so, output the position of the globally optimal particle, i.e. the optimal number of layers, nodes, learning rate and batch size, and determine the final LSTM neural network prediction model. Otherwise, continue to adjust the particle position according to the particle velocity and position update formula, generate a new particle swarm, and continue to iterate and update.

[0033] Existing technologies often rely on a single method for pollutant generation prediction, that is, establishing a pollutant generation prediction model solely based on the pollutant generation mechanism or big data-driven methods, failing to consider the coupling advantages of both. This invention proposes a method for predicting NOx emission concentrations under boiler variable load conditions, based on both mechanism-based and big data-driven modeling.

[0034] The present invention has the following positive effects:

[0035] (1) The mechanistic model (physical equation) is combined with the data-driven model (LSTM). The physical interpretability of the mechanistic model is used to make up for the black box defect of the data-driven model, while the dynamic adaptability of the data-driven model is used to optimize the static error of the mechanistic model. Compared with the single model, the fusion model has higher robustness and prediction accuracy under varying operating conditions, sparse data or noise interference.

[0036] (2) By using the calculated values ​​of the fuel-type NOx formation mechanism equation as input variables and combining them with the predicted values ​​of the data-driven model, a joint input feature is constructed to solve the problem that traditional data-driven models are insufficient in capturing the dynamic characteristics of the mechanism. Feature expansion enhances the model's ability to represent the dynamic correlation of the combustion process. Attached Figure Description

[0037] Figure 1 A schematic flowchart of the prediction method of this invention.

[0038] Figure 2 Distribution of prediction results from a single prediction model.

[0039] Figure 3 Distribution of prediction results from the fusion prediction model. Detailed Implementation

[0040] A dynamic NOx prediction method that integrates mechanism and data-driven approaches includes the following process:

[0041] 1. A model of NOx formation mechanism at the boiler furnace outlet was established based on the NOx formation mechanism, using the fuel nitrogen conversion rate (ζ) as an example. N), volatile matter release coefficient (k) v ) and coke combustion rate (R c The relationship between boiler combustion parameters and NOx generation is analyzed.

[0042] The NOx formation mechanism is based on the De Soete model (see reference DE SOETE G G. Heterogeneous N2O and NO formation from bound nitrogen atoms during coal charcombustion[J / OL]. Symposium (International) on Combustion, 1991, 23(1): 1257-1264. DOI:10.1016 / S0082-0784(06)80388-7), and its expression is:

[0043]

[0044] In the formula, J （NOx） NOx generation, kg / s; ζ N The percentage of fuel nitrogen conversion is %; N ar The mass fraction of nitrogen received for fuel, %; k v k is the volatile nitrogen release coefficient. c W is the nitrogen emission coefficient of coke. c Coal feed rate, kg / s; V daf The mass fraction of volatile matter on a dry, ash-free basis, is %; R c The value of C is the coke combustion rate, expressed in kg / s. ar The carbon mass fraction received from the fuel is %.

[0045] 2. A data-driven model for NOx generation at the boiler furnace outlet was established using a Long Short-Term Memory (LSTM) neural network optimized by the particle swarm optimization algorithm.

[0046] S1: Obtain historical data of the boiler at Hanchuan Power Plant under different operating conditions, including the NOx concentration at the SCR inlet and various operating parameters of the boiler. Perform outlier detection and completion on the data, and preprocess the data using wavelet threshold denoising.

[0047] S2: Based on the NOx generation mechanism, the model input parameters are initially screened. Then, the mutual information (MI) statistical method is used to calculate the mutual information value between NOx concentration and each operating parameter. Key influencing factors of NOx generation are comprehensively selected as model input feature variables, and the screened data are normalized.

[0048] The model input feature variables include: load, coal quality parameters (M ad V ad F cad A ad Q net，ad， These represent moisture, volatile matter, fixed carbon, ash, and lower heating value, respectively; total coal quantity; total mass air volume; coal feed rate of each coal mill; primary air flow rate of each coal mill; secondary air damper opening of each layer; SOFA damper opening of each layer; and secondary air temperature at the outlet of air preheater A / B.

[0049] The mutual information (MI) method calculates the nonlinear correlation between variables using the following formula: The larger the mutual information coefficient, the stronger the correlation between variables:

[0050]

[0051] In the formula, I(x,y) is the mutual information coefficient of variables X and Y; p(x,y) is the joint probability density function of X and Y; p(x) is the marginal probability density function of input sample X; and p(y) is the marginal probability density function of input sample Y.

[0052] The normalization process uses the Z-score standardization method, as shown in the following formula:

[0053]

[0054] In the formula, x scaled For the normalized data of variable X; μ x σ is the sample mean; x The standard deviation of the sample.

[0055] S3: A NOx emission prediction model is established based on a Long Short-Term Memory (LSTM) neural network. The optimal NOx emission prediction model is obtained by optimizing the number of layers, nodes, learning rate, and batch size of the LSTM neural network using a particle swarm optimization algorithm. The entire modeling and optimization process is as follows:

[0056] Initialize the parameters of the particle swarm optimization algorithm, including the inertia weight w, learning factors c1 and c2, and the maximum number of iterations. For each particle, determine the number of layers, nodes, learning rate, and batch size of the corresponding LSTM neural network. Train the model based on historical data from different boiler operating conditions. Calculate the root mean square error (RMSE) as the fitness value, and update the particle state according to the fitness value.

[0057] (1) Adjust the particle velocity according to the velocity update formula:

[0058]

[0059] In the formula, w is the inertia weight; c1 and c2 are learning factors; pbesti is the individual optimal position; gbest is the global optimal position; and r1 and r2 are random numbers.

[0060] (2) Adjust the particle position according to the position update formula:

[0061]

[0062] (3) Determine whether the maximum number of iterations has been reached. If so, output the position of the globally optimal particle, i.e. the optimal number of layers, nodes, learning rate and batch size, and determine the final LSTM neural network prediction model. Otherwise, continue to adjust the particle position according to the particle velocity and position update formula, generate a new particle swarm, and continue to iterate and update.

[0063] The final parameters for the LSTM neural network prediction model are as follows:

[0064] Number of LSTM layers: 2; Number of nodes: 64 for the first layer and 32 for the second layer; Learning rate: 0.001; Batch size: 32.

[0065] 3. Based on support vector machines, the NOx mechanism model and the NOx data-driven model are fused to obtain a NOx emission fusion prediction model. The entire fusion process is as follows:

[0066] Boiler operating parameters were obtained, and the calculated NOx mechanism values ​​at the furnace outlet were obtained based on the constructed NOx formation mechanism model and NOx formation prediction model, respectively. Compared with model predictions As a new joint input feature, construct the feature vector X=[ A NOx emission fusion prediction model based on support vector machine (SVM) was established to achieve accurate prediction of NOx emission concentration from boilers.

[0067] The parameters of the support vector machine are as follows:

[0068] Kernel function: Radial basis function (RBF); Penalty parameter C: 1.0; Kernel function coefficient γ: 0.01.

[0069] Parameters of the fusion prediction model:

[0070]

[0071] The errors of the NOx prediction model training set and test set are respectively as follows: Figure 2 and Figure 3 As shown.

[0072] according to Figure 2 , Figure 3 Comparative analysis of model prediction results:

[0073]

[0074] Compared to a single prediction model, the root mean square error of the fusion prediction model on the training set and the test set decreased by 9.68 mg / m². 3 11.45 mg / m 3 Decreased to 5.62 mg / m³ 3 6.90 mg / m 3 The mean absolute percentage errors decreased from 4.51% and 6.74% to 2.30% and 3.92%, respectively, indicating that the fusion prediction model significantly improved the accuracy and robustness of NOx emission prediction under variable load conditions of thermal power plants through the synergistic prediction of mechanism and data-driven approaches.

Claims

1. A method for predicting NOx emissions from thermal power plant boilers based on model fusion, characterized in that, Specifically, the steps include the following: (1) A model of NOx formation mechanism at the boiler furnace outlet is established based on the NOx formation mechanism. The De Soete model is adopted, and its expression is: ； In the formula, J （NOx） NOx generation, kg / s; ζ N The percentage of fuel nitrogen conversion is %; N ar The mass fraction of nitrogen received for fuel, %; k v k is the volatile nitrogen release coefficient. c W is the nitrogen emission coefficient of coke. c Coal feed rate, kg / s; V daf The mass fraction of volatile matter on an ash-free basis for dried fuel, %; R c The value of C is the coke combustion rate, expressed in kg / s. ar The carbon mass fraction of the fuel received, % . (2) A data-driven model for NOx generation at the furnace outlet of a thermal power plant boiler was established based on a long short-term memory neural network (LSTM) optimized by particle swarm optimization. (3) The NOx generation mechanism model and the NOx generation data-driven model are fused based on support vector machines to obtain the NOx emission fusion prediction model. The fusion process is as follows: Boiler operating parameters are obtained, and the calculated NOx mechanism values ​​at the furnace outlet are obtained based on the constructed NOx formation mechanism model and the NOx formation data-driven model, respectively. Compared with model predictions As a new joint input feature, construct the feature vector X=[ A NOx emission fusion prediction model based on support vector machine (SVM) was established to achieve accurate prediction of NOx emission concentration from boilers. The parameters of the support vector machine are as follows: kernel function: radial basis function; penalty parameter C: 1.0; kernel coefficient γ: 0.

01.

2. The method for predicting NOx emissions from thermal power plant boilers according to claim 1, characterized in that, Step 2 specifically includes the following steps: S1: Obtain historical data of the boiler in the thermal power plant under different operating conditions, including the NOx concentration at the SCR inlet and various operating parameters of the boiler. Perform outlier detection and completion on the data, and preprocess the data using wavelet threshold denoising. S2: Based on the NOx generation mechanism, the model input parameters are initially screened. Then, the mutual information value between NOx concentration and each operating parameter is calculated using the statistical method of mutual information. Key influencing factors of NOx generation are comprehensively selected as model input feature variables, and the screened data are normalized. S3: A NOx emission prediction model is established based on a long short-term memory neural network, and the particle swarm optimization algorithm is used to optimize the number of layers, number of nodes, learning rate and batch size of the LSTM neural network to obtain the optimal NOx emission prediction model.

3. The method for predicting NOx emissions from thermal power plant boilers according to claim 2, characterized in that, In step S2, the model input feature variables include: load, coal quality parameters, total coal quantity, total mass air volume, coal feed rate of each coal mill, primary air flow rate of each coal mill, secondary air damper opening of each layer, SOFA damper opening of each layer, and secondary air temperature at the outlet of air preheater A / B. The mutual information (MI) method calculates the nonlinear correlation between variables using the following formula: The larger the mutual information coefficient, the stronger the correlation between variables. ； In the formula, I(x,y) is the mutual information coefficient of variables X and Y; p(x,y) is the joint probability density function of X and Y; p(x) is the marginal probability density function of input sample X; p(y) is the marginal probability density function of input sample Y. The normalization process uses the Z-score standardization method, as shown in the following formula: ； In the formula, x scaled The normalized data for variable X; μ x σ is the sample mean; x The standard deviation of the sample.

4. The method for predicting NOx emissions from thermal power plant boilers according to claim 2, characterized in that, In step S3, the entire modeling and optimization process is as follows: Initialize the parameters of the particle swarm optimization algorithm, including the inertia weight w, learning factors c1 and c2, and the maximum number of iterations. For each particle, determine the number of layers, nodes, learning rate, and batch size of the corresponding LSTM neural network. Train the model based on historical data of the boiler operating under different conditions. Calculate the root mean square error as the fitness value and update the particle state according to the fitness value. (1) Adjust the particle velocity according to the velocity update formula: ； In the formula, w is the inertia weight; c1 and c2 are learning factors; p besti For the individual's optimal position; g best The global optimal position is represented by r1 and r2, which are random numbers. x represents the particle position, v represents the particle velocity, i represents the particle index, and k represents the iteration number. (2) Adjust the particle position according to the position update formula: ； (3) Determine whether the maximum number of iterations has been reached. If so, output the position of the globally optimal particle, i.e. the optimal number of layers, nodes, learning rate and batch size, and determine the final LSTM neural network prediction model. Otherwise, continue to adjust the particle position according to the particle velocity and position update formula, generate a new particle swarm, and continue to iterate and update.

Images