Soft measurement method and device for outlet NOx concentration of SCR system

Abstract

This invention provides a method and apparatus for soft measurement of NOx concentration at the outlet of an SCR system, belonging to the field of thermal power generation. The method includes: acquiring historical operating data of a coal-fired unit under different load conditions; training a NOx concentration prediction model based on the historical operating data; acquiring real-time operating data of the coal-fired unit; acquiring real-time characteristic variable data based on the real-time operating data; and inputting the real-time characteristic variable data into the NOx concentration prediction model to obtain the NOx concentration prediction result. By using the above technical means, a NOx concentration prediction model is trained from the historical operating data of the coal-fired unit. This model can improve the accuracy of NOx concentration prediction at the outlet of the SCR system, achieve soft measurement, save on labor maintenance and equipment costs, and simultaneously improve denitrification efficiency.

Description

Technical Field

[0001] This invention relates to the field of thermal power generation, and more specifically to a soft measurement method and a soft measurement device for NOx concentration at the outlet of an SCR system. Background Technology

[0002] Against the backdrop of the proposed "dual carbon" goals, achieving more efficient low-NOx emission reduction is even more urgent for thermal power plants. SCR (Selective Catalytic Reduction) denitrification technology, as the most mature NOx removal technology at present, has been widely used both domestically and internationally. SCR denitrification technology essentially involves injecting ammonia and other catalysts into the flue gas upstream of the catalyst, where they react with nitrogen oxides in the flue gas to reduce NOx emissions during the combustion process of thermal power units. However, injecting too little ammonia will not completely react with the NOx in the flue gas, while injecting too much will cause ammonia escape, producing new harmful substances such as ammonium sulfate ((NH4)2SO4) and ammonium bisulfate (NH4HSO4), and easily leading to waste and increased costs. Obtaining accurate NOx outlet concentration in advance allows for adjustment of the appropriate ammonia injection rate, improving denitrification efficiency.

[0003] Currently, thermal power plants commonly use flue gas analyzers to monitor NOx concentrations in real time. However, the detection environment for these analyzers is extremely harsh, requiring them to withstand high temperatures and dust erosion. This necessitates frequent calibration and maintenance, resulting in high labor costs and low efficiency. Furthermore, for large coal-fired boilers, there is a significant delay between changes in operating parameters and changes in NOx concentration, hindering optimized combustion adjustments. Therefore, the thermal power industry places great emphasis on developing soft-sensing methods for NOx concentration, and research into soft-sensing of NOx concentrations incorporating intelligent algorithms is becoming increasingly widespread.

[0004] Currently, many scholars and researchers have applied support vector machines (SVMs) with different improvements to the prediction of NOx concentration at the outlet of SCR systems, realizing soft measurement of NOx concentration and applying them to ammonia injection control systems to form a complete precision ammonia injection equipment.

[0005] Existing intelligent algorithms for predicting NOx concentrations rely on limited optimization methods, resulting in only minor improvements. Support vector machine models for predicting NOx concentrations at SCR system outlets exhibit poor generalization capabilities, only capable of predicting NOx concentrations under a single load condition, making it difficult to achieve accurate predictions for SCR system outlets under various load conditions. Summary of the Invention

[0006] The purpose of this invention is to provide a soft measurement method and device for NOx concentration at the outlet of an SCR system. This method applies a machine learning model to the measurement of NOx at the outlet of an SCR system, replacing the traditional flue gas analyzer for real-time monitoring of NOx concentration, thus achieving soft measurement, reducing the cost of manual maintenance and equipment investment, and improving denitrification efficiency.

[0007] To achieve the above objectives, a first aspect of the present invention provides a soft measurement method for NOx concentration at the outlet of an SCR system, the soft measurement method for NOx concentration at the outlet of an SCR system comprising: Obtain actual historical operating data of coal-fired power units under different load conditions; A NOx concentration prediction model was trained based on actual historical operational data. Obtain real-time data on the actual operation of coal-fired power units; Obtain real-time feature variable data based on actual operational real-time data; Real-time characteristic variable data are input into the NOx concentration prediction model to obtain NOx concentration prediction results.

[0008] Based on the above technical means, a NOx concentration prediction model is trained using historical data from actual operation of coal-fired power units. This model can improve the accuracy of NOx concentration prediction at the outlet of the SCR system, achieve soft measurement, reduce the cost of manual maintenance and equipment funds, and improve denitrification efficiency.

[0009] In this embodiment of the application, a NOx concentration prediction model is trained based on actual historical operating data, including: Historical characteristic variable data are obtained based on actual operational history data; Construct a least-squares support vector machine model; Based on historical characteristic variable data, an improved particle swarm optimization algorithm is used to optimize the least squares support vector machine model to obtain a NOx concentration prediction model.

[0010] Based on the above technical means, historical characteristic variable data with higher correlation to NOx concentration at the SCR system outlet are selected from the actual historical operation data and used to train the model. At the same time, an improved particle swarm optimization algorithm is used to optimize the least squares support vector machine model during the training process to reduce the amount of training computation.

[0011] In this embodiment of the application, historical feature variable data is obtained based on actual operational history data, including: The actual historical data is preprocessed to obtain preprocessed data; Identify historical characteristic variables; Historical feature variable data are extracted from the preprocessed data based on historical feature variables.

[0012] Based on the above technical means, preprocessing the actual historical data can make the extraction of historical feature variables and the training model more accurate. Determining historical feature variables from all feature variables can reduce the computational load of model training while ensuring the accuracy of the model.

[0013] In this embodiment of the application, the actual historical operation data is preprocessed to obtain preprocessed data, including: The Raida criterion is used to filter and delete abnormal data in the actual operation history data; Linear interpolation is used to complete missing data; A filter is used for filtering, smoothing, and noise reduction. The filtered data is then normalized to obtain the preprocessed data.

[0014] Based on the above technical means, by filtering and deleting abnormal data and then completing it, complete normal data can be obtained. Filtering can remove noise from the data, and normalization can eliminate the differences in dimensions between data.

[0015] In this application embodiment, determining historical feature variables includes: The maximum information coefficient method was used to analyze the maximum information coefficients between each characteristic variable and NOx emissions in the preprocessed data. The correlation coefficients between each characteristic variable and NOx emissions in the preprocessed data were calculated using the Pearson correlation coefficient. A predetermined number of feature variables are selected as historical feature variables based on the maximum information coefficient and the correlation coefficient.

[0016] Based on the above technical means, the maximum information coefficient and Pearson correlation coefficient are combined to screen out the feature variables most correlated with the NOx concentration at the SCR system outlet, thereby reducing the amount of training computation while ensuring the accuracy of the model.

[0017] In this embodiment of the application, based on historical feature variable data, an improved particle swarm optimization algorithm is used to optimize the least squares support vector machine model to obtain a NOx concentration prediction model, including: Initialize the particle swarm's position and velocity; Optimize the kernel width and normalization parameters of the least squares support vector machine model; Historical NOx concentration predictions were calculated using the calculated kernel width parameter, normalization parameter, and historical characteristic variable data. The fitness value of each particle is calculated based on the historical predicted NOx concentration and the historical actual NOx concentration. Determine the position of the particle with the best fitness from the fitness values ​​of each particle, and update the historical best solution and the global best solution of the particle swarm. Dynamically adjust the inertia weights and learning factors of the particle swarm optimization algorithm; The particle swarm position and velocity are updated based on the adjusted particle swarm algorithm; Repeat the above steps until the particle swarm optimization algorithm reaches the termination condition, and obtain the optimal kernel width parameter and the optimal normalization parameter of the least squares support vector machine model; By assigning the optimal kernel width parameter and the optimal normalization parameter to the least squares support vector machine model, a NOx concentration prediction model is obtained.

[0018] Based on the above technical means, the particle swarm optimization algorithm is used to optimize the least squares support vector machine model, thereby improving the model training convergence speed. Dynamic inertial weights and learning factors are used to improve the accuracy of the trained model.

[0019] In this embodiment of the application, the fitness value of each particle is calculated based on the predicted historical NOx concentration and the actual historical NOx concentration, including: The fitness value of each particle is calculated using the following formula: Where N is the number of sets of historical feature variable data. The historical NOx concentration of the k-th group of historical characteristic variable data. , where is the predicted historical NOx concentration value for the k-th group of historical characteristic variable data.

[0020] In this embodiment of the application, dynamically adjusting the inertia weights and learning factors of the particle swarm optimization algorithm includes: Adjust the inertia weights according to the following function: ; In the formula, This is the initial inertia weight; The minimum inertia weight is k; k is the current iteration number. Let V be the variance of the fitness values ​​of the previous particle swarm. For a custom minimum value, Let be the fitness value of the current particle in the i-th iteration. This represents the average fitness value of the population at the current iteration. Adjust the learning factor according to the following formula: ; Where w is the inertia weight, The value is , The value is .

[0021] Based on the above technical means, adjusting the inertia weight using a nonlinear decreasing function can prevent the algorithm from getting stuck in local optima and achieve the ideal optimization result. The dynamically changing learning factor can adapt to different evolutionary stages of particles, thereby strengthening the role of the linear decreasing weight.

[0022] In this embodiment of the application, the method further includes: initializing the initial position of the particle swarm using the optimal Latin hypercube sampling method.

[0023] Based on the above technical means, the initial position distribution of the generated particle swarm can be more reasonable, and it can reasonably cover all areas.

[0024] A second aspect of this application provides a soft measurement device for NOx concentration at the outlet of an SCR system, the device comprising: The historical data acquisition unit is used to acquire historical data on the actual operation of coal-fired power units; The model training unit is used to train a NOx concentration prediction model based on actual historical data. The real-time data acquisition unit is used to acquire real-time data on the actual operation of the coal-fired power unit. The real-time feature variable data acquisition unit is used to acquire real-time feature variable data based on actual operational real-time data. The prediction unit is used to input real-time characteristic variable data into the NOx concentration prediction model to obtain the NOx concentration prediction result.

[0025] Based on the above technical means, a NOx concentration prediction model is obtained by training the actual historical operating data of coal-fired power units. This model can improve the accuracy of NOx concentration prediction at the outlet of the SCR system, realize soft measurement, save the cost of manual maintenance and equipment funds, and improve denitrification efficiency.

[0026] The above technical solution applies machine learning models to the measurement of NOx at the outlet of the SCR system, replacing traditional flue gas analyzers for real-time monitoring of NOx concentration. This achieves soft measurement, saving costs on manual maintenance and equipment, while simultaneously improving denitrification efficiency. Furthermore, improvements to the traditional machine learning model for predicting NOx concentration enhance its generalization ability, enabling it to predict NOx concentration under various load conditions.

[0027] Other features and advantages of the embodiments of the present invention will be described in detail in the following detailed description section. Attached Figure Description

[0028] The accompanying drawings are provided to further illustrate embodiments of the present invention and form part of the specification. They are used together with the following detailed description to explain the embodiments of the present invention, but do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a soft measurement method for NOx concentration at the outlet of an SCR system provided by one embodiment of the present invention; Figure 2a This is a schematic diagram of sampling points obtained by a random method provided in one embodiment of the present invention; Figure 2b This is a schematic diagram of sampling points obtained by the optimal Latin hypercube method provided in one embodiment of the present invention; Figure 3 This is a comparison graph of the fitness curves of the improved PSO algorithm and the original PSO algorithm as a function of the number of iterations, provided by one embodiment of the present invention. Figure 4 This is a Pearson correlation heatmap among 13 characteristic variables provided in one embodiment of the present invention; Figure 5a This is a NOx concentration prediction curve for a 200MW load provided by one embodiment of the present invention; Figure 5b This is a NOx concentration prediction curve for a 300MW load provided by one embodiment of the present invention; Figure 5c This is a NOx concentration prediction curve for a 410MW load provided by one embodiment of the present invention; Figure 5d This is a NOx concentration prediction curve for a 560MW load provided by one embodiment of the present invention; Figure 6 This is a block diagram of a soft measurement device for NOx concentration at the outlet of an SCR system provided in one embodiment of the present invention. Detailed Implementation

[0029] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0030] Figure 1 This is a flowchart of a soft measurement method for NOx concentration at the outlet of an SCR system provided by one embodiment of the present invention. Figure 1 As shown, the soft measurement method for NOx concentration at the SCR system outlet includes: S1: Obtain the actual operating history data of the coal-fired power unit under different load states. In this embodiment, the actual operating history data of the coal-fired power unit is obtained from the distributed control system of the coal-fired power unit. The load state of the coal-fired power unit is generally selected according to the load state that corresponds to the most frequent operation of the coal-fired power unit.

[0031] S2: A NOx concentration prediction model is trained based on actual historical operating data.

[0032] In this embodiment of the application, a NOx concentration prediction model is trained based on actual historical operating data, including: S201: Obtain historical characteristic variable data based on actual operating history data. The collected actual operating history data contains multiple characteristic variables, each with a different impact on the NOx concentration at the SCR system outlet. To reduce information redundancy and improve prediction efficiency, it is necessary to extract historical characteristic variable data containing only the key characteristic variables from the actual operating history data containing multiple characteristic variables.

[0033] Specifically, the process begins by preprocessing the historical data to obtain preprocessed data; then, historical feature variables are determined; and finally, historical feature variable data is extracted from the preprocessed data based on these historical feature variables.

[0034] Based on the above technical means, preprocessing the actual historical data can make the extraction of historical feature variables and the training model more accurate. Determining historical feature variables from all feature variables can reduce the computational load of model training while ensuring the accuracy of the model.

[0035] In this embodiment of the application, the actual historical operation data is preprocessed to obtain preprocessed data, including: The Raida criterion is used to filter and delete abnormal data in the actual operation history data in order to obtain normal data.

[0036] Linear interpolation is used to complete the missing data to obtain the complete data.

[0037] A filter is used for filtering, smoothing, and denoising to remove noise from the data. In this embodiment, a Butterworth filter is used for filtering, smoothing, and denoising.

[0038] The filtered data is then normalized to eliminate differences in unit of measurement between data points, resulting in preprocessed data. In this embodiment, the Butterworth-processed dataset is normalized using Min-Max standardization, as follows: ; in, For the new data obtained through normalization, X is the initial data. It is the minimum value of the original data sample. It is the maximum value in the original data sample.

[0039] In this application embodiment, determining historical feature variables includes: The maximum information coefficient (MIC) method was used to analyze the maximum information coefficients between each characteristic variable and NOx emissions in the preprocessed data. The MIC method can analyze the correlation between characteristic variables. Based on mutual information, the MIC can be used to calculate linear and nonlinear correlations between variables. For variables x and y, the formula for calculating the MIC is: ; ; In the formula: I(x , y) For variables x and y Mutual information values ​​between them; p ( x ), p ( y These are the marginal probability distributions; p ( x , y ) represents the joint probability distribution; B ( n It is usually set to the power of 0.6 of the data size n.

[0040] The maximum information coefficient (MIC) method can identify characteristic variables that are strongly correlated with NOx emissions. However, selecting all strongly correlated variables would lead to information redundancy. Since MIC is applicable to both linear and nonlinear correlations, this paper uses the Pearson product-moment correlation coefficient (PCCs) to further refine the selection of characteristic variables.

[0041] The correlation coefficient between each characteristic variable and NOx emissions in the preprocessed data was calculated using the Pearson correlation coefficient.

[0042] For variables x and y The formula for calculating the Pearson correlation coefficient is: ; In the formula: r xy Pearson correlation coefficient; n The sample length; x i and y i All are the actual values ​​of the variables.

[0043] Finally, a preset number of feature variables are determined as historical feature variables based on the maximum information coefficient and the correlation coefficient.

[0044] Based on the above technical means, the maximum information coefficient and Pearson correlation coefficient are combined to determine the feature variable most correlated with the NOx concentration at the outlet of the SCR system, thereby reducing the amount of training computation while ensuring the accuracy of the model.

[0045] S202: Construct a least-squares support vector machine model.

[0046] The general form of the regression function for Least Squares Support Vector Machine (LSSVM) is: ; In the formula, Represents a weight vector. For mapping functions, For paranoia. The LSSVM parameter optimization function is: ; In the formula, Represents the error variable, where Generate dataset , To predict the expected value, To normalize the parameters, Lagrange multipliers are introduced. Then the above formula becomes: ; Solving using the KKT conditions yields: ; eliminate and The solution to the above equation is: ; In the formula, , , ,and ; ; For radial basis kernel functions, Where is the kernel function width. To simplify the calculation, the nonlinear prediction is transformed into: ; In LSSVM, the kernel width parameter Normalization parameters affect the computational efficiency of the model These parameters determine the minimization and smoothing of the adaptation error. High-quality optimization of both parameters is crucial for obtaining more accurate predictions.

[0047] S203: Based on historical characteristic variable data, the least squares support vector machine model is optimized using an improved particle swarm optimization algorithm to obtain the NOx concentration prediction model.

[0048] In this embodiment of the application, based on historical feature variable data, an improved particle swarm optimization algorithm is used to optimize the least squares support vector machine model to obtain a NOx concentration prediction model, including: Initialize the particle swarm's position and velocity; Optimize the kernel width and normalization parameters of the least squares support vector machine model; Historical NOx concentration predictions were calculated using the calculated kernel width parameter, normalization parameter, and historical characteristic variable data. The fitness value of each particle is calculated based on the historical predicted NOx concentration and the historical actual NOx concentration. Determine the position of the particle with the best fitness from the fitness values ​​of each particle, and update the historical best solution and the global best solution of the particle swarm. Dynamically adjust the inertia weights and learning factors of the particle swarm optimization algorithm; The particle swarm position and velocity are updated based on the adjusted particle swarm algorithm; Repeat the above steps until the particle swarm optimization algorithm reaches the termination condition, and obtain the optimal kernel width parameter and the optimal normalization parameter of the least squares support vector machine model; By assigning the optimal kernel width parameter and the optimal normalization parameter to the least squares support vector machine model, a NOx concentration prediction model is obtained.

[0049] In the embodiments of this application, the optimal Latin hypercube sampling method (OLHD) is used to initialize the particle swarm positions during the first optimization calculation.

[0050] For a given design, its distance list and distance from the index list Yes, it can be obtained. Represents the Euclidean distance between two points, where , Represents the distance between design points The subscripts for the segmentation are s, where s is the number of distances between different points. It can be represented as: ; in p It is a positive integer. Thanks to the space-filling characteristics of the optimal Latin hypercube sampling technique, the initial particles can be distributed relatively regularly throughout the design region, such as... Figures 2a-2b As shown, Figure 2a The distribution of experimental points generated by random sampling. Figure 2b The experimental points generated by the optimal Latin hypercube sampling have a more reasonable distribution.

[0051] Based on the above technical means, the particle swarm optimization algorithm is used to optimize the least squares support vector machine model, thereby improving the model training convergence speed. Dynamic inertial weights and learning factors are used to improve the accuracy of the trained model.

[0052] Inertia weights in the PSO algorithm As one of the determining factors for particle velocity updates, it can maintain the particle's inertia, thus tending to expand the search space. And when... When the coefficient is large, the particle swarm optimization exhibits superior global optimization performance. When the coefficients are small, the global optimization performance is weakened, while the local optimization effect is better. Therefore, the global and local optimization capabilities of the PSO algorithm also need to be adjusted by adjusting the inertia weights. To balance, to fix The relationship between the two cannot be balanced.

[0053] Based on the aforementioned technical methods, historical characteristic variable data with a higher correlation to the NOx concentration at the SCR system outlet are selected from actual operational historical data for model training. Simultaneously, an improved particle swarm optimization algorithm is used to optimize the least squares support vector machine model during training, reducing the computational load. In this application, the inertia weights are adjusted according to the following function: ; In the formula, This is the initial inertia weight; The minimum inertia weight is k; k is the current iteration number. Let V be the variance of the fitness values ​​of the previous particle swarm. For a custom minimum value, Let be the fitness value of the current particle in the i-th iteration. This represents the average fitness value of the population at the current iteration. Using a non-linear decreasing function to adjust the inertia weights can prevent the algorithm from getting trapped in local optima and achieve the desired optimization result.

[0054] In traditional particle swarm optimization, the learning factor is generally set in the interval (0, 4). In order to make the learning factor adapt to different evolutionary stages of particles and thus strengthen the effect of linearly decreasing weights, a dynamically changing learning factor is proposed. In this application, the learning factor is adjusted according to the following formula: ; Where w is the inertia weight, The value is , The value is In the early stages of particle swarm evolution, When the value is relatively large, the particle's self-learning ability, i.e., its local optimization ability, is strong. According to the above formula... The value of will gradually decrease to ensure the particles' optimization ability on a global scale and improve the overall search speed. In the later stages of particle swarm evolution, The value of the learning factor gradually increases, at which point the particles possess a stronger global search capability. This change helps enhance the particles' ultimate local optimization ability through information sharing among the group. Therefore, they can search for the global optimum, achieving higher accuracy and efficiency. A dynamically changing learning factor allows it to adapt to different evolutionary stages of the particles, thereby strengthening the effect of the linearly decreasing weights.

[0055] like Figure 3 This graph compares the fitness curves of the improved PSO algorithm and the original PSO algorithm as a function of iterations. In the graph, the solid line represents the improved PSO algorithm, and the dashed line represents the original PSO algorithm. Observation reveals that the dashed line, due to getting trapped in a local optimum, begins to converge around the 30th iteration, with a fitness of 35.1; while the improved PSO algorithm converges around the 150th iteration, with a fitness of 1.2. The improved algorithm avoids the problem of premature convergence of the original PSO algorithm, significantly reduces the fitness, and improves the optimization effect of the algorithm.

[0056] In this embodiment of the application, the fitness value of each particle is calculated based on the predicted historical NOx concentration and the actual historical NOx concentration, including: The fitness value of each particle is calculated using the following formula: Where N is the number of sets of historical feature variable data. The historical NOx concentration of the k-th group of historical characteristic variable data. , where is the predicted historical NOx concentration value for the k-th group of historical characteristic variable data.

[0057] In some embodiments, the trained model needs to be validated. This application uses MAE, RMSE, and MAPE to evaluate the performance of the NOx concentration prediction model, and the expressions are as follows: ; ; ; in, These are measured values. is the model's predicted value, and N is the total number of samples.

[0058] S3: Obtain real-time data on the actual operation of coal-fired power units. The real-time data on actual operation is obtained from the distributed system in real time.

[0059] S4: Obtain real-time feature variable data based on actual operational real-time data. The feature variables contained in the real-time feature variable data are the same as those contained in the historical feature variable data.

[0060] S5: Input the real-time characteristic variable data into the NOx concentration prediction model to obtain the NOx concentration prediction results.

[0061] Based on the above technical means, a NOx concentration prediction model is trained using historical data from actual operation of coal-fired power units. This model can improve the accuracy of NOx concentration prediction at the outlet of the SCR system, achieve soft measurement, save on labor maintenance and equipment costs, and improve denitrification efficiency.

[0062] The method of this application will be explained below in conjunction with practical applications.

[0063] Taking a 600 MW coal-fired power unit of a certain plant as the research object, the actual operation history data of its distributed control system (DCS) was collected. Four most frequently operating load states (200MW, 300MW, 410MW, and 560MW) were selected, and 14 characteristic variables were collected, totaling more than 60,000 data samples. Outliers in the original dataset were filtered and deleted using the Laida criterion, and missing values ​​were filled using linear interpolation. A Butterworth filter was used for filtering, smoothing, and denoising. Considering the differences in the units of measurement between data, the Butterworth-processed dataset was normalized using Min-Max standardization. Finally, 48,000 data sets were obtained for steady-state conditions corresponding to the 14 characteristic variables for each load state (200MW, 300MW, 410MW, and 560MW), i.e., 1200 data sets for each load state, with each data set containing 14 characteristic variables.

[0064] Each data sample contains 13 characteristic variables reflecting the SCR denitrification process, such as load, ammonia injection rate, and total air volume, as well as the NOx concentration at the outlet of the target object. This serves as a database for NOx concentration prediction. Taking a 300MW load as an example, some of the data is shown in Table 1.

[0065] Table 1 Historical Operation Data

[0066] Then, the correlation between the characteristic variables is analyzed using the maximum information coefficient. Taking a 300MW load as an example, Table 2 shows the maximum information coefficients between 14 characteristic variables and NOx emissions.

[0067] Table 2. Maximum information coefficient between characteristic variables and NOx emissions

[0068] As shown in Table 2, 10 characteristic variables, including load value and ammonia injection flow rate, are strongly correlated with NOx emissions. Selecting all strongly correlated variables would lead to information redundancy. Since the MIC (Minimum Injection Micrometry) is applicable to both linear and nonlinear correlations, this paper uses the Pearson correlation coefficient to further filter the characteristic variables. Figure 4 A heatmap of Pearson correlations among 13 characteristic variables.

[0069] from Figure 4 It can be seen that the characteristic variables such as load value, ammonia injection valve opening, ammonia injection rate, inlet NOx concentration, flue gas temperature, main steam temperature, reheat steam temperature, total air volume, denitrification efficiency, and coal feed rate have a strong linear correlation with NOx emission mass concentration. The O2 concentrations at the SCR inlet and outlet have a strong negative linear correlation with the outlet NOx concentration; therefore, the O2 concentrations at the SCR inlet and outlet are not used as characteristic variables for NOx concentration prediction training. The main steam temperature and reheat steam temperature have a very strong positive linear correlation with the load value and inlet flue gas temperature. The correlation coefficient between the ammonia injection valve opening and the ammonia injection rate reaches 0.99. To reduce information redundancy and improve prediction efficiency, the main steam temperature, reheat steam temperature, and ammonia injection valve opening are removed. (Combined with Table 2 and...) Figure 4 Six characteristic variables were determined to be used as training characteristic variables for the prediction model: load value, ammonia injection rate, inlet NOx concentration, flue gas temperature, denitrification efficiency, and coal feed rate.

[0070] Based on the training feature variables of the prediction model, the corresponding historical feature variable data are obtained for model training, and finally a trained NOx concentration prediction model is obtained.

[0071] This application selects three prediction models—LSSVM, PSO-LSSVM, and PrePSO-LSSVM—to predict NOx concentration sequences, and compares and analyzes the prediction results using model evaluation metrics. Simulation results for the three prediction models are shown below. Figures 5a to 5d As shown. Figures 5a to 5d The black solid line represents the actual NOx concentration sequence, the blue and green solid lines represent the predicted sequences of individual LSSVM and PSO-LSSVM models, and the red solid line represents the PrePSO-LSSVM combined model. Combining the prediction curves, it can be seen that the blue and green solid lines fit the NOx concentration sequence much worse than the red solid line, indicating that the proposed PrePSO-LSSVM combined model performs better in predicting NOx concentration sequences.

[0072] To more intuitively compare the prediction accuracy of the three models, the error calculation results of the prediction results of the three models based on the model evaluation indicators MAE, RMSE, and MAPE are shown in Table 3.

[0073] Table 3 Prediction Error (a)

[0074] (b)

[0075] (c)

[0076] (d)

[0077] As shown in Table 3, the PrePSO-LSSVM model reduces RMSE, MAPE, and MAE by nearly 30% compared to other models. While the PSO-LSSVM model improves prediction accuracy compared to the standalone LSSVM model, the PrePSO-LSSVM model further enhances accuracy based on the PSO-LSSVM model.

[0078] The above method uses intelligent algorithms to predict NOx concentrations through soft measurement instead of traditional flue gas analyzers, reducing the cost of detection and saving manpower for equipment maintenance and repair. The model can adapt to NOx concentration prediction under more load conditions, has high generalization, and has a wider range of applications.

[0079] The second aspect of this application provides a soft measurement device for NOx concentration at the outlet of an SCR system, such as... Figure 6 As shown, the device includes: The historical data acquisition unit is used to acquire historical data on the actual operation of coal-fired power units; The model training unit is used to train a NOx concentration prediction model based on actual historical data. The real-time data acquisition unit is used to acquire real-time data on the actual operation of the coal-fired power unit. The real-time feature variable data acquisition unit is used to acquire real-time feature variable data based on actual operational real-time data. The prediction unit is used to input real-time characteristic variable data into the NOx concentration prediction model to obtain the NOx concentration prediction result.

[0080] Based on the above technical means, a NOx concentration prediction model is trained using historical data from actual operation of coal-fired power units. This model can improve the accuracy of NOx concentration prediction at the outlet of the SCR system, achieve soft measurement, save on labor maintenance and equipment costs, and improve denitrification efficiency.

[0081] The above technical solution applies machine learning models to the measurement of NOx at the outlet of the SCR system, replacing traditional flue gas analyzers for real-time monitoring of NOx concentration. This achieves soft measurement, saving costs on manual maintenance and equipment, while simultaneously improving denitrification efficiency. Furthermore, improvements to the traditional machine learning model for predicting NOx concentration enhance its generalization ability, enabling it to predict NOx concentration under various load conditions.

[0082] Those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by a program instructing related hardware. This program is stored in a storage medium and includes several instructions to cause a microcontroller, chip, or processor to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as a USB flash drive, a portable hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0083] The optional embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the embodiments of the present invention are not limited to the specific details described above. Within the scope of the technical concept of the embodiments of the present invention, various simple modifications can be made to the technical solutions of the embodiments of the present invention, and these simple modifications all fall within the protection scope of the embodiments of the present invention. It should also be noted that the various specific technical features described in the above specific embodiments can be combined in any suitable manner without contradiction. To avoid unnecessary repetition, the embodiments of the present invention will not further describe the various possible combinations.

[0084] Furthermore, various different embodiments of the present invention can be combined in any way, as long as they do not violate the spirit of the embodiments of the present invention, they should also be regarded as the content disclosed by the embodiments of the present invention.

Claims

1. A soft measurement method for NOx concentration at the outlet of an SCR system, characterized in that, The soft measurement method for NOx concentration at the SCR system outlet includes: Obtain actual historical operating data of coal-fired power units under different load conditions; A NOx concentration prediction model was trained based on actual historical operational data. Obtain real-time data on the actual operation of coal-fired power units; Obtain real-time feature variable data based on actual operational real-time data; Input real-time characteristic variable data into the NOx concentration prediction model to obtain NOx concentration prediction results; The NOx concentration prediction model was trained based on actual historical operating data, including: Historical characteristic variable data are obtained based on actual operational history data; Construct a least-squares support vector machine model; Based on historical characteristic variable data, an improved particle swarm optimization algorithm is used to optimize the least squares support vector machine model to obtain a NOx concentration prediction model. Based on historical characteristic variable data, an improved particle swarm optimization algorithm is used to optimize the least squares support vector machine model, resulting in a NOx concentration prediction model, including: Initialize the particle swarm's position and velocity; Optimize the kernel width and normalization parameters of the least squares support vector machine model; Historical NOx concentration predictions were calculated using the calculated kernel width parameter, normalization parameter, and historical characteristic variable data. The fitness value of each particle is calculated based on historical predicted NOx concentrations and historical actual NOx concentrations, using the following formula: Where N is the number of sets of historical feature variable data. The historical NOx concentration of the k-th group of historical characteristic variable data. For the historical NOx concentration prediction value of the kth group of historical characteristic variable data; Determine the position of the particle with the best fitness from the fitness values ​​of each particle, and update the historical best solution and the global best solution of the particle swarm. Dynamically adjust the inertia weights and learning factors of the particle swarm optimization algorithm, including: Adjust the inertia weights according to the following function: ； In the formula, This is the initial inertia weight; The minimum inertia weight is k; k is the current iteration number. Let V be the variance of the fitness values ​​of the previous particle swarm. For a custom minimum value, For the current particle in the th i The fitness value of the next iteration. This represents the average fitness value of the population at the current iteration. Adjust the learning factor according to the following formula: ； in, w For inertial weights, The value is , The value is ; The particle swarm position and velocity are updated based on the adjusted particle swarm algorithm; Repeat the above steps until the particle swarm optimization algorithm reaches the termination condition, and obtain the optimal kernel width parameter and the optimal normalization parameter of the least squares support vector machine model; By assigning the optimal kernel width parameter and the optimal normalization parameter to the least squares support vector machine model, a NOx concentration prediction model is obtained.

2. The soft measurement method for NOx concentration at the outlet of an SCR system according to claim 1, characterized in that, Historical characteristic variable data are obtained based on actual operational history data, including: The actual historical data is preprocessed to obtain preprocessed data; Identify historical characteristic variables; Historical feature variable data are extracted from the preprocessed data based on historical feature variables.

3. The soft measurement method for NOx concentration at the outlet of an SCR system according to claim 2, characterized in that, Preprocessing of historical operational data yields preprocessed data, including: The Raida criterion is used to filter and delete abnormal data in the actual operation history data; Linear interpolation is used to complete missing data; A filter is used for filtering, smoothing, and noise reduction. The filtered data is then normalized to obtain the preprocessed data.

4. The soft measurement method for NOx concentration at the outlet of an SCR system according to claim 2, characterized in that, Identify historical characteristic variables, including: The maximum information coefficient method was used to analyze the maximum information coefficients between each characteristic variable and NOx emissions in the preprocessed data. The correlation coefficients between each characteristic variable and NOx emissions in the preprocessed data were calculated using the Pearson correlation coefficient. A predetermined number of feature variables are selected as historical feature variables based on the maximum information coefficient and the correlation coefficient.

5. The soft measurement method for NOx concentration at the outlet of an SCR system according to claim 1, characterized in that, The method further includes: initializing the initial position of the particle swarm using the optimal Latin hypercube sampling method.

6. A soft-sensor device for measuring NOx concentration at the outlet of an SCR system, characterized in that, The device includes: The historical data acquisition unit is used to acquire historical data on the actual operation of coal-fired power units; The model training unit is used to train a NOx concentration prediction model based on actual historical data. The real-time data acquisition unit is used to acquire real-time data on the actual operation of the coal-fired power unit. The real-time feature variable data acquisition unit is used to acquire real-time feature variable data based on actual operational real-time data. The prediction unit is used to input real-time characteristic variable data into the NOx concentration prediction model to obtain the NOx concentration prediction result. The NOx concentration prediction model was trained based on actual historical operating data, including: Historical characteristic variable data are obtained based on actual operational history data; Construct a least-squares support vector machine model; Based on historical characteristic variable data, an improved particle swarm optimization algorithm is used to optimize the least squares support vector machine model to obtain a NOx concentration prediction model. Based on historical characteristic variable data, an improved particle swarm optimization algorithm is used to optimize the least squares support vector machine model, resulting in a NOx concentration prediction model, including: Initialize the particle swarm's position and velocity; Optimize the kernel width and normalization parameters of the least squares support vector machine model; Historical NOx concentration predictions were calculated using the calculated kernel width parameter, normalization parameter, and historical characteristic variable data. The fitness value of each particle is calculated based on historical predicted NOx concentrations and historical actual NOx concentrations, using the following formula: Where N is the number of sets of historical feature variable data. The historical NOx concentration of the k-th group of historical characteristic variable data. For the historical NOx concentration prediction value of the kth group of historical characteristic variable data; Determine the position of the particle with the best fitness from the fitness values ​​of each particle, and update the historical best solution and the global best solution of the particle swarm. Dynamically adjust the inertia weights and learning factors of the particle swarm optimization algorithm, including: Adjust the inertia weights according to the following function: ； In the formula, This is the initial inertia weight; The minimum inertia weight is k; k is the current iteration number. Let V be the variance of the fitness values ​​of the previous particle swarm. For a custom minimum value, For the current particle in the th i The fitness value of the next iteration. This represents the average fitness value of the population at the current iteration. Adjust the learning factor according to the following formula: ； in, w For inertial weights, The value is , The value is ; The particle swarm position and velocity are updated based on the adjusted particle swarm algorithm; Repeat the above steps until the particle swarm optimization algorithm reaches the termination condition, and obtain the optimal kernel width parameter and the optimal normalization parameter of the least squares support vector machine model; By assigning the optimal kernel width parameter and the optimal normalization parameter to the least squares support vector machine model, a NOx concentration prediction model is obtained.

Images