SNCR outlet NOx concentration prediction method and system based on random forest

Abstract

The invention discloses an SNCR outlet NOx concentration prediction method and system based on a random forest, and the method comprises the steps: S1, obtaining the operation data of an SNCR system, selecting a plurality of input variables affecting the outlet NOx concentration, and carrying out the normalization processing of the input variables; s2, carrying out maximum information coefficient analysis on the normalized data set, determining the delay time of each input variable relative to the NOx concentration of the outlet, and reconstructing the data set according to the delay time; s3, dividing the reconstructed data set into a training set and a test set, constructing a random forest model, and initializing model parameters; s4, optimizing parameters of the random forest model to obtain an optimal parameter combination; and S5, predicting by using the random forest model of the optimal parameter combination to obtain a predicted value of the NOx concentration at the outlet of the SNCR system. The method has the advantages of high prediction precision and the like.

Description

Technical Field

[0001] This invention mainly relates to the field of waste incineration technology, specifically to a method and system for predicting NOx concentration at SNCR outlets based on random forests. Background Technology

[0002] To provide early warning of NOx concentration at the outlet of an SNCR (Selective Non-Catalytic Reduction) flue gas denitrification system, accurate modeling of NOx emissions from the SNCR system was performed. This involved proactive adjustment of ammonia injection and optimization of combustion to achieve a NOx concentration of 80 mg / m³. 3 The following is a summary of recent research findings on NOx emission modeling by numerous experts and scholars both domestically and internationally, highlighting significant progress made in practical applications. Currently, the most commonly used modeling methods are mechanistic modeling and data modeling. Compared to mechanistic modeling, data modeling does not require in-depth understanding of the object's mechanistic characteristics.

[0003] Random forests can evaluate the importance of a large number of input variables; when building a forest, they can assess internal induction errors without bias; they can predict and maintain good accuracy even with missing data; errors can be offset in imbalanced classification datasets; calculating correlations in each instance helps with data mining, outlier detection, and data visualization, and can also detect biases and analyze data. Due to these advantages, random forest algorithms have been widely used in recent years in various industries such as the stock market, pharmaceuticals, and e-commerce. However, the application of random forest algorithms in power generation companies is very limited, especially for predicting NOx mass concentration at the SNCR reaction outlet.

[0004] A single random forest algorithm is not perfect. Some classification or regression problems in random forests are very noisy. For data with different attribute levels, the data will be biased if there are many level divisions. Therefore, the attributes generated from these data in random forests are unreliable. Summary of the Invention

[0005] To address the technical problems existing in the prior art, this invention provides a method and system for predicting SNCR outlet NOx concentration with high prediction accuracy based on random forest.

[0006] To solve the above-mentioned technical problems, the technical solution proposed by this invention is as follows: A method for predicting NOx concentration at SNCR outlets based on random forests includes the following steps: S1. Obtain the operating data of the SNCR system, select multiple input variables that affect the outlet NOx concentration, normalize the input variables, and obtain the normalized dataset; S2. Perform maximum information coefficient analysis on the normalized dataset to determine the delay time of each input variable relative to the outlet NOx concentration, and reconstruct the dataset according to the delay time to obtain a time-aligned reconstructed dataset. S3. Divide the reconstructed dataset into a training set and a test set, construct a random forest model, and initialize the model parameters; S4. Optimize the parameters of the random forest model to obtain the optimal parameter combination; S5. Using the random forest model with the optimal parameter combination, predict the NOx concentration at the outlet of the SNCR system, and perform inverse normalization on the prediction results to output the final prediction results.

[0007] Preferably, in step S1, the multiple input variables include inlet flue gas velocity, outlet flue gas velocity, outlet NH3 concentration, inlet O2 concentration, outlet O2 concentration, inlet flue gas temperature, outlet flue gas temperature, inlet NOx concentration, ammonia leak detection value, unit load, and cumulative fuel quantity.

[0008] Preferably, in step S1, the normalization formula is:

[0009] in It is the data value corresponding to the normalized variable x; This refers to the maximum value of a variable across all samples in the entire dataset; This refers to the minimum value of the variable across all samples in the entire dataset.

[0010] Preferably, the specific process of the maximum information coefficient (MIC) analysis in step S2 is as follows: S201. Divide the two-dimensional variables x and y into m and n different regions, discretize the variables, construct an m×n grid Q, and use the number of samples in the grid and the proportion of the number of samples in the region to the total number of samples to obtain the mutual information under the grid.

[0011] in Let x and y be the mutual information values ​​of variables x and y when discretized in grid Q; p(x, y) Let be the joint probability density function. p(x) and p(y) The edge density function; S202. Select the maximum mutual information under different partitioning patterns and express it as:

[0012] S203. Establish m×n grids of varying sizes, calculate the maximum mutual information value for each grid in sequence, and normalize it to obtain the final MIC value. The specific calculation method is as follows:

[0013] Where m×n represents the total number of grid divisions, and the B value is required to be the power of 0.6 of the total data.

[0014] Preferably, in step S3, the model parameters include the number of decision trees, the maximum depth of the decision trees, the minimum number of samples required for the child nodes to split downwards, the minimum number of samples for the leaf nodes, and the number of features to be considered when finding the optimal cut.

[0015] Preferably, in step S4, the parameters of the random forest model are optimized using a grid search method with cross-validation.

[0016] Preferably, the parameter range used in the grid search method is: Number of decision trees: 600; Maximum depth of decision trees: 25; Minimum number of samples required for a child node to split downwards: 2; Minimum number of samples for a leaf node: 1; Number of features to consider when finding the best cut: 0.8.

[0017] Preferably, in step S5, the model prediction performance is evaluated by mean square error, root mean square error, and mean absolute percentage error.

[0018] The present invention also discloses a computer-readable storage medium having a computer program stored thereon, the computer program performing the steps of the method described above when run by a processor.

[0019] The present invention further discloses a random forest-based SNCR outlet NOx concentration prediction system, comprising an interconnected memory and a processor, wherein the memory stores a computer program, and the computer program executes the steps of the method described above when run by the processor.

[0020] Compared with the prior art, the advantages of the present invention are as follows: This invention effectively solves the problem of NOx concentration prediction bias in SNCR systems caused by reaction delay and measurement lag through five steps: data preprocessing, MIC delay time analysis, random forest modeling, parameter optimization, and concentration prediction. By introducing MIC analysis, the nonlinear time-delay relationship between each input variable and the outlet NOx is accurately quantified, and the dataset is reconstructed accordingly, greatly improving data timeliness and model input accuracy. Furthermore, the random forest model is optimized by combining a grid search algorithm with cross-validation, significantly improving prediction accuracy and generalization ability. Attached Figure Description

[0021] Figure 1 This is a flowchart of the NOx concentration prediction method of the present invention in an embodiment.

[0022] Figure 2 This is a comparison chart of the actual NOx concentration values ​​and the prediction results corresponding to the six sets of model parameters in this invention.

[0023] Figure 3 This is a comparison chart of the actual NOx concentration values ​​and the prediction results corresponding to the optimal model parameters in this invention. Detailed Implementation

[0024] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0025] like Figure 1 As shown in the figure, the SNCR exit NOx concentration prediction method based on random forest provided in this embodiment of the invention includes the following steps: S1. Data preprocessing: Obtain the operating data of the SNCR system, select multiple input variables that affect the outlet NOx concentration, normalize the input variables, and obtain the normalized dataset; Using the SNCR outlet NOx concentration as the output variable of the prediction model, 11 variables affecting the output were selected as candidate input variables through mechanistic analysis. Due to differences in data specifications and structures among different application systems, data inconsistencies arise, making it difficult to obtain accurate data during collection, or resulting in missing values ​​due to human error during system operation. Therefore, data preprocessing is necessary.

[0026] Because differences in data between variables can affect model training, data normalization is performed to unify the variable range to (0, 1). This not only overcomes the adverse effects of singular sampling but also accelerates the gradient descent algorithm's search for the optimal solution, thereby improving the algorithm's accuracy. The calculation formula is as follows: (1) in It is the data value corresponding to the normalized variable x; This refers to the maximum value of a certain variable (such as inlet flue gas velocity, outlet NOx, etc.) across all samples in the entire dataset; This refers to the minimum value of the variable across all samples in the entire dataset. Partially normalized data is shown in Table 1.

[0027] Table 1. Normalized data

[0028] S2. Delay time analysis based on MIC: Perform maximum information coefficient (MIC) analysis on the normalized dataset to determine the delay time of each input variable relative to the outlet NOx concentration, and reconstruct the dataset based on the delay time to obtain a time-aligned reconstructed dataset; Because of the lag time in NOx concentration measurement at the outlet and the relatively long delay time in the denitrification reaction and adjustment of SNCR, the SNCR system exhibits a significant delay characteristic. The lag time of each variable in the original data is obtained using the maximum information coefficient (MIC), and the nonlinear relationship between the reconstructed data and the outlet NOx concentration is studied to reduce the impact of the lag time on the modeling accuracy.

[0029] The Maximum Information Coefficient (MIC) is proposed based on mutual information and can be used to assess linear or nonlinear correlations between variables. The MIC is calculated for two-dimensional variables x and y using the following method: S201. Divide x and y into m and n different regions, discretize the variables, construct an m×n grid Q, and use the number of samples in the grid and the proportion of the number of samples in the region to the total number of samples to obtain the mutual information under the grid.

[0030] (2) in Let x and y be the mutual information values ​​of variables x and y when discretized in grid Q; p(x, y) Let be the joint probability density function. p(x) and p(y) This is the edge density function.

[0031] S202. Since discretization methods may differ at the same scale, the maximum mutual information under different partitioning patterns is selected and expressed as: (3) S203. Establish m×n grids of varying sizes. Calculate the maximum mutual information value for each grid according to equations (2) and (3) in sequence, and normalize it to obtain the final MIC value. The specific calculation method is as follows: (4) Where m×n represents the total number of grid divisions, and the B value is required to be the power of 0.6 of the total data.

[0032] Based on field experience, the maximum delay time of variables affecting the operation of the SCR system was limited to within 10 minutes. The outlet NOx concentration was set as the output variable, and other selected variables were used as input variables. The original data was recorded at 5-second intervals; now, 2000 sets of data were collected at 30-second intervals, and the minimum MIC (Minimum Interval) value between each variable and the outlet NOx was calculated. The delay time corresponding to the maximum MIC was taken as the delay time of the corresponding input variable. The delay time between each input variable and the outlet NOx and the corresponding maximum MIC value are shown in Table 2.

[0033] Table 2 Delay time and MIC value of input variables

[0034] Table 2 shows that the variables in the SNCR system have a larger delay range than other parameters of the unit, with the largest delays observed at the inlet and outlet of the SCNR zone for Unit 1: O2, inlet flue gas temperature, and inlet NOx. Based on Table 2, the 11 input variables were shifted relative to the outlet NOx to overcome model inaccuracies caused by the delay. Specifically, the inlet flue gas velocity in the SNCR zone of Unit 1 was shifted downwards by 7 data points, the outlet flue gas velocity by 13 data points, the outlet NH3 by 8 data points, the inlet O2 by 19 data points, the outlet O2 by 20 data points, the inlet flue gas temperature by 20 data points, the outlet flue gas temperature by 11 data points, the inlet NOx by 19 data points, the ammonia leak detection in the SNCR zone of Unit 1 by 6 data points, the unit load by 9 data points, and the cumulative total fuel by 12 data points. This completed the reconstruction of the dataset.

[0035] S3. Establish an export NOx concentration prediction model: Divide the reconstructed dataset into a training set and a test set, construct a random forest model, and initialize the model parameters; Modeling process: To establish an export NOx concentration prediction model, first select appropriate variables from the original data and perform data preprocessing, namely, handling missing data values ​​and normalization. Secondly, 2000 sets of data were selected for MIC-based delay time analysis. The data was reconstructed for different variables to avoid interference of delay time with model accuracy. Then, the dataset is divided, different parameters are selected in the random forest, the model is trained, the model is predicted, the model is evaluated, and the prediction results are visualized. The data is then denormalized to obtain the NOx concentration at the outlet, and the impact of different parameters on the prediction results in the random forest is compared. The flowchart for predicting NOx concentration using Python is as follows: Figure 1 As shown.

[0036] We selected 2000 sets of data after data processing and, based on the idea of ​​random forest modeling, divided the dataset into 8:2 sets to obtain 1600 sets of training data and 400 sets of test data.

[0037] The constructed RF model is trained using the training set data. Five parameters are selected: the number of decision trees (n_estimators), the maximum depth of the decision tree (max_depth), the minimum number of samples required for a child node to split down (min_samples_split), the minimum number of samples for a leaf node (min_samples_leaf), and the number of features to consider when finding the best cut (max_features).

[0038] S4. Parameter Optimization Finally, in the established random forest model, a grid search method with cross-validation is used to optimize the specified parameters, obtain the optimal parameters, obtain the best-performing prediction model, and obtain the prediction value that is closest to the true value.

[0039] Six different parameter sets were created by selecting different parameters while keeping other parameters unchanged. For example, the modeling parameters selected for number A were: n_estimators=100, max_depth=10, min_samples_split=2, min_samples_leaf=1, max_features=1. The six selected parameter sets are summarized in Table 3.

[0040] Table 3 RF Training Model Parameters

[0041] Grid Search with Cross Verification (GSC) is an algorithm that, given parameter values, sorts and merges the probabilities of each parameter, lists all possible combinations to generate a "grid," and then optimizes the parameters using cross-validation to obtain the best performance. Using GSC with cross-validation, the parameters of the established Random Forest model are optimized, resulting in a better trained model. The obtained RF model parameters are shown in Table 4.

[0042] Table 4 Optimal parameters for the RF training model

[0043] Finally, three evaluation metrics were used to measure the model's predictive performance on NOx emissions: Mean Square Error (MSE), Root Mean Square Error (RMSE), and Mean Absolute Percentage Error (MAPE). The formulas are as follows: (5) In the formula Indicates the predicted value. represents the true value; n represents the number of samples.

[0044] The root mean square error (RMSE), also known as the standard error, effectively reflects the deviation between the actual and predicted values ​​of the established model. Its calculation principle is that in the test set data, as the average distance from the model decreases, the model's accuracy improves. The calculation formula is: (6) Mean Absolute Percentage Error (MAPE) represents the absolute error of the current model. A smaller MAPE indicates a smaller loss and a better model. The calculation formula is: (7).

[0045] S5. NOx Concentration Prediction The random forest model with the optimal parameter combination is used to predict the test set to obtain the predicted value of the NOx concentration at the outlet. The prediction results are then inversely normalized to output the final prediction result.

[0046] This invention effectively solves the problem of NOx concentration prediction bias in SNCR systems caused by reaction delay and measurement lag through five steps: data preprocessing, MIC delay time analysis, random forest modeling, parameter optimization, and concentration prediction. By introducing MIC analysis, the nonlinear time-delay relationship between each input variable and the outlet NOx is accurately quantified, and the dataset is reconstructed accordingly, greatly improving data timeliness and model input accuracy. Furthermore, the random forest model is optimized by combining a grid search algorithm with cross-validation, significantly improving prediction accuracy and generalization ability.

[0047] Analysis of RF prediction model results: Different models were built in the random forest to train and predict the model, resulting in different predictions. Table 5 summarizes the data for the true values, predictions with different parameters, and the predictions for the optimal parameters. The prediction results are visualized, showing the true values ​​and the predictions for the six sets of parameters. Figure 2 As shown. The predicted values ​​obtained from the RF model established using the true values ​​and optimal parameters are as follows. Figure 3As shown in Table 6, the model evaluation indices corresponding to the six sets of parameters and the optimal parameters are summarized in Table 6.

[0048] Table 5 Comparison of Actual Values ​​and Predicted Values

[0049] Table 6. Performance Comparison of RF Models with Different Parameters

[0050] Depend on Figures 2-3 As shown in Tables 5 and 6, the selection of random forest parameters has a significant impact on the model's prediction accuracy; different parameter combinations lead to different prediction results. Tests have demonstrated that three parameters have a significant impact on the prediction results: the number of decision trees, the maximum depth of the decision trees, and the number of features considered when finding the optimal cut. Theoretically, a larger number of decision trees is better, but this also increases computation time. Therefore, the number of trees with the best prediction performance is limited. As shown in Table 4, the optimal number of decision trees is 600, which is the most suitable number for this dataset. The maximum depth of the decision trees reflects the complexity of a single tree; a larger depth results in a better fit, but it also slows down computation and can lead to overfitting. Therefore, the maximum depth of the decision trees with the optimal parameters obtained in grid search is 25. To reduce correlation, random forest introduces randomness into the ensemble, randomly selecting features from within the decision trees. The number of features selected is related to the correlation between the decision trees and the strength of the decision trees. Reducing the number of features decreases the correlation between trees, and the reduced complexity of each tree weakens its strength; increasing the number of features increases the strength of the trees, and the connectivity between them also improves. Among the specified parameters, it is most suitable to consider 80% of the total number of features when searching for the optimal cut.

[0051] from Figures 2-3As shown in Tables 5 and 6, the lower the accuracy of the NOx concentration prediction model at the outlet, the closer the curve between the predicted and actual values ​​will be, and the closer the predicted NOx concentration at the outlet will be to the actual value. In Table 6, the models built with parameters from groups A and F show significant deviations from the actual NOx concentration at the outlet, with a bias of 2-3%. Evaluating these two models, the average absolute percentage error for group A is 28.023%, and for group F it is 33.560%. In contrast, the model built with the optimal parameters shows a deviation of less than 0.5 from the actual NOx concentration at the outlet, with an average absolute percentage error of only 2.950%. Through parameter optimization, the accuracy of the prediction model was specifically improved, demonstrating that the model based on random forests has small prediction errors, high prediction accuracy, and strong generalization ability, achieving accurate and stable prediction of NOx emissions. This provides reliable technical support for the intelligent control and environmental compliance operation of the SNCR system in waste-to-energy incineration.

[0052] This invention also discloses a computer-readable storage medium storing a computer program thereon, which, when run by a processor, performs the steps of the method described above. This invention further discloses a random forest-based SNCR outlet NOx concentration prediction system, comprising an interconnected memory and a processor, wherein the memory stores a computer program that, when run by a processor, performs the steps of the method described above. The medium and system of this invention, corresponding to the methods described above, also possess the advantages described above.

[0053] The present invention can implement all or part of the processes in the methods of the above embodiments, or it can be implemented by hardware related to computer program instructions. The computer program can be stored in a computer-readable storage medium. When the computer program is executed by a processor, it can implement the steps of the above method embodiments. The computer program includes computer program code, which can be in the form of source code, object code, executable file, or some intermediate form. The computer-readable storage medium includes: any entity or device capable of carrying computer program code, recording media, USB flash drive, portable hard drive, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. The memory is used to store computer programs and / or modules. The processor implements various functions by running or executing the computer programs and / or modules stored in the memory, and by calling data stored in the memory. The memory may include high-speed random access memory, as well as non-volatile memory, such as hard disks, RAM, plug-in hard disks, smart media cards (SMC), secure digital (SD) cards, flash cards, at least one disk storage device, flash memory device, or other volatile solid-state storage devices.

[0054] The above are merely preferred embodiments of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should be considered within the scope of protection of the present invention.

Claims

1. A method for predicting NOx concentration at SNCR outlets based on random forests, characterized in that, Including the following steps: S1. Obtain the operating data of the SNCR system, select multiple input variables that affect the outlet NOx concentration, normalize the input variables, and obtain the normalized dataset; S2. Perform maximum information coefficient analysis on the normalized dataset to determine the delay time of each input variable relative to the outlet NOx concentration, and reconstruct the dataset according to the delay time to obtain a time-aligned reconstructed dataset. S3. Divide the reconstructed dataset into a training set and a test set, construct a random forest model, and initialize the model parameters; S4. Optimize the parameters of the random forest model to obtain the optimal parameter combination; S5. Using the random forest model with the optimal parameter combination, predict the NOx concentration at the outlet of the SNCR system, and perform inverse normalization on the prediction results to output the final prediction results.

2. The SNCR-based NOx concentration prediction method for outlets according to claim 1, characterized in that, In step S1, multiple input variables include inlet flue gas velocity, outlet flue gas velocity, outlet NH3 concentration, inlet O2 concentration, outlet O2 concentration, inlet flue gas temperature, outlet flue gas temperature, inlet NOx concentration, ammonia leak detection value, unit load, and cumulative fuel quantity.

3. The SNCR-based NOx concentration prediction method for outlets according to claim 1, characterized in that, In step S1, the normalization formula is: in It is the data value corresponding to the normalized variable x; This refers to the maximum value of a variable across all samples in the entire dataset; This refers to the minimum value of the variable across all samples in the entire dataset.

4. The method for predicting SNCR exit NOx concentration based on random forest according to claim 1, 2, or 3, characterized in that, In step S2, the specific process of the maximum information coefficient (MIC) analysis is as follows: S201. Divide the two-dimensional variables x and y into m and n different regions, discretize the variables, construct an m×n grid Q, and use the number of samples in the grid and the proportion of the number of samples in the region to the total number of samples to obtain the mutual information under the grid. in Let x and y be the mutual information values ​​of variables x and y when discretized in grid Q; p(x, y) Let be the joint probability density function. p (x) and p(y) The edge density function; S202. Select the maximum mutual information under different partitioning patterns and express it as: S203. Establish m×n grids of varying sizes, calculate the maximum mutual information value for each grid in sequence, and normalize it to obtain the final MIC value. The specific calculation method is as follows: Where m×n represents the total number of grid divisions, and the B value is required to be the power of 0.6 of the total data.

5. The method for predicting SNCR exit NOx concentration based on random forest according to claim 1, 2, or 3, characterized in that, In step S3, the model parameters include the number of decision trees, the maximum depth of the decision trees, the minimum number of samples required for a child node to split downwards, the minimum number of samples for a leaf node, and the number of features to be considered when finding the best cut.

6. The method for predicting SNCR exit NOx concentration based on random forest according to claim 1, 2, or 3, characterized in that, In step S4, the parameters of the random forest model are optimized using a grid search method with cross-validation.

7. The SNCR-based NOx concentration prediction method for outlets according to claim 6, characterized in that, The parameter range used in the grid search method is: Number of decision trees: 600; Maximum depth of decision trees: 25; Minimum number of samples required for a child node to split downwards: 2; Minimum number of samples for a leaf node: 1; Number of features to consider when finding the best cut: 0.

8.

8. The method for predicting SNCR exit NOx concentration based on random forest according to claim 1, 2, or 3, characterized in that, In step S5, the model's prediction performance is evaluated using mean square error, root mean square error, and mean absolute percentage error.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, The computer program, when run by a processor, performs the steps of the method as described in any one of claims 1-8.

10. A random forest-based SNCR outlet NOx concentration prediction system, comprising interconnected memory and processor, wherein the memory stores a computer program, characterized in that, The computer program, when run by a processor, performs the steps of the method as described in any one of claims 1-8.

Images