XGBoost-based boiler nitrogen oxide emission concentration prediction method and application

Abstract

The invention discloses a boiler nitrogen oxide emission concentration prediction method based on XGBoost and an application. The method comprises the following steps: step (1), collecting a data sample of an operation parameter related to the boiler nitrogen oxide emission concentration; (2) decomposing the data sample acquired in the step (1) by adopting a characteristic mode decomposition method to obtain decomposed data; step (3), training the constructed XGBoost model by using the decomposed data, and optimizing hyper-parameters of the XGBoost model by using a fruit fly optimization algorithm in the model training process; after training is completed, an FMD-FOA-XGBoost model is obtained, and model test evaluation is carried out; and step (4), utilizing the FMD-FOA-XGBoost model to predict the emission concentration of the nitrogen oxide of the boiler. By using the method, the boiler nitrogen oxide emission concentration under the deep peak regulation working condition can be predicted, and the prediction accuracy can be effectively improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of nitrogen oxide emission concentration prediction. In particular, it is a boiler nitrogen oxide emission concentration prediction method based on XGBoost and application. BACKGROUND

[0002] Nitrogen oxide (NOx) is one of the main pollutants in the atmosphere, and coal-fired power plants are the main source of its emission. Energy saving and emission reduction, especially the prediction of NOx emission, are crucial for coal-fired power plants. The NOx generated in the combustion process of coal-fired power plants is mainly divided into fuel type, thermal type and fast type. In the pulverized coal furnace, the generation of fast type NOx is more related to pressure than temperature, and can be ignored for large-scale boilers. Fuel type NOx accounts for about 80 % to 90 % of the total amount of NOx generated, while thermal type NOx accounts for about 10 % to 20 %. Accurate prediction of NOx concentration before the selective catalytic reduction (SCR) system helps to determine the optimal amount of ammonia injection required to control pollutant emissions. However, deep peaking will cause the boiler to operate under fluctuation between minimum load and rated load. The change of boiler load increases the difficulty of pollutant emission control and combustion optimization. Therefore, it is necessary to build a NOx concentration prediction model at the inlet of the SCR denitration system that can adapt to changing conditions to achieve precise ammonia injection control during deep peaking.

[0003] In machine learning algorithms, such as long short-term memory network (LSTM) and gated recurrent unit (GRU), have been applied to predict the pollutant emissions of coal-fired boilers, providing a new solution for emission reduction. Among the ten traditional data-driven algorithms, the GRU algorithm is recognized as the best choice for predicting the NOx emission rate of the SCR system, with an average root mean square error (RMSE) of 0.01. However, the prediction accuracy of the model will decrease as the accumulated prediction bias increases over time. Convolutional neural networks (CNN) exhibit stronger ability in capturing time series features. Lightweight CNN reduces the depth of the model while ensuring the prediction accuracy of the model, with an RMSE of 13.53 and an average absolute error of 9.89. To improve data quality, the mutual information (MI) method is used to determine the delay time. The prediction results of the least squares support vector machine show that the average absolute percentage error (MAPE) of the test set is less than 1 %, and the maximum error and RMSE are less than 8.0 mg / m³ and 4.0 mg / m³, respectively.

[0004] However, the coal-fired boiler data presents strong nonlinearity and coupling characteristics. The special deep peak regulation operation condition increases the system complexity, resulting in insufficient generalization ability and accuracy of the NOx prediction model. For non-stationary changes, researchers have proposed methods such as variational mode decomposition (VMD), improved adaptive noise complete ensemble empirical mode decomposition (CEEMDAN), and feature mode decomposition (FMD). After modal decomposition, the error results of the NOx prediction model fall within the minimum interval [0, 4], and most of the errors are concentrated in [0, 1]. Ensemble learning methods can capture complex patterns and improve overall prediction accuracy by integrating multiple algorithms. However, the NOx prediction accuracy still has a lot of room for improvement in the face of fluctuating data under deep peak regulation conditions. Currently, few studies use SHapley Additive exPlanations (SHAP) analysis to clarify the influence of input parameters on prediction results, and due to the lack of physical meaning support, the interpretability of the NOx prediction model is poor.

[0005] The extreme gradient boosting (XGBoost) algorithm has high accuracy, strong complex pattern capturing ability, fast computing efficiency and speed, excellent overfitting control ability, and strong flexibility in handling real data. The XGBoost prediction results show that the root mean square error (RMSE) is 2.11, which is more than 11.61% higher than other models. Based on XGBoost, a number of gas emission prediction studies have been completed, successfully realizing the ability to capture nonlinear information. However, directly using XGBoost to predict NOx emissions under deep peak regulation conditions is still not accurate enough. Therefore, it is necessary to further improve the existing XGBoost prediction model. SUMMARY

[0006] To this end, the technical problem to be solved by the present application is to provide a boiler nitrogen oxide emission concentration prediction method and application based on XGBoost, to solve the problem of inaccurate prediction results of existing boiler nitrogen oxide emission concentration prediction methods for NOx emissions from boilers under deep peak regulation conditions.

[0007] To solve the above technical problems, the present application provides the following technical solutions:

[0008] The boiler nitrogen oxide emission concentration prediction method based on XGBoost comprises the following steps:

[0009] Step (1), collect data samples of operating parameters related to boiler nitrogen oxide emission concentration;

[0010] Step (2), decompose and process the data samples collected in step (1) using the feature mode decomposition method to obtain the decomposed data;

[0011] Step (3), the decomposed data is used to train the constructed XGBoost model, and the fruit fly optimization algorithm is used to optimize the hyperparameters of the XGBoost model in the model training process; after training, the FMD-FOA-XGBoost model is obtained and model test evaluation is performed;

[0012] Step (4), the FMD-FOA-XGBoost model is used to predict the concentration of boiler nitrogen oxide emissions.

[0013] The above-mentioned XGBoost-based boiler nitrogen oxide emission concentration prediction method, in step (1), the data samples include boiler load, oxygen quantity, primary air quantity, secondary air damper opening, overfire air damper opening, total air quantity, coal supply quantity and boiler nitrogen oxide emission concentration.

[0014] The above-mentioned XGBoost-based boiler nitrogen oxide emission concentration prediction method, in step (1), the data samples are collected under deep peak regulation conditions, with an interval of 1s; the span of the data samples is 50%~100% of the rated load of the boiler; the number of data in the data samples is 15000-20000.

[0015] The above-mentioned XGBoost-based boiler nitrogen oxide emission concentration prediction method, in step (2), the method for processing the data samples by feature modal decomposition is:

[0016] Step (2-1), input the data sample, set the mode number n and the FIR filter length L; initialize the FIR filter set, and the FIR filter set has K filters;

[0017] Step (2-2), filter the input data sample to obtain a filtered signal;

[0018] Step (2-3), update the filter coefficients and iteratively optimize the filtering effect; determine whether the number of iterations reaches the preset value; if not, return to "step (2-2)" for continuous iteration until the number of iterations reaches the preset value;

[0019] Step (2-4), when the number of iterations reaches the preset value, calculate the K×K modal correlation matrix CC, and determine whether the mode K reaches the specified mode number n; if not, return to "step (2-2)" for continuous iteration until the mode K reaches the specified mode number n;

[0020] Step (2-5), when the mode K reaches the specified mode number n, the final decomposition mode is obtained, and the decomposed data is output.

[0021] The XGBoost-based boiler nitrogen oxide emission concentration prediction method, in step (3), the decomposed data is divided into a training set and a test set, and the data amount ratio in the training set and the test set is 80:20; the training process of the FMD-FOA-XGBoost model is as follows:

[0022] Step (3-1), initial parameter setting is performed on the constructed XGBoost model;

[0023] Step (3-2), the fruit fly optimization algorithm is used to optimize the hyperparameters of the XGBoost model, and the hyperparameters of the XGBoost model are updated;

[0024] Step (3-3), the XGBoost model with updated hyperparameters is trained and the model parameters are iterated using the training set;

[0025] Step (3-4), after the training is completed, it is judged whether the XGBoost model loss converges, if not, it is returned to step (3-2) to continue iteration and optimization until the model loss converges; if it converges, the final FMD-FOA-XGBoost model is obtained;

[0026] Step (3-5), the FMD-FOA-XGBoost model is evaluated and tested using the test set.

[0027] The XGBoost-based boiler nitrogen oxide emission concentration prediction method, in step (3-2), the method for optimizing the hyperparameters of the XGBoost model by the fruit fly optimization algorithm is as follows: first, the number of fruit fly populations and the number of iterations are initialized, then the position coordinates of the individual with the best taste concentration are found, and finally the best concentration value and the fruit fly position coordinates are input; in step (3-5), the evaluation index used in the evaluation test is the mean absolute percentage error, the root mean square error and the determination coefficient.

[0028] The XGBoost-based boiler nitrogen oxide emission concentration prediction method, in step (3-2), the search boundaries of each parameter when the fruit fly optimization algorithm optimizes the hyperparameters of the XGBoost model are set as follows:

[0029] Max_depth is [3, 16], Learning_rate is [0.01, 0.5], n_estimators is [50, 1000], Min_child_weight is [1, 10], Gamma is [0, 5], Subsample is [0.5, 1], Colsample_bytree is [0.5, 1], Reg_lambda is [0, +∞], and Reg_alpha is [0, +∞].

[0030] The super parameter settings of the FMD-FOA-XGBoost model in the XGBoost-based boiler nitrogen oxide emission concentration prediction method are as follows: Max_depth is 10, Learning_rate is 0.119, n_estimators is 200, Min_child_weight is 4, Gamma is 0.001, Subsample is 0.998, Colsample_bytree is 0.998, Reg_lambda is 0, and Reg_alpha is 2.671.

[0031] In the XGBoost-based boiler nitrogen oxide emission concentration prediction method, the time delay between the parameter variable and the nitrogen oxide concentration is evaluated by MI when the XGBoost model is constructed.

[0032] (1);

[0033] In formula (1), X and Y are two random variables; p (x, y) is the joint probability distribution of X and Y; p (x) and p (y) are the marginal probability distributions of X and Y, respectively.

[0034] The application of the XGBoost-based boiler nitrogen oxide emission concentration prediction method is used for predicting the boiler nitrogen oxide emission concentration under the deep peak regulation condition.

[0035] The technical scheme of the present application has the following beneficial technical effects:

[0036] 1. The boiler nitrogen oxide emission concentration prediction method based on XGBoost of the present application uses an optimized extreme gradient boosting (XGBoost) model combined with SHapley Additive exPlanations (SHAP) interpretable analysis, which can accurately predict the concentration of nitrogen oxides. The FMD-FOA-XGBoost prediction model constructed realizes the indicators of mean absolute percentage error (MAPE) 0.379% and root mean square error (RMSE) 2.412 mg / m³, which can be used for accurate prediction of NOx emission at the SCR inlet under the condition of boiler deep peak shaving. In the test set, feature mode decomposition and fruit fly optimization algorithm promote performance improvement, showing strong versatility, with RMSE reduced by 20.2% and 4.2% respectively, and MAPE reduced by 35.9% and 26.8% respectively. The SHAP analysis method is used to analyze the correlation between NOx emission and combustion variables. SHAP analysis shows that the main influencing factors of the feature mode decomposition-fruit fly optimization algorithm-XGBoost model prediction result are the overfire air, the load and the O2 value, while the influence of the primary air volume and the total air volume is smaller. The generalization ability of the proposed model is verified by using a new NOx emission data set, i.e. using 350 MW boiler nitrogen oxide emission data to prove the generalization ability of the model, and the results show that the relative error of 87.85% of the data points is less than 5%, and the relative error of 70.56% of the data points is less than 0.5%.

[0037] 2. The model constructed in the present application shows the best effect in the prediction accuracy and training efficiency of nitrogen oxide concentration, while improving the prediction accuracy and interpretability. The present application provides a new way to improve the NOx emission reduction effect during the deep peak shaving of the boiler, and the method can also be applied to the prediction and control of other pollutants. BRIEF DESCRIPTION OF DRAWINGS

[0038] Figure 1 The overall framework diagram of the prediction model constructed in the embodiment of the present application;

[0039] Figure 2 The nitrogen oxide concentration prediction results of different models in the training set in the embodiment of the present application;

[0040] Figure 3 The nitrogen oxide concentration prediction results of different models in the test set in the embodiment of the present application;

[0041] Figure 4a and Figure 4b The relative error results (violin plot) of the FMD-FOA-XGBoost training set and test set of the five different models in the embodiment of the present application are shown in Figs. 1 to 5, respectively;

[0042] Figure 5a and Figure 5brespectively are the feature importance ranking based on SHAP value and SHAP summary plot of the embodiment of the present application;

[0043] Figure 6 the SHAP dependency relationship diagram of the FMD-FOA-XGBoost model constructed in the embodiment of the present application;

[0044] Figure 7a and Figure 7b respectively are the prediction result distribution plot and the relative error distribution plot compared with the measured value of the model constructed in the embodiment of the present application on the new data set. DETAILED DESCRIPTION

[0045] 1. Construction of a boiler nitrogen oxide emission concentration prediction model under deep peak regulation conditions

[0046] 1.1 Boundary conditions

[0047] A 660 MW pulverized coal boiler is selected as the research object in this embodiment. The boiler adopts Π type arrangement and once intermediate reheat system, the furnace width is 23,500 mm, the depth is 15,400 mm, and the height is about 66,000 mm. The rated steam pressure of the boiler is 25.4 MPa, the maximum continuous evaporation capacity is 2,250 t / h, the rated evaporation capacity is 2,155 t / h, and the superheater and reheater steam temperatures are 571 ℃ and 569 ℃, respectively.

[0048] 1.2 Model verification and parameters

[0049] The formation of nitrogen oxides in coal-fired boilers is influenced by multiple factors, including load, O2 value, primary air volume, secondary air volume, overfire air volume, total air volume, coal feed rate, and coal properties. In coal-fired power plants, the type of coal remains unchanged within a certain period of time, so the properties of coal are also a constant value. Parameters such as coal feed rate, overfire air volume, and O2 value can also reflect coal quality information to some extent. Therefore, the properties of coal are not included in the candidate variables of this embodiment. The coal burned in the boiler in this embodiment is quasi-Gegengqi bituminous coal from China. Seven operating parameters related to the SCR inlet nitrogen oxide concentration were selected, and 18,400 one-second interval samples were collected under peak regulation conditions, spanning 50% ~ 100% of the rated load. 114,720 samples were used for training, and 3,680 samples were used for testing. The variables and detailed statistical parameters are shown in Table 1.

[0050] Table 1. Variables and statistical parameters for model input

[0051]

[0052] 1.3 Model construction

[0053] FMD (Feature Modal Decomposition) uses an adaptive finite impulse response filter bank, in which the filter coefficients are iteratively updated to make the filtered signal close to the deconvolution objective function. This process can decompose the signal into different modes. For large-scale industrial data sets such as boiler nitrogen oxide emissions, FMD provides an effective noise attenuation method, thereby improving the robustness of nitrogen oxide prediction. FOA algorithm (fruit fly optimization algorithm) is used to optimize the parameters of XGBoost model, which can improve its prediction ability. Nitrogen oxide emission data is divided into two parts, one part is used for training, and the other part is used for testing. Then set the hyperparameters according to the initial value of the XGBoost model, and continue to optimize the hyperparameters using the fruit fly optimization model, and the training effect of the XGBoost model reaches the best. The model construction and prediction process is shown in Figure 1 as follows:

[0054] Step (1), collect data samples of operating parameters related to the concentration of boiler nitrogen oxide emissions; the data samples include boiler load, oxygen amount, primary air volume, secondary air damper opening, overfire air damper opening, total air volume, coal supply amount and boiler nitrogen oxide emission concentration;

[0055] Step (2), decompose the data samples collected in step (1) using feature modal decomposition method to obtain decomposed data; the method for decomposing the data samples by feature modal decomposition is as follows:

[0056] Step (2-1), input the data samples, set the mode number n and the FIR filter length L; initialize the FIR filter bank, and the FIR filter bank has K filters;

[0057] Step (2-2), filter the input data samples to obtain a filtered signal;

[0058] Step (2-3), update the filter coefficients and iteratively optimize the filtering effect; determine whether the number of iterations reaches the preset value; if not, return to "step (2-2)" for continuous iteration until the number of iterations reaches the preset value;

[0059] Step (2-4), when the number of iterations reaches the preset value, calculate the KxK mode correlation matrix CC, and determine whether the mode K reaches the specified mode number n; if not, return to "step (2-2)" for continuous iteration until the mode K reaches the specified mode number n;

[0060] Step (2-5), when the mode K reaches the specified mode number n, the final decomposition mode is obtained, and the decomposed data is output.

[0061] Step (3), the decomposed data is used to train the constructed XGBoost model, and the fruit fly optimization algorithm is used to optimize the hyperparameters of the XGBoost model during the model training process; after the training is completed, the FMD-FOA-XGBoost model is obtained and the model test evaluation is performed; specifically, the decomposed data is divided into a training set and a test set, and the data amount ratio in the training set and the test set is 80:20; the training process of the FMD-FOA-XGBoost model is as follows:

[0062] Step (3-1), the initial parameter setting of the constructed XGBoost model is performed; step (3-2), the fruit fly optimization algorithm is used to optimize the hyperparameters of the XGBoost model, and the hyperparameters of the XGBoost model are updated; the method for the fruit fly optimization algorithm to optimize the hyperparameters of the XGBoost model is as follows: firstly, the number of fruit fly populations and the number of iterations are initialized, then the position coordinates of the individual with the best taste concentration are found out, and finally the best concentration value and the fruit fly position coordinates are input; step (3-3), the XGBoost model with updated hyperparameters is trained by using the training set to update the model parameters; step (3-4), after the training is completed, it is judged whether the loss of the XGBoost model converges or not, if not, it is returned to step (3-2) to continue the iteration optimization until the model loss converges; if the convergence is achieved, the final FMD-FOA-XGBoost model is obtained; step (3-5), the FMD-FOA-XGBoost model is evaluated and tested by using the test set

[0063] Step (4), the FMD-FOA-XGBoost model is used to predict the concentration of nitrogen oxide emissions of the boiler.

[0064] 1.4 Time lag processing

[0065] The time delay between the 7 input variables and the nitrogen oxide concentration is evaluated by MI, as shown in Table 2. The MI formula is as follows:

[0066] (1);

[0067] Wherein: X and Y are two random variables; p (x, y) is the joint probability distribution of X and Y; p (x) and p (y) are the marginal probability distributions of X and Y, respectively.

[0068] Table 2 MI and time delay analysis (time step = 1 second)

[0069]

[0070] 1.5 Evaluation index

[0071] The performance of the model in predicting the concentration of nitrogen oxides at the inlet of the SCR was verified by three evaluation indicators. The mean absolute percentage error MAPE, the root mean square error RMSE and the determination coefficient R 2 , and the relevant formulas are as follows:

[0072] (2);

[0073] (3);

[0074] (4);

[0075] wherein: n represents the number of sample points; and represent the predicted value and the actual value; represents the average value of the samples in the test set.

[0076] 2、Results and discussion

[0077] 2.1 Evaluation of the prediction accuracy of the model

[0078] The preprocessed data was applied to the four models (VMD, REMD, CEEMDAN and FMD) to train and evaluate their fault diagnosis performance. The prediction results of different methods are shown in Table 3, and FMD shows higher accuracy in capturing the nonlinear characteristics of the primary air volume, the main steam temperature and the pressure. Compared with VMD, TCN and CEEMDAN, FMD has smaller errors (training set: MAPE 1.291, RMSE 5.364; test set: MAPE 1.345, RMSE 5.618) and higher R² (both 0.988), highlighting its advantages in feature extraction and model prediction.

[0079] Table 3 Prediction results of different methods based on FMD-XGBoost

[0080]

[0081] Table 4 shows the prediction results of the XGBoost model using different optimization algorithms on the training and test sets, including the sparrow search algorithm (SSA), dung beetle optimization algorithm (DBO), particle swarm optimization algorithm (PSO), and fruit fly optimization algorithm (FOA). FOA-XGBoost achieved the lowest error (training set: MAPE 0.324, RMSE 2.204; test set: MAPE 0.379, RMSE 2.412) and the highest R² (training set 0.998; test set 0.998), while PSO-XGBoost also showed strong prediction accuracy (training set: MAPE 0.342, RMSE 2.209; test set: MAPE 0.395, RMSE 2.517), indicating that the prediction accuracy was improved and the hyperparameter optimization was more robust.

[0082] Table 4 Prediction results after demonstration of different optimization algorithms

[0083]

[0084] Table 5 shows the results of the algorithm after optimizing the hyperparameters. As shown in Table 5, the FOA-optimized XGBoost model has the best prediction ability. Through multiple rounds of iterative search of the optimization algorithm, the best hyperparameter combination is determined, thereby improving the performance of XGBoost in predicting nitrogen oxide concentration.

[0085] Table 5 Best hyperparameters of different optimization algorithms

[0086]

[0087] This embodiment selects five prediction models, namely FOA-XGBoost, LSTM, BiLSTM, GRU, and SVM, for comparison with the proposed model. Table 6 summarizes the prediction indicators of the models, indicating the prediction ability of the models for nitrogen oxide emission concentration under peak-shaving conditions. The proposed model in this embodiment has a clear advantage in terms of MSE, RMSE, and R 2

[0088] Table 6 Performance of different methods in predicting nitrogen oxide concentration

[0089]

[0090] Figure 2 ​The (a-d) in FIG. 1 show that LSTM and BiLSTM can capture the trend of the concentration of nitrogen oxide emissions regularly, but cannot fit some peak sample points with large values, and cannot fit the change of the concentration of nitrogen oxide, which shows that the combustion condition of the boiler is more complex and variable under the peak regulation condition. Although LSTM and BiLSTM have an advantage in capturing long-term dependence of sequences, they cannot effectively extract important features of the concentration of nitrogen oxide emissions data under the peak regulation condition.

[0091] As shown in Figure 2 (e-f) in FIG. 1, on the training set, the effect of FOA-XGBoost is better than that of the traditional time series model alone. Compared with other single models, the FOA-XGBoost model can fully extract the features of the concentration of nitrogen oxide emissions data, but cannot extract the long-term dependence of the features. The FMD-FOA-XGBoost model proposed in the embodiment can track the trend of the concentration of nitrogen oxide emissions, and improves the immunity of the model for data peak points, especially in the case of non-compliance of the concentration of nitrogen oxide emissions. In the latter part of the test set, the concentration of nitrogen oxide shows frequent fluctuations, and compared with the remaining five models on the training set, the FMD-FOA-XGBoost model proposed in the embodiment achieves better fitting.

[0092] The accuracy of the prediction model can be evaluated by judging whether the predicted value and the actual value are aligned with the y=x line. Figure 3 (f) in FIG. 1 shows the closest alignment of the actual value and the predicted value on the test set, and the points on the diagonal line are very concentrated, indicating that the prediction accuracy is very high. The analysis of Figure 3 (e) and (f) in FIG. 1 shows that the regression performance of the FMD-FOA-XGBoost test set is significantly improved compared with the XGBoost model. Other models, such as LSTM ( Figure 3 (d) in FIG. 1) and GRU ( Figure 3 (b) in FIG. 1), show significant performance, but lag behind the FMD-FOA-XGBoost in terms of accuracy and consistency. When the optimized model predicts the concentration of nitrogen oxide, the deviation between the predicted value and the actual value is small, so the fitting degree is high. This shows that the FOA optimization strategy can effectively improve the performance of XGBoost in predicting the concentration of nitrogen oxide.

[0093] Figure 4a and Figure 4b shows the relative error and error distribution range of the different models used in the embodiment in predicting the concentration of nitrogen oxide emissions of the training set and the test set under the peak regulation condition. The relative error is defined as:

[0094] (5);

[0095] wherein is the measured value, is the predicted value.

[0096] In violin plots, the length represents the error range, and the width reflects the density of the relative error distribution. The rectangular box inside the violin corresponds to the range between quartiles, with the upper and lower edges representing the third and first quartiles, respectively, and the center point representing the median. As Figure 4a The relative error in the test set is shown. The relative error range of the FMD-FOA-XGBoost model is limited to [-15.88, 17.87], indicating that its error control is stable. It is worth noting that the relative error ranges of FOA-XGBoost, LSTM, BiLSTM, GRU, and SVM models are [-15.85, 20.12], [-29.09, 26.81], [-29.97, 29.23], [-30.97, 28.55], and [-27.32, 33.22], respectively, indicating that they have good prediction performance in the training set. From Figure 4b It can be seen from the table that the relative error range of the FMD-FOA-XGBoost model is [-13.42, 15.82], and the relative error ranges of FOA-XGBoost, LSTM, BiLSTM, GRU, and SVM models are [-13.45, 16.85], [-24.08, 26.81], [-24.72, 25.93], [-30.61, 30.25], and [-22.63, 33.22], respectively, showing the prediction ability in the test set.

[0097] 2.2 SHAP analysis

[0098] SHAP is an analysis method based on the game theory framework. SHAP values can quantify the contribution of input features to the prediction target and intuitively show the influence of various input features on nitrogen oxide concentration. The SHAP value of a feature is calculated as follows:

[0099] (6);

[0100] where S represents the input feature, F represents the input feature without index I; N represents the number of features; and f represents the prediction function.

[0101] SHAP value analysis can explain the relationship between input features and model prediction, and the results show the relationship in the test set. In general, Figure 5a and Figure 5bThe significant difference in the contribution of each input feature to the prediction of nitrogen oxide concentration is shown. The overfire air and load become the main factors affecting the prediction ability of the model, playing a control role in generating output values. The influence of excess air on the prediction result is consistent with the dependence of oxygen value on the generation of thermal nitrogen oxides in the combustion process. The boiler load is consistent with the parameters such as the amount of primary air and the amount of coal, which have a significant impact on the generation of fuel nitrogen oxides under peak regulation conditions.

[0102] When the overfire air damper opening is greater than 85.42%, it is positively correlated with the NO x concentration, and less than 85.28%, it is negatively correlated; when the load is between 505.03 MW and 591.68 MW, it is positively correlated with the nitrogen oxide concentration, and in the remaining range, it is negatively correlated; when the O2 value is less than 0.68% and within the range of 2.02% to 3.42%, it is negatively correlated with the nitrogen oxide concentration, and when the O2 concentration is within the range of 0.68% to 2.02% and 3.05% to 4.02%, it is positively correlated with the nitrogen oxide concentration. When the coal supply is greater than 1998.37 t / h, it is positively correlated with the nitrogen oxide concentration, and the remaining compliance is negatively correlated. This indicates that the coal supply rate and fuel NO x between 69.89% and 80.92%. In contrast, the importance of the reactor inlet temperature is lower because its influence on the nitrogen oxide formation process is weak and is strictly controlled by the primary and secondary air ratio and oxygen value. The influence of total air flow is the smallest, showing a negative effect when its value is lower than 1510 t / h and a positive correlation when it is higher than 1510 t / h. The principles of combustion and nitrogen oxide formation indicate that air volume and coal supply rate are indicators of combustion intensity, which in turn reflects the depth of boiler peak regulation.

[0103] 2.3 Model generalization ability

[0104] To further verify the generalization ability of the model, this embodiment collects a DCS and CEMS dataset of a 350 MW boiler in Inner Mongolia, China, totaling 8,642 groups of data, and uses it as the input of the proposed model. The nitrogen oxide concentration prediction results based on the FMD-FOA-XGBoost model are shown in Figure 6. The results show that the predicted values of the proposed model are basically close to the actual values. It can be seen that the nitrogen oxide concentration prediction values of the proposed model are very close to the measured values, and can accurately reflect the trend of nitrogen oxide emissions under peak regulation conditions, and can also be well fitted when there are large fluctuations.

[0105] The prediction graph of the 350 MW coal-fired boiler is shown in Figure 7a and Figure 7b. The horizontal axis represents the measured value, the vertical axis represents the predicted value, the black line represents the ideal curve, and the red line represents the fitting line of the sample points. The blue points represent the distribution of the prediction results. The closer the scatter points are to the ideal curve, the higher the prediction accuracy. As can be seen from Figure 7a , the scatter point distribution of the proposed model prediction results is closer to the ideal curve, indicating that the prediction results of the proposed model are better.

[0106] Error distribution analysis is used to evaluate the statistical significance of model prediction. Figure 7b The error distribution of the proposed model on a 350 MW coal-fired boiler is described. The relative error distribution shows a sharp, symmetric peak, with most errors concentrated around zero, reflecting the best prediction accuracy and generalization ability of the proposed model. The relative error of 7592 data points is less than 5%, and the relative error of 6098 data points is less than 0.5%. This result shows that the proposed model can capture nonlinear and linear relationships and manage the inherent complexity of the data.

[0107] 3. Conclusion

[0108] This embodiment proposes a boiler SCR inlet nitrogen oxide emission prediction method based on FOA optimized XGBoost model. It is verified by data from a 660 MW coal-fired boiler. In the test set, the FMD-FOA-XGBoost model has a higher R 2 (0.998), lower MAPE (0.379%) and RMSE (2.412 mg / m 3 ) than other widely used nitrogen oxide prediction models (GRU, LSTM and XGBoost). In addition, the FMD and FOA optimization algorithms promote performance improvement, showing strong versatility. In the test set, the RMSE of the GRU and LSTM models is reduced by 20.2% and 4.2%, respectively, and the MAPE is reduced by 35.9% and 26.8%, respectively.

[0109] SHAP analysis shows that the main influencing factors of the FMD-FOA-XGBoost model prediction results are superheated air, load and oxygen value, while the influence of primary air distribution and total airflow is smaller. The model proposed in this embodiment is consistent with the nitrogen oxide generation mechanism and regulation strategy of each system. To verify the generalization ability of the model, a new nitrogen oxide emission data set collected from a 350 MW boiler is used. The results show that the relative error of 87.85% of the data points is less than 5%, and the relative error of 70.56% of the data points is less than 0.5%. The FMD-FOA-XGBoost model proposed in this embodiment performs well in terms of prediction accuracy and training efficiency, and takes into account both prediction accuracy and interpretability.

Claims

1. A method for predicting boiler nitrogen oxide emission concentration based on XGBoost, characterized in that, Includes the following steps: Step (1): Collect data samples of operating parameters related to the concentration of nitrogen oxide emissions from the boiler; Step (2): Use the Eigenmode Decomposition method to decompose the data samples collected in step (1) to obtain the decomposed data; Step (3): Train the constructed XGBoost model using the decomposed data, and optimize the hyperparameters of the XGBoost model using the fruit fly optimization algorithm during the model training process; after training is completed, obtain the FMD-FOA-XGBoost model and perform model testing and evaluation. Step (4): Predict the concentration of nitrogen oxide emissions from the boiler using the FMD-FOA-XGBoost model.

2. The boiler nitrogen oxide emission concentration prediction method based on XGBoost according to claim 1, characterized in that, In step (1), the data sample includes boiler load, oxygen quantity, primary air volume, secondary air damper opening, burnout air damper opening, total air volume, coal feed rate, and boiler nitrogen oxide emission concentration.

3. The boiler nitrogen oxide emission concentration prediction method based on XGBoost according to claim 2, characterized in that, In step (1), data samples are collected under deep peak shaving conditions, with a collection interval of 1 second; the range of data samples is 50% to 100% of the boiler's rated load; and the number of data samples is 15,000 to 20,000.

4. The boiler nitrogen oxide emission concentration prediction method based on XGBoost according to claim 3, characterized in that, In step (2), the method for performing feature mode decomposition on the data samples is as follows: Step (2-1): Input data samples, set the number of modes n and the FIR filter length L; initialize the FIR filter bank, which has a total of K filters; Step (2-2): Filter the input data samples to obtain the filtered signal; Step (2-3): Update the filter coefficients and iteratively optimize the filtering effect; determine whether the number of iterations has reached the preset value; if not, return to "Step (2-2)" to continue iterating until the number of iterations reaches the preset value; Step (2-4): When the number of iterations reaches the preset value, calculate the K×K modal correlation matrix CC and determine whether mode K has reached the specified number of modes n; if not, return to "Step (2-2)" to continue iterating until mode K reaches the specified number of modes n. Step (2-5): When pattern K reaches the specified number of patterns n, the final decomposition pattern is obtained, and the decomposed data is output.

5. The boiler nitrogen oxide emission concentration prediction method based on XGBoost according to claim 3, characterized in that, In step (3), the decomposed data is divided into a training set and a test set, with a data volume ratio of 80:

20. The training process of the FMD-FOA-XGBoost model is as follows: Step (3-1): Set the initial parameters for the constructed XGBoost model; Step (3-2): Optimize the hyperparameters of the XGBoost model using the fruit fly optimization algorithm and update the hyperparameters of the XGBoost model. Step (3-3): Use the training set to train the updated hyperparameter XGBoost model and iterate the model parameters. Step (3-4): After training, determine whether the XGBoost model loss has converged. If it has not converged, return to step (3-2) to continue iterative optimization until the model loss converges. If it has converged, the final FMD-FOA-XGBoost model is obtained. Steps (3-5): Evaluate and test the FMD-FOA-XGBoost model using the test set.

6. The boiler nitrogen oxide emission concentration prediction method based on XGBoost according to claim 5, characterized in that, In step (3-2), the method for optimizing the hyperparameters of the XGBoost model using the fruit fly optimization algorithm is as follows: First, initialize the fruit fly population size and the number of iterations, then find the location coordinates of the individual with the best taste concentration, and finally input the best concentration value and the fruit fly location coordinates; In step (3-5), the evaluation indicators used in the evaluation test are mean absolute percentage error, root mean square error and coefficient of determination.

7. The boiler nitrogen oxide emission concentration prediction method based on XGBoost according to claim 6, characterized in that, In step (3-2), the search boundaries for each parameter are set as follows when the fruit fly optimization algorithm optimizes the hyperparameters of the XGBoost model: Max_depth is [3,16], Learning_rate is [0.01,0.5], n_estimators is [50,1000], Min_child_weight is [1,10], Gamma is [0,5], Subsample is [0.5,1], Colsample_bytree is [0.5,1], Reg_lambda is [0, +∞], and Reg_alpha is [0, +∞].

8. The boiler nitrogen oxide emission concentration prediction method based on XGBoost according to claim 7, characterized in that, The hyperparameters of the FMD-FOA-XGBoost model are set as follows: Max_depth = 10, Learning_rate = 0.119, n_estimators = 200, Min_child_weight = 4, Gamma = 0.001, Subsample = 0.998, Colsample_bytree = 0.998, Reg_lambda = 0, and Reg_alpha = 2.

671.

9. The boiler nitrogen oxide emission concentration prediction method based on XGBoost according to claim 3, characterized in that, When constructing the XGBoost model, the time delay between the parameter variables and the nitrogen oxide concentration is evaluated using MI; the MI formula is as follows: (1)； In equation (1), X and Y are two random variables; p(x, y) is the joint probability distribution of X and Y; p(x) and p(y) are the marginal probability distributions of X and Y, respectively.

10. Application of a boiler nitrogen oxide emission concentration prediction method based on XGBoost, characterized in that, The boiler nitrogen oxide emission concentration prediction method based on XGBoost as described in any one of claims 1-9 is used to predict the boiler nitrogen oxide emission concentration under deep peak shaving conditions.

Images