SAO-Ensemble Weighted Fusion Dynamic Prediction Model NO x Prediction Methods

Abstract

The present invention belongs to the technical field of denitrification in glass factories. The present invention proposes a dynamic prediction model of SAO-integrated weighted fusion for NO x The prediction method comprises: collecting real-time combustion working data of the glass factory; inputting the real-time combustion working data into a preset dynamic prediction model to predict the SCR inlet NO x concentration; wherein, the dynamic prediction model is constructed based on the integrated weighted fusion of LightGBM sub-model, XGBoost sub-model and CatBoost sub-model, and is obtained through data set training; the data set includes: combustion working data after time lag compensation and the corresponding SCR inlet NO x The combustion working data include: unit load, natural gas flow, oxygen content in the kiln, average kiln temperature, flue gas oxygen content and total air volume. The present invention can achieve SCR inlet NO x Accurate prediction of concentration.

Description

Technical Field

[0001] The present invention belongs to the technical field of denitration in glass factories, and in particular to a dynamic prediction model of NO based on SAO-integrated weighted fusion. x Prediction method. Background Art

[0002] Although glass furnaces use natural gas as their furnace fuel, these fuels still produce a large amount of NO x , SO2, CO2 and dust, etc. Among them, NO x Excessive emissions can lead to natural disasters such as greenhouse effect and acid rain, which seriously affect the human living environment and endanger human health. x There are several basic types of emission control: pre-combustion control technology, combustion control technology and post-combustion control technology. Among the three, the selective catalytic reduction technology (SCR) in the post-combustion control technology has become the most widely used denitrification method at home and abroad. The denitrification principle of SCR technology is that under the condition of working temperature of about 300-400°C, the nitrogen oxides and ammonia in the flue gas emitted by boiler combustion undergo redox reaction under the action of catalyst, and the reactants are non-toxic and pollution-free nitrogen and water, thereby achieving denitrification of boiler flue gas.

[0003] Ideally, the denitrification efficiency of SCR technology can exceed 90%. However, since the glass furnace is a nonlinear system with large delay and inertia, and the actual installation location of the CEMS detection instrument is generally far away from the inlet of the SCR system, it takes some time for the flue gas to flow to the CEMS, and the measurement also takes some time, so the measured value is different from the SCR inlet NO x The actual concentration value has a time lag of 40 to 50 seconds. This makes it impossible for us to obtain accurate SCR inlet NO through the instrument. x If the ammonia concentration is too low, it will be difficult to accurately control the amount of ammonia sprayed, which will result in the denitrification system being unable to spray ammonia in a timely, effective and accurate manner. If the amount of ammonia sprayed is too low, the denitrification effect of the SCR system will be poor, which will increase the emission of NO in the exhaust gas. x The amount of NO x Emissions do not meet standards; excessive ammonia injection will affect the denitrification efficiency, causing huge resource consumption and increasing operating costs. Therefore, we need to establish an algorithm model to analyze the SCR inlet NO x The concentration can be predicted and analyzed, which can not only help the denitrification system to accurately control the amount of ammonia sprayed and improve the denitrification quality, but also reduce the denitrification cost of the glass factory.

[0004] In establishing accurate and effective SCR inlet NO xBefore the prediction model was developed, the SCR denitrification reaction was easily affected by the activity of the catalyst, and there were many interference factors in the unit, so the mechanism model could not fully describe the real object. Therefore, how to determine the relationship between the input and output of the system has always been the key to research. x The concentration is affected by many factors. Therefore, in the process of building the SCR system, it is necessary to simplify the structure of the model by selecting appropriate input variables.

[0005] In addition, the SCR system is a typical large time-delay controlled object, and many input variables to the inlet NO x The response delay of concentration was more than 45s, and NO x The measurement of concentration itself also has a lag time, which causes the entire response process to take several minutes. All of these will affect the accuracy of the prediction model. Therefore, in the modeling process, the impact of input variable lag on input variable selection and model prediction accuracy should be considered.

[0006] Finally, in terms of model building, due to the very complex working conditions of boiler combustion, a single model often cannot meet the requirements for NOx concentration prediction at the SCR inlet under different working conditions. Therefore, it is a very good breakthrough to design a multi-model hybrid integrated model with high adaptability and robustness, which can maintain high prediction accuracy under different working conditions.

[0007] In summary, by selecting appropriate modeling methods, an accurate, efficient, and highly adaptive SCR inlet NO model can be constructed under different working conditions. x The concentration prediction model can accurately adjust the key parameters such as the amount of ammonia injection in the denitrification process, thereby improving the denitrification efficiency and reducing the operating costs of the enterprise. x The key to the present invention is to develop a prediction model for NOx content and improve the monitoring of NOx concentration. Summary of the invention

[0008] The purpose of this invention is to propose a dynamic prediction model of SAO-integrated weighted fusion. x Prediction method: A single model for SCR system cannot meet the needs of NOx at SCR inlet under different working conditions. x The problem of concentration prediction is solved by identifying the NO x Concentration-related auxiliary variables reduce the complexity of the model while ensuring the accuracy of the model. The traditional single model is packaged and integrated into an integrated model through historical data, and the prediction results are weighted and integrated using an optimization algorithm to complete the prediction of NO inlet of the glass plant SCR. x Concentration prediction model is built to achieve SCR inlet NO x Accurate prediction of concentration.

[0009] To achieve the above object, the present invention provides a SAO-integrated weighted fusion dynamic prediction model of NO x Prediction methods, including:

[0010] Collect real-time combustion data of glass factories;

[0011] The real-time combustion working data is input into the preset dynamic prediction model, and the SCR inlet NO is output. x Predicted concentration; wherein the dynamic prediction model is constructed based on the integrated weighted fusion of LightGBM sub-model, XGBoost sub-model and CatBoost sub-model, and is obtained through data set training; the data set includes: combustion working data after time lag compensation and the corresponding SCR inlet NO x concentration; the combustion working data include: unit load, natural gas flow, oxygen content in the kiln, average kiln temperature, flue gas oxygen content and total air volume.

[0012] Optionally, constructing the dynamic prediction model includes:

[0013] The snow melting optimization algorithm is used to optimize the weights of the LightGBM sub-model, the XGBoost sub-model, and the CatBoost sub-model to obtain an optimal weight distribution scheme;

[0014] Based on the optimal weight distribution scheme, the LightGBM sub-model, the XGBoost sub-model and the CatBoost sub-model are integrated and weighted fused to obtain the dynamic prediction model.

[0015] Optionally, obtaining the optimal weight allocation scheme includes:

[0016] Generate a set of initial individuals, each individual represents a weight allocation scheme in the LightGBM sub-model, the XGBoost sub-model, and the CatBoost sub-model;

[0017] Based on the preset indicators, the performance of each individual is evaluated and the fitness value is calculated;

[0018] According to the fitness value of the individual, its position is continuously updated; the position represents the current weight distribution scheme of the individual.

[0019] When the maximum number of iterations is reached, the individual with the best fitness value is output as the optimal solution, which is the optimal weight allocation scheme.

[0020] Optionally, the preset indicators include: root mean square error, mean absolute percentage error and mean absolute error.

[0021] Optionally, the dynamic prediction model is:

[0022]

[0023] in, k is the number of models used, G m is the predicted value corresponding to the integrated model, G i is the predicted value of the ith model, w i is the weight of the i-th surrogate model.

[0024] Optionally, constructing the data set includes:

[0025] The glass factory DCS system is used to collect auxiliary variable data; the auxiliary variable data includes: the combustion working data;

[0026] The auxiliary variable data are screened by using the mutual information and mRMR method with the output accuracy of the dynamic prediction model as a screening criterion, and the combustion working data are determined as the input variable of the model;

[0027] SCR inlet NO x The concentration is used as the output variable, and the mutual information of the input variable and the output variable within a preset time is analyzed to obtain the mutual information of the two variables and the corresponding time delay;

[0028] Based on the corresponding time delay, the input variable is compensated for the time delay;

[0029] The data set is constructed based on the input variables and output variables after time lag compensation.

[0030] Optionally, performing time lag compensation on the input variable includes:

[0031] Performing a 215s time lag compensation on the unit load;

[0032] Performing a time delay compensation of 166 seconds on the natural gas flow;

[0033] Performing a time lag compensation of 105s on the oxygen content in the kiln;

[0034] A time lag compensation of 148 seconds is performed on the average temperature of the kiln;

[0035] Performing a time lag compensation of 121 seconds on the oxygen content of the flue gas;

[0036] The total air volume is compensated for a time lag of 175 seconds.

[0037] Optionally, after collecting the auxiliary variable data, the method further includes:

[0038] Using the box plot method to identify and remove abnormal values ​​in the auxiliary variable data;

[0039] The data gaps after elimination are filled with the mean, and the sliding average method is used for filtering and noise reduction, and normalization is performed.

[0040] The present invention has the following beneficial effects:

[0041] The present invention aims at the problem that a single model of the SCR system cannot meet the requirements of the NO at the SCR inlet under different working conditions. x The problem of concentration prediction is solved by identifying the NO x Concentration-related auxiliary variables reduce the complexity of the model while ensuring the accuracy of the model. The traditional single model is packaged and integrated into an integrated model through historical data, and the prediction results are weighted and integrated using an optimization algorithm to complete the prediction of NO inlet of the glass plant SCR. x The concentration prediction model is built to achieve precise adjustment of key parameters such as the amount of ammonia sprayed during the denitrification process, thereby improving the denitrification efficiency and reducing the operating costs of the enterprise. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] The drawings constituting a part of the present application are used to provide a further understanding of the present application. The illustrative embodiments and descriptions of the present application are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0043] Figure 1 A dynamic prediction model based on SAO-integrated weighted fusion is described in an embodiment of the present invention. x Schematic diagram of the prediction method flow chart. DETAILED DESCRIPTION

[0044] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features in the embodiments can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0045] It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and that, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0046] like Figure 1 As shown, this embodiment proposes a dynamic prediction model of SAO-integrated weighted fusion. x Prediction methods, including:

[0047] Collect real-time combustion data of glass factories;

[0048] The real-time combustion working data is input into the preset dynamic prediction model, and the SCR inlet NO is output. x Predicted concentration; the dynamic prediction model is constructed based on the integrated weighted fusion of LightGBM sub-model, XGBoost sub-model and CatBoost sub-model, and is obtained through data set training; the data set includes: combustion working data after time lag compensation and the corresponding SCR inlet NO x concentration; combustion working data include: unit load, natural gas flow, oxygen content in the kiln, average kiln temperature, flue gas oxygen content and total air volume.

[0049] Furthermore, building a dynamic prediction model includes:

[0050] The snow melting optimization algorithm is used to optimize the weights of the LightGBM sub-model, XGBoost sub-model, and CatBoost sub-model to obtain the optimal weight distribution scheme;

[0051] Based on the optimal weight distribution scheme, the LightGBM sub-model, XGBoost sub-model and CatBoost sub-model are integrated and weighted fused to obtain a dynamic prediction model.

[0052] Furthermore, obtaining the optimal weight allocation scheme includes:

[0053] Generate a set of initial individuals, each of which represents a weight distribution scheme in the LightGBM sub-model, the XGBoost sub-model, and the CatBoost sub-model;

[0054] Based on the preset indicators, the performance of each individual is evaluated and the fitness value is calculated;

[0055] Specifically, in this embodiment, the fitness value in the exploration phase is calculated as:

[0056] In the exploration phase, the algorithm simulates the process of converting snow or liquid water from snow into steam. In this phase, the search individuals are highly dispersed, so Brownian motion is used to simulate this situation. The one-dimensional distribution of Brownian motion and the trajectory in the two-dimensional and three-dimensional search space can explore some potential areas in the search space. In this phase, the calculation of fitness values ​​involves the selection of elite individuals, which are randomly selected from the current optimal solution and other individuals in the group.

[0057] Fitness value calculation in the development phase:

[0058] In the development phase, the algorithm simulates the process of snow turning into liquid water through melting behavior. In this phase, search agents are encouraged to develop high-quality solutions around the current best solution. This phase uses the degree-day method to simulate the snowmelt process, encouraging search agents to find high-quality solutions near the current best solution. In the development phase, the calculation of fitness values ​​involves the calculation of the snowmelt rate, which is a key parameter for simulating snowmelt behavior in the development phase.

[0059] According to the fitness value of the individual, its position is continuously updated; the position represents the current weight distribution scheme of the individual;

[0060] Specifically, in this embodiment, the position update formula of each individual in the exploration phase is as follows:

[0061]

[0062] in, Indicates The individual in The position in the iteration, Indicates The step size and direction in the iteration, represents the average position of the entire population, is a random number, represents a random individual, It means multiplying rows. represents the current optimal solution;

[0063] The position update formula at the development stage is as follows:

[0064]

[0065] in, represents the snowmelt rate, Expressed as the degree-day factor in the degree-day method, represents the daily average temperature, Indicates from The random number selected in max represents the maximum number of iterations, Represents the operation of the current number of iterations and the maximum number of iterations. Indicates the current iteration number, represents the daily average temperature, represents the snowmelt rate, Indicates from A random number selected from .

[0066] When the maximum number of iterations is reached, the individual with the best fitness value is output as the optimal solution, which is the optimal weight allocation scheme.

[0067] Furthermore, the preset indicators include: root mean square error, mean absolute percentage error and mean absolute error.

[0068] Furthermore, constructing the dataset includes:

[0069] The glass factory DCS system is used to collect auxiliary variable data; the auxiliary variable data include: combustion working data;

[0070] The output accuracy of the dynamic prediction model is used as the screening criterion through the mutual information and mRMR method to screen the auxiliary variable data and determine the combustion working data as the input variable of the model;

[0071] SCR inlet NO x Concentration is used as the output variable. By analyzing the mutual information between the input variable and the output variable within a preset time, the mutual information of the two variables and the corresponding time delay are obtained;

[0072] Based on the corresponding time delay, the input variables are compensated for the time delay;

[0073] A data set is constructed based on the input variables and output variables after time lag compensation.

[0074] Specifically, in this embodiment, the data is collected by the glass factory DCS system, with 8600 sets of continuous data at intervals of 10 seconds. The unit load range of the data corresponding to the time sequence is 30% to 100% load segment, including load fluctuation segment and stable segment. Due to the influence of equipment performance, temperature, humidity, etc., the data collected by DCS is prone to abnormal values, noise points and individual missing.

[0075] The box plot method is used to identify outliers and remove them. The data gaps generated after removal are filled with the mean of the two data points before and after the point to ensure the integrity of the data time series. The same method is used to fill the missing values. For noise points, the sliding average method is used for filtering and noise reduction, and the window value size is set to 10. For each auxiliary variable, the numerical differences are large due to different measurement units, so all auxiliary variable data are normalized to the range of [0,1].

[0076] In this embodiment, 36 auxiliary variables were collected on-site in the DCS control room of the glass factory by consulting literature, among which the main ones were unit load, natural gas flow, primary air volume, etc. The output accuracy of the prediction model was used as the screening standard by the mutual information and (maximum relevance-minimum redundancy) mRMR method, and 6 auxiliary variables were determined as the final model input variables. xThe prediction model with concentration as the dominant variable. The selected variables are: unit load, natural gas flow, oxygen content in the kiln, average kiln temperature, flue gas oxygen content, total air volume, a total of 6 input variables. In order to ensure the consistency of the input and output of the system in timing, corresponding time lag compensation must be made for each input. Because mutual information can reflect the correlation between two variables, and when the input and output sequences are the same, the correlation between the two is the highest. Considering that NO is generated from boiler combustion, x It takes about 300 seconds for the flue gas to enter the SCR denitrification system, so after completing the screening of the auxiliary variables, the maximum delay time is set to 300s. By analyzing the mutual information of the input and output variables within 300 seconds, the mutual information of the two variables and the corresponding time delay are found, and the time lag at that time is compensated. The time delay compensation values ​​of each variable determined by the mutual information algorithm are shown in Table 1.

[0077] Table 1 Time delay compensation values ​​of various variables

[0078]

[0079] Specifically, in this embodiment, the steps for constructing the SAO-integrated weighted fusion dynamic prediction model are as follows:

[0080] First, a set of initial individuals are generated, each of which represents a weight distribution scheme in LightGBM, XGBoost, and CatBoost.

[0081] Then, the root mean square error (RMSE), mean absolute percentage error (MAPE), and mean absolute error (MAE) are used to evaluate the performance of each individual and calculate the fitness value. Then, according to the fitness value of the individual, its position is continuously updated, and the position represents the current weight distribution scheme of the individual. Finally, when the maximum number of iterations is reached, the individual with the best fitness value is output as the optimal solution, which is the weight of LightGBM, XGBoost, and CatBoost.

[0082] The dynamic prediction model is:

[0083]

[0084] in, k is the number of models used, G m is the predicted value corresponding to the integrated model, G i is the predicted value of the ith model, w i is the weight of the i-th surrogate model.

[0085] The LightGBM, XGBoost and CatBoost used in this example are as follows:

[0086] XGBoost is a gradient boosting tree algorithm. Compared with the original Gradient Boosting Decision Tree (GBDT), it has the following advantages: First, it combines the gradient boosting decision tree with a regularization term that can reduce the complexity of the model, which improves the accuracy and generalization of the model. Secondly, it speeds up the solution process by using a second-order Taylor expansion in the objective function. Finally, XGBoost learns a new decision tree in each iteration through gradient descent and accumulates it with the previous tree to obtain a strong classifier, achieving the goal of minimizing the loss function.

[0087] Below is the mathematical formula of the objective function:

[0088] (1)

[0089] In the formula, represents the loss function of the entire strong learner, represents the complexity of K weak learners in the strong learner.

[0090] After derivation and Taylor expansion, the specific formula of the objective function is as follows:

[0091] (2)

[0092] and are the first-order derivative and second-order derivative of the loss function respectively.

[0093] LightGBM is a machine learning algorithm based on the gradient boosted decision tree (GBDT). As an open source distributed gradient boosting framework, it can efficiently perform data modeling and prediction. Its calculation formula is:

[0094] (3)

[0095] The basic idea of ​​LightGBM is to gradually merge multiple weak regression trees to form a strong regression tree, thereby reducing the prediction error rate. Compared with XGBoost, it has the following advantages:

[0096] 1) LightGBM uses the Gradient-based One-Side Sampling (GOSS) algorithm. GOSS discards data instances with small gradients because data instances with small gradients will provide marginal contributions when estimating information gain. Through the GOSS algorithm, LightGBM reduces the number of data instances, improves the utilization rate of high-gradient data instances, and greatly reduces the sampling time.

[0097] 2) LightGBM introduces the Exclusive Feature Bundling (EFB) method, which aims to reduce the number of features by bundling mutually exclusive features together. In a general feature space, there are many mutually exclusive features that cannot have non-zero values ​​at the same time. Therefore, these mutually exclusive features can be reasonably bundled into a single feature to achieve feature dimensionality reduction, thereby improving computational efficiency. In other words, the EFB method reduces the number of features by merging mutually exclusive features while ensuring that the meaning of the features remains unchanged.

[0098] 3) LightGBM uses a histogram method to divide the feature values ​​into multiple small intervals for data segmentation and modeling on the decision tree. This method greatly reduces the calculation time and usually obtains better prediction results.

[0099] 4) LightGBM uses a leaf-wise growth algorithm. This algorithm can locate the leaves with the highest gain and split them, while limiting the leaves with low gain to avoid wasting resources. This leaf-oriented growth method can improve training efficiency. In addition, LightGBM can also limit the maximum depth of the leaves to reduce the risk of overfitting while ensuring efficiency.

[0100] CatBoost is a new machine learning model based on the GBDT algorithm, which outperforms XGBoost and LightGBM on some public datasets. The main advantage of CatBoost is that it effectively handles classification features, solves the problems of gradient bias and prediction bias, thus effectively avoiding overfitting problems and further improving accuracy and generalization ability.

[0101] When dealing with categorical features, CatBoost uses a numerical encoding method called "Ordered Target Statistics". , i=1, 2, ..., n are randomly sorted, and the new sorting sequence is For each sample’s category feature, the value of the feature is the previous The average value of the categorical feature for the records.

[0102] (4)

[0103] in represents the prior term, represents the weight coefficient of the prior term, and [·] represents the indicator function.

[0104] When dealing with the problem of gradient deviation and prediction deviation, CatBoost randomly shuffles the samples, and each sample only uses the samples in front of it for model training. Such a model is used to estimate the first-order and second-order gradients of the sample prediction results. These gradients are then used to build a tree structure, and the value of each leaf node of the final tree is calculated using all samples. This is the core idea of ​​Ordered Boosting, which can effectively reduce gradient estimation errors and alleviate prediction deviations. However, this method will increase the amount of calculations and affect the training speed.

[0105] In addition to the above three main outstanding issues, CatBoost also uses a symmetrical binary tree structure. This constraint on the tree structure has a certain regularization effect, which enables the CatBoost model to predict multiple samples at the same time, greatly shortening the prediction time. At the same time, in each iteration, CatBoost uses a greedy method to select the best features and split points to minimize the loss function. It effectively solves the exponential growth problem of feature combinations and improves the performance and accuracy of the model.

[0106] The snow melting optimization algorithm is a novel meta-heuristic optimization algorithm. Compared with other optimization algorithms, this algorithm achieves a balance between exploration and development, enabling it to discover potential and promising areas in the exploration phase and develop high-quality solutions around the current best solution in the development phase. The snow melting optimization algorithm simulates the evaporation and melting behavior of snow by introducing a dual population mechanism, allowing the algorithm to strike a balance between development and exploration.

[0107] Exploration Phase

[0108] When snow or liquid water converted from snow is converted into steam, it will show highly dispersed characteristics due to irregular motion. Brownian motion can simulate the diffusion behavior of steam in the search space. In standard Brownian motion, the step size is obtained based on the probability density function based on the normal distribution, with a mean of zero and a variance of one. Using dynamic and uniform step sizes in the algorithm, Brownian motion can explore some potential areas in the search space. In the SAO algorithm, the position update formula for each individual is as follows:

[0109] (5)

[0110] in, is the optimal solution including the current , the second best individual , the third best individual and the set of centroid positions of the leader individual group Randomly selected from Indicates The individual in The position in the iteration, Indicates The step size and direction in the iteration, represents the average position of the entire population, is a random number used to balance global and local searches.

[0111] By using Brownian motion to simulate the behavior of snow, the SAO algorithm is able to explore more potential solutions in the search space and maintain the diversity of the population, thereby improving the global search capability of the algorithm.

[0112] Development Phase

[0113] The development phase of the SAO algorithm simulates the behavior of snow melting into liquid water, and uses the degree-day method, one of the most classic snowmelt models, to reflect the snowmelt process. The position update formula is as follows:

[0114] (6)

[0115] in, represents the snowmelt rate, Expressed as the degree-day factor in the degree-day method, represents the daily average temperature, Indicates from The random number selected in. In this stage, the SAO algorithm focuses the search on high-quality solutions near the current optimal solution. It accelerates convergence by utilizing known high-quality solutions and further optimizes the quality of the solution.

[0116] Dual population mechanism

[0117] In the optimization algorithm, it is very important to achieve a balance between exploitation and exploration. The dual population mechanism is an optimization strategy in the SAO algorithm, which corresponds to the snow melting process. In the SAO algorithm, the entire population is randomly divided into two sub-populations Pa and Pb. Similar to the existence of liquid water and solid snow in the snow melting process, the process of melting snow into water is defined as the exploitation phase, while the process of evaporating water into steam and sublimating snow into steam is defined as the exploration phase. Therefore, the Pa sub-population is mainly responsible for exploration, and the Pb sub-population is mainly responsible for exploitation. The dual population mechanism operates as follows:

[0118] In the initialization phase season ;when When , then let , , until or The population iteration stops when .

[0119] in, Indicates the population size; Respectively represent the number of current Pa sub-population and Pb sub-population; represents the number of iterations, Indicates the maximum number of iterations.

[0120] In each iteration, by randomly dividing the entire population, individuals in the Pa subpopulation update their positions according to the position update equation (5) in the exploration phase. In this way, individuals in the Pa subpopulation tend to explore more potential optimal solutions in the search space. On the other hand, individuals in the Pb subpopulation update their positions according to the position update equation (6) in the development phase. This part includes the current best individual and the leader individual, so the individuals in the Pb subpopulation tend to move closer to the positions of the current best individual and the leader individual.

[0121] This embodiment analyzes the prediction results as follows:

[0122] The data used in the above method are 8600 sets of data sampled at 10s intervals provided by its DCS system. The unit load corresponding to the time series of the data used has a large range of changes, including load fluctuation segments and stable segments. The first 7000 sets of data are used as training sets, and the last 1600 sets are used as test sets.

[0123] The traditional model CatBoost and the SAO-integrated weighted fusion dynamic prediction model proposed in this embodiment are compared with the actual values. The comparison shows that the integrated model proposed in this embodiment after parameter optimization using the SAO algorithm has a better performance on NO than the traditional CatBoost model without optimization. x The predicted value of concentration is closer to the actual value, and the root mean square error and maximum error are better than those of the ordinary CatBoost model. The root mean square error (RMSE), mean absolute percentage error (MAPE) and mean absolute error (MAE) are shown in Table 2.

[0124] Table 2 Root mean square error (RMSE), mean absolute percentage error (MAPE) and mean absolute error (MAE) values

[0125]

[0126] The above are only preferred specific implementations of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed in the present application should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. NO of SAO-integrated weighted fusion dynamic prediction model x The prediction method is characterized in that include: Collect real-time combustion data of glass factories; The real-time combustion working data is input into the preset dynamic prediction model, and the SCR inlet NO is output. x Predicted concentration; wherein the dynamic prediction model is constructed based on the integrated weighted fusion of LightGBM sub-model, XGBoost sub-model and CatBoost sub-model, and is obtained through data set training; the data set includes: combustion working data after time lag compensation and the corresponding SCR inlet NO x concentration; the combustion working data include: unit load, natural gas flow, oxygen content in the kiln, average kiln temperature, flue gas oxygen content and total air volume; Constructing the dataset includes: The auxiliary variable data is collected by using the glass factory DCS system, wherein the auxiliary variable data includes: the combustion working data; The auxiliary variable data are screened by using the mutual information and mRMR method with the output accuracy of the dynamic prediction model as a screening criterion, and the combustion working data are determined as the input variable of the model; SCR inlet NO x The concentration is used as the output variable, and the mutual information of the input variable and the output variable within a preset time is analyzed to obtain the mutual information of the two variables and the corresponding time delay; Based on the corresponding time delay, the input variable is compensated for the time delay; constructing the data set based on the input variables and output variables after time lag compensation; Constructing the dynamic prediction model includes: The snow melting optimization algorithm is used to optimize the weights of the LightGBM sub-model, the XGBoost sub-model, and the CatBoost sub-model to obtain an optimal weight distribution scheme; Based on the optimal weight distribution scheme, the LightGBM sub-model, the XGBoost sub-model and the CatBoost sub-model are integrated and weighted fused to obtain the dynamic prediction model; The time lag compensation for the input variable includes: Performing a 215s time lag compensation on the unit load; Performing a time delay compensation of 166 seconds on the natural gas flow; Performing a time lag compensation of 105s on the oxygen content in the kiln; A time lag compensation of 148 seconds is performed on the average temperature of the kiln; Performing a time lag compensation of 121 seconds on the oxygen content of the flue gas; Performing a time lag compensation of 175 seconds on the total air volume; After collecting the auxiliary variable data, the method further includes: Using the box plot method to identify and remove abnormal values ​​in the auxiliary variable data; The data gaps after elimination are filled with the mean, and the sliding average method is used for filtering and noise reduction, and normalization is performed.

2. NO of the dynamic prediction model of SAO-integrated weighted fusion according to claim 1 x The prediction method is characterized in that Obtaining the optimal weight allocation scheme includes: Generate a set of initial individuals, each individual represents a weight allocation scheme in the LightGBM sub-model, the XGBoost sub-model, and the CatBoost sub-model; Based on the preset indicators, the performance of each individual is evaluated and the fitness value is calculated; According to the fitness value of the individual, its position is continuously updated; the position represents the current weight distribution scheme of the individual; When the maximum number of iterations is reached, the individual with the best fitness value is output as the optimal solution, which is the optimal weight allocation scheme.

3. NO of the dynamic prediction model of SAO-integrated weighted fusion according to claim 2 x The prediction method is characterized in that The preset indicators include: root mean square error, mean absolute percentage error and mean absolute error.

4. NO of the SAO-integrated weighted fusion dynamic prediction model according to claim 1 x The prediction method is characterized in that The dynamic prediction model is: in, k is the number of models used, G m is the predicted value corresponding to the integrated model, G i is the predicted value of the ith model, w i is the weight of the i-th surrogate model.

Images