A Multi-Output Prediction Method for NOx Formation Concentration and Ammonia Slip in Coal-fired Boilers Based on Deep Tree Ensemble Model

Abstract

This invention discloses a multi-output prediction method for NOx generation concentration and ammonia slip in coal-fired boilers based on a deep tree ensemble model, relating to the field of NOx emission control technology for coal-fired boilers. The method includes the following steps: S100, data processing; acquiring data, processing the data, and outputting the results; S200, using the output of step S100 as model input, and taking NOx concentration and ammonia slip concentration as outputs, constructing a multi-output prediction model for NOx concentration and ammonia slip concentration based on a deep tree ensemble. This invention overcomes the influence of large delays and hysteresis in the SCR denitrification system of coal-fired boilers through time delay analysis and the construction of time series data; it uses a deep tree ensemble model to predict NOx generation concentration and ammonia slip in coal-fired boilers, learns information from the complete dataset through a sliding window, and obtains new datasets with different feature scales for training through multi-granularity scanning; thus improving the prediction accuracy and generalization ability of the model for predicting NOx generation concentration and ammonia slip multi-outputs.

Description

Technical Field

[0001] This invention belongs to the field of NOx emission control technology for coal-fired boilers, specifically involving a method for predicting NOx generation concentration and ammonia slip output of coal-fired boilers based on a deep tree ensemble model. Background Technology

[0002] With increasingly stringent environmental protection requirements, NOx emission control from coal-fired boilers has become particularly important. Selective Catalytic Reduction (SCR) systems are one of the most widely used denitrification methods for coal-fired boilers. This process uses ammonia (NH3) as a reducing agent, reacting it with NOx in the flue gas to reduce it into N2 and H2O, thereby reducing NOx emissions. In actual operation, the denitrification efficiency and ammonia slip rate of the SCR system are two key indicators for evaluating its performance. However, to meet stringent NOx emission standards, it is often necessary to increase the ammonia injection rate to improve denitrification efficiency, but this may also lead to excessive ammonia slip rates. Excessively high ammonia slip rates can cause a series of problems, such as reacting with SO3 in the flue gas to form ammonium bisulfate, leading to air preheater blockage, decreased dust removal efficiency, and catalyst damage, seriously affecting the safety and economy of the unit.

[0003] Continuous emission monitoring systems are installed at the inlet and outlet of the SCR denitrification system to continuously monitor the NOx concentration entering the system, which is the basis for ammonia injection control. However, actual continuous emission monitoring systems experience a time delay due to the measurement method itself, resulting in a slow response and an inability to promptly reflect rapid changes in system operating status. This leads to inaccurate control of the ammonia-to-nitrogen ratio in the SCR system. Insufficient ammonia injection may result in incomplete NOx removal, with the NOx concentration at the outlet failing to meet emission standards, leading to unstable denitrification efficiency. Conversely, excessive ammonia injection may cause the ammonia slip rate in the denitrification system to exceed limits.

[0004] For example, the Chinese invention application with publication number CN116741301A, entitled "A Method and System for Real-time Prediction of NOx Generation Concentration in Coal-fired Boilers," obtains multiple independent variable features of NOx generation concentration in the boiler, uses the ammonia injection time of the boiler's denitrification system as the current time, constructs multiple NOx concentration prediction data sets with time series, trains multiple NOx concentration prediction data sets on a linear model based on the least squares method, constructs a NOx generation concentration prediction model, and uses the NOx generation concentration prediction model to predict the NOx generation concentration at the ammonia injection time in real time.

[0005] For example, the Chinese invention patent with announcement number CN109304086B, entitled "A Refined Ammonia Injection Control Method for SCR Denitrification in Power Plant Boilers," describes a method that combines a long short-term memory neural network model with an intelligent optimization algorithm to control the opening of the main valve, and controls the opening of the sub-valve based on the detection data from multiple NOx detection points. The method predicts the outlet NOx concentration at the next moment using the long short-term memory neural network model, and iteratively updates the optimization index to obtain the optimal ammonia injection strategy.

[0006] For example, Chinese invention patent CN113433911B, entitled "Precise Control System and Method for Ammonia Injection in Denitrification Unit Based on Accurate Concentration Prediction," describes a system comprising a power plant information system, a NOx concentration prediction model at the denitrification unit inlet, a multi-model predictive control module, and a controlled object for the denitrification unit. This method, designed for CFB boiler units, clusters operating data to establish a global LSTM neural network prediction model adaptable to a wide range of variable loads and operating conditions in the boiler, enabling accurate prediction of the NOx concentration at the denitrification system inlet.

[0007] For example, the Chinese invention application with publication number CN116417094A, entitled "An Intelligent Ammonia Injection Denitrification Prediction System Based on a Time-Series Neural Network Model," discloses that feature selection uses a LASSO regression model to eliminate collinearity in the model, the prediction model adopts feature-based difference and baseline-based models, the model uses a long short-term memory network as the main body in the non-backflushing stage, and a time-series Transformer model in the backflushing stage, and ablation experiments are used to determine the importance of features.

[0008] For example, the Chinese invention patent with announcement number CN109062053B and invention title "A Denitrification Ammonia Injection Control Method Based on Multivariate Correction" targets the denitrification system of thermal power generating units. It uses SPSS to screen and process the operating information data of the generating unit to extract the principal components, uses an SVM prediction model to predict the NOx content at the inlet of the SCR denitrification system at the current moment, and performs feedforward control and prediction correction of the ammonia injection quantity based on the predicted NOx content at the inlet of the SCR system and the measurement data to generate the ammonia injection quantity control command at the current moment, control the ammonia injection regulating valve, and adjust the ammonia injection quantity.

[0009] For example, the Chinese invention application publication number CN117270387A, entitled "A Method and System for Controlling Low Ammonia Slip in an SCR Denitrification System Based on Deep Learning," utilizes the maximum information coefficient (MIC) to determine the delay time m. Based on the delay time, the corrected NOx concentration value, and other relevant operational data, training data is constructed to train a deep learning model (LSTM) to predict the outlet NOx concentration. The predicted outlet NOx concentration at time t+m is then corrected based on the difference between the predicted and actual concentration values ​​from the previous time step. Based on the corrected prediction and the economic benefits of the power plant, a multi-objective optimization function is constructed. This function is then solved using a differential evolution algorithm to obtain the target ammonia injection rate at time t+m, thereby reducing ammonia slip.

[0010] For example, Chinese invention patent CN107158946B, entitled "A Method for Real-Time Online Prediction and Control of Ammonia Slip Concentration," discloses a method for real-time online prediction and control of ammonia slip concentration in SCR denitrification systems of coal-fired power units. This method uses the potential of the denitrification reactor, which is tested periodically on-site, to predict the ammonia slip concentration. It can predict the actual ammonia slip concentration in real time based on the flue gas conditions at the SCR reactor inlet and the operating denitrification efficiency. Based on this, it can reasonably control the ammonia injection flow rate and denitrification efficiency, and provide timely warnings to improve denitrification performance, control and reduce ammonia slip concentration, and mitigate the impact of ammonium bisulfate blockage on the downstream air preheater.

[0011] For example, the Chinese invention application with publication number CN112461995A, entitled "A Method for Predicting Ammonia Slip in Thermal Power Plants," discloses the measured NOx concentrations at the inlet and outlet of denitrification reactor A, the NOx concentrations at the inlet and outlet of denitrification reactor B, the ammonia slip at the outlet of denitrification reactor A, and the ammonia slip at the outlet of denitrification reactor B. When all measured NOx concentrations meet the requirements, DCS data is collected. The collected DCS data is preprocessed to obtain standard data, which is then divided into training data and validation data. The training data is used as input, and the ammonia slip data is used as output data to train ammonia slip prediction model A and ammonia slip prediction model B. The trained models are then used to predict the real-time ammonia slip concentration.

[0012] There are two main ways to obtain NOx concentration and ammonia slip concentration in existing coal-fired boiler SCR denitrification systems: (1) monitoring NOx concentration and ammonia slip rate through a continuous emission monitoring system (CEMS) and ammonia slip meter installed in the flue. (2) predicting NOx concentration using a machine learning model, with NOx concentration as the target output of the training model or with the slip rate as the target output of the training model to predict ammonia slip concentration.

[0013] Existing methods for predicting NOx concentration and ammonia slip concentration are prone to delays and low accuracy due to their measurement methods. Most existing machine learning prediction methods for NOx concentration or ammonia slip rate focus on a single output, such as NOx concentration or ammonia slip concentration, and lack effective methods for simultaneously predicting multiple related outputs, such as NOx concentration and ammonia slip concentration. Therefore, in order to solve the technical problems existing in the above-mentioned technologies, it is necessary to develop a multi-output prediction method for NOx generation concentration and ammonia slip in coal-fired boilers based on a deep tree ensemble model. Summary of the Invention

[0014] The purpose of this invention is to provide a method for predicting NOx generation concentration and ammonia escape multiple outputs of coal-fired boilers based on a deep tree ensemble model, in order to solve the aforementioned technical problems. This method uses a deep tree ensemble model to predict NOx generation concentration and ammonia escape in coal-fired boilers. It learns information from the complete dataset through a sliding window and acquires new datasets with different feature scales through multi-granularity scanning for training. Each layer of the deep forest includes both random forest and fully random forest models. Through forest cascading, each level receives feature information processed by the previous level and outputs its processing results to the next level, improving the prediction accuracy and generalization ability of the model for NOx generation concentration and ammonia escape multiple outputs.

[0015] To achieve the above-mentioned objectives, the technical solution adopted by the present invention is as follows:

[0016] A multi-output prediction method for NOx formation concentration and ammonia slip in coal-fired boilers based on a deep tree ensemble model includes the following steps:

[0017] S100, Data Processing; Acquire data, process the data, and then output the data;

[0018] S200: Using the output of step S100 as model input and NOx concentration and ammonia slip concentration as output, construct a deep tree ensemble multi-output prediction model for NOx concentration and ammonia slip concentration. The specific processing steps are as follows:

[0019] S201; The original input feature variables are first subjected to multi-granularity scanning: The first scanning window has n-dimensional features. After sliding the window, a new matrix is ​​obtained, which is then processed through a random forest and a fully random forest model to obtain the enhanced feature matrix A; The second scanning window has 2n-dimensional features. After sliding the window, a new matrix is ​​obtained, which is then processed through a random forest and a fully random forest model to obtain the enhanced feature matrix B; The third scanning window has 3n-dimensional features. After sliding the window, a new matrix is ​​obtained, which is then processed through a random forest and a fully random forest model to obtain the enhanced feature matrix C. Finally, three sets of enhanced feature matrices A, B, and C with different granularities are obtained.

[0020] S202; The eight prediction results obtained after the enhanced feature matrix A is processed through two random forests and two fully random forests, along with the enhanced feature matrix A, are input into layer 1. A ,

[0021] S203; Passing through layer 1 A The eight predicted results, along with the enhanced feature matrix B, are input into layer 1. B ,

[0022] S204; Passing through layer 1 B The eight predicted results, along with the enhanced feature matrix C, are input into layer 1. C ,

[0023] S205; Repeat steps S2, S3, and S4 until layer N is reached. C Complete the forest cascade.

[0024] S206; Prediction result calculation: Layer N C The results were then processed through two random forests and two completely random forests to obtain eight prediction results. The average of the four prediction results for NOx generation concentration is the final predicted NOx generation concentration, and the average of the four prediction results for ammonia escape concentration is the final predicted ammonia escape concentration.

[0025] Preferably, the scanning window size n, 2n, 3n and the number of layers N used in the deep tree-integrated NOx concentration and ammonia escape concentration multi-output prediction model are the hyperparameters used for training the model.

[0026] Preferably, in step S201, after sliding the window with different granularities, a dataset with a different number of input feature variables than the original dataset will be obtained. The original dataset has X features. After sliding the window with Y features as the window size, (X-Y+1) datasets containing Y features can be created, where Y is less than X.

[0027] Preferably, each layer of the deep forest in step S200 includes two models: random forest and fully random forest.

[0028] Preferably, the steps of acquiring data, processing the data, and then outputting the data include the following:

[0029] S101, data for model training is obtained through data monitored by the continuous emission monitoring system and ammonia slip instrument installed in the flue;

[0030] S102, Process the outliers in the acquired data, including removing purging conditions and removing outliers using the 3σ rule, that is, identifying data points with values ​​outside the range of [μ-3σ,μ+3σ] as outliers and removing them, where μ is the mean of the data and σ is the standard deviation of the data;

[0031] S103, after outlier processing, feature variables were selected. A set of feature variables was selected as the first feature variable group (Group1) based on NOx concentration as the target and combined with the maximum information coefficient and mechanism analysis. A set of feature variables was selected as the second feature variable group (Group2) based on ammonia escape concentration as the target and combined with the maximum information coefficient and mechanism analysis.

[0032] S104, the first feature variable group Group1 and the second feature variable group Group2 obtained in step S103 are fused to form a new feature variable group as the third feature variable group Group3;

[0033] S105, perform time lag analysis on the characteristic variable parameters of the selected third characteristic variable group Group3, and use the MIC method to calculate the correlation between each characteristic variable parameter in the third characteristic variable group Group3 after time shift and the NOx emission concentration, obtain the time shift value at the maximum correlation, and use it as the lag time of the variable. Use the data after time lag alignment as the data at time t.

[0034] S106, construct time series data containing the previous k time steps from the new data at time t after time delay alignment;

[0035] S107, standardize the constructed time series input and output data before outputting, that is, standardize the input and output data of the prediction model.

[0036] Preferably, the determination of purging conditions follows the formula below:

[0037] ;

[0038] In the formula, The measurement at the inlet of the SCR denitrification system at time i concentration; The measurement at the inlet of the SCR denitrification system at time i+j Concentration, where j = 1, 2, 3... 20; To determine the inlet of the SCR denitrification system during purging Minimum change in concentration.

[0039] Preferably, in step S107, Z-Score normalization is used, and its calculation formula is as follows:

[0040] ;

[0041] In the formula, These are the normalized data values. For the input data, The mean of the data. denoted as the standard deviation of the data.

[0042] Preferably, the SCR denitrification system inlet for purging in step S102 is determined. Minimum change in concentration The value is determined by empirical historical data measured in actual power plants under actual purging conditions.

[0043] Preferably, in step S103, the feature variable selection analysis targeting NOx concentration first uses the maximum information coefficient to analyze the correlation between operating variables and NOx emission concentration, and selects variables with higher correlation coefficients as data features. Then, the denitrification reaction mechanism is analyzed as mechanistic features. Finally, the determined mechanistic features and data features are merged together as the first feature variable group Group1 targeting NOx concentration. In step S103, the feature variable selection analysis targeting ammonia slip concentration first uses the maximum information coefficient to analyze the correlation between operating variables and ammonia slip concentration, and selects variables with higher correlation coefficients as data features. Then, the feature variables affecting ammonia slip concentration in the denitrification reaction mechanism of the SCR system are analyzed as mechanistic features. Finally, the determined mechanistic features and data features are merged together as the second feature variable group Group2 targeting ammonia slip concentration.

[0044] Preferably, in the time delay analysis of step S105, the process time m from fuel input to final combustion and flue gas discharge is first calculated based on actual boiler operating experience. Then, for each variable in the third characteristic variable group Group3 after feature fusion... Take time t, time t-1, time t-2 up to time tm to form a new variable. Time series matrix Calculate the maximum mutual information coefficient (MIC) between the matrix and the predicted target NOx formation concentration, and take the value at the moment when the MIC is the largest. As the time-delay aligned input of this variable at time t, where .

[0045] Preferably, the data from the first k time steps in step S106 are used, where k is the largest value among the variables finally determined in the time delay analysis in step S105. value.

[0046] Preferably, the feature variable fusion operation in step S104 includes deleting duplicate feature variables.

[0047] This application has achieved beneficial technical effects:

[0048] This invention overcomes the effects of large delays and lags in SCR denitrification systems of coal-fired boilers by using time-delay analysis and constructing time-series data. It employs a deep tree ensemble model to predict NOx generation concentration and ammonia slip in coal-fired boilers, learning information from the complete dataset through a sliding window and acquiring new datasets with different feature scales for training through multi-granularity scanning. Each layer of the deep forest includes both random forest and fully random forest models. Forest cascading allows each level to receive feature information from the previous level and output its results to the next level, improving the prediction accuracy and generalization ability of the model in predicting NOx generation concentration and ammonia slip. Attached Figure Description

[0049] Figure 1 The diagram shows the flowchart for establishing the NOx generation concentration and ammonia escape multi-output prediction model for coal-fired boilers according to the present invention.

[0050] Figure 2 The diagram shown illustrates the principle of the model training and prediction process of the deep tree ensemble model used in this invention. Detailed Implementation

[0051] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the specific implementation methods of the present invention will be described below with reference to the accompanying drawings. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings and other implementation methods can be obtained based on these drawings without any creative effort.

[0052] The technical solution of the present invention will be described in detail below with specific embodiments.

[0053] Reference Figure 1 , Figure 2 A multi-output prediction method for NOx generation concentration and ammonia slip in coal-fired boilers based on a deep tree ensemble model is proposed. This method applies the deep forest ensemble model to predict NOx generation concentration and ammonia slip concentration in an SCR denitrification system, and includes the following steps:

[0054] Step 1: Data Acquisition; Data for model training is obtained through data monitored by the Continuous Emission Monitoring System (CEMS) and ammonia slip meter installed in the flue.

[0055] Step 2: Outlier Handling; Outliers in the acquired data are processed, primarily including removing purging conditions and using the 3σ rule to remove outliers. The determination of purging conditions follows the formula below:

[0056] ;

[0057] In the formula, The measurement at the inlet of the SCR denitrification system at time i concentration; The measurement at the inlet of the SCR denitrification system at time i+j Concentration, where j = 1, 2, 3... 20; To determine the inlet of the SCR denitrification system during purging Minimum change in concentration.

[0058] Among them, the inlet of the SCR denitrification system to be purged is determined. Minimum change in concentration The value is determined by empirical historical data measured in actual power plants under actual purging conditions.

[0059] The 3σ rule for outlier removal specifically involves identifying data points whose values ​​fall outside the range of [μ-3σ, μ+3σ] as outliers and removing them, where μ is the mean of the data and σ is the standard deviation of the data.

[0060] Step 3: Feature variable selection; Feature variables are selected for the data after outlier processing. A set of feature variables is selected as the first feature variable group (Group1) based on NOx concentration as the target and combined with the maximum information coefficient (MIC) and mechanism analysis. A set of feature variables is selected as the second feature variable group (Group2) based on ammonia escape concentration as the target and combined with the maximum information coefficient (MIC) and mechanism analysis.

[0061] The feature variable selection analysis targeting NOx concentration first uses the maximum information coefficient to analyze the correlation between operating variables and NOx emission concentration, selecting variables with higher correlation coefficients as data features. Then, the denitrification reaction mechanism is analyzed as mechanistic features. Finally, the determined mechanistic features and data features are merged to form the first feature variable group, Group1, targeting NOx concentration. Similarly, the feature variable selection analysis targeting ammonia slip concentration first uses the maximum information coefficient to analyze the correlation between operating variables and ammonia slip concentration, selecting variables with higher correlation coefficients as data features. Then, the feature variables affecting ammonia slip concentration in the SCR system's denitrification reaction mechanism are analyzed as mechanistic features. Finally, the determined mechanistic features and data features are merged to form the second feature variable group, Group2, targeting ammonia slip concentration.

[0062] Step 4: Feature Variable Fusion; The first feature variable group (Group1) and the second feature variable group (Group2) obtained in Step 3 are fused to form a new feature variable group (Group3). The main operation of feature variable fusion is to remove duplicate feature variables.

[0063] Step 5: Time lag analysis; Perform time lag analysis on the selected characteristic variable parameters, and use the MIC method to calculate the correlation between each characteristic variable parameter in the third characteristic variable group (Group3) after time shift and the NOx emission concentration. Obtain the time shift value at which the maximum correlation is obtained, which is used as the lag time of the variable. Use the data after time lag alignment as the data at time t.

[0064] In the time-delay analysis, the process time *m* from fuel input to final combustion and flue gas discharge is first calculated based on actual boiler operating experience. Then, for each variable in the third characteristic variable group (Group3) after feature fusion... Take time t, time t-1, time t-2 up to time tm to form a new variable. Time series matrix Calculate the maximum mutual information coefficient (MIC) between the matrix and the predicted target 1, where the predicted target 1 is specifically the NOx formation concentration at time t+1. The value of the MIC at the time with the largest MIC is taken. As the time-delay aligned input of this variable at time t, where .

[0065] Step 6: Construct time series data; construct time series data containing the previous k time steps from the new data at time t after time delay alignment. In the data at the previous k time steps, k is the largest value among the variables finally determined in the time delay analysis in step 5. value.

[0066] Step 7: Data Standardization; Standardize the constructed time series input and output data using Z-Score normalization, the calculation formula of which is as follows:

[0067] ;

[0068] In the formula, These are the normalized data values. For the input data, The mean of the data. denoted as the standard deviation of the data.

[0069] Step 8: Prediction Model Construction; Construct a deep tree ensemble multi-output prediction model for NOx concentration and ammonia slip concentration, using NOx concentration and ammonia slip concentration as outputs, and the normalized third feature variable group (Group3) data after steps 5, 6, and 7 as inputs. Simultaneously, predict NOx concentration and ammonia slip concentration. The main calculation steps of this model are as follows:

[0070] S1; The original input feature variables are first scanned at multiple granularities: The first scanning window has n-dimensional features. After sliding the window, a new matrix is ​​obtained, which is then processed through a random forest and a fully random forest model to obtain the enhanced feature matrix A; the second scanning window has 2n-dimensional features. After sliding the window, a new matrix is ​​obtained, which is then processed through a random forest and a fully random forest model to obtain the enhanced feature matrix B; the third scanning window has 3n-dimensional features. After sliding the window, a new matrix is ​​obtained, which is then processed through a random forest and a fully random forest model to obtain the enhanced feature matrix C. Finally, three sets of enhanced feature matrices A, B, and C with different granularities are obtained.

[0071] S2; The eight prediction results obtained after the enhanced feature matrix A is processed through two random forests and two fully random forests, along with the enhanced feature matrix A, are input into layer 1. A .

[0072] S3; Passing through layer 1 A The eight predicted results, along with the enhanced feature matrix B, are input into layer 1. B .

[0073] S4; after layer 1 B The eight predicted results, along with the enhanced feature matrix C, are input into layer 1. C .

[0074] S5; Repeat steps S2, S3, and S4 until layer N is reached. C Complete the forest cascade.

[0075] S6; Prediction result calculation: Layer N C The results were then processed through two random forests and two completely random forests to obtain eight prediction results. The average of the four prediction results for NOx generation concentration is the final predicted NOx generation concentration, and the average of the four prediction results for ammonia escape concentration is the final predicted ammonia escape concentration.

[0076] The scanning window sizes n, 2n, 3n and the number of layers N used in the prediction model in step 8 are the hyperparameters used in training the model.

[0077] In step S1 of step 8, after sliding the window with different granularities, a dataset with a different number of features than the original input will be obtained. For example, if the original dataset has 100 features, sliding with a window size of 20 features will create 81 datasets containing 20 features; sliding with a window size of 40 features will create 61 datasets containing 40 features; and sliding with a window size of 60 features will create 41 datasets containing 60 features. That is, if the original dataset has X features, sliding with a window size of Y features will create (X-Y+1) datasets containing Y features, where Y is less than X.

[0078] Each layer of the deep forest in step 8 includes two models: random forest and fully random forest. By default, each model includes 500 weak classifiers. That is, the random forest model contains 500 decision tree models, while the fully random forest model contains 500 fully random decision trees.

[0079] This invention presents a method for predicting NOx formation concentration and ammonia slip multiple outputs in coal-fired boilers based on a deep tree ensemble model. Its purpose is to address the problems of existing NOx and ammonia slip measurement methods in coal-fired power plant denitrification systems, which suffer from delays and low accuracy due to measurement limitations, and the lack of effective methods for simultaneously predicting these two crucial SCR denitrification system indicators. The method overcomes the large delays and lags of coal-fired boiler SCR denitrification systems by using time-delay analysis and constructing time-series data. It learns information from the complete dataset through a sliding window and acquires new datasets with different feature scales for training through multi-granularity scanning. Each layer of the deep forest includes both random forest and fully random forest models. Forest cascading allows each level to receive feature information processed by the previous level and output its results to the next level, further improving the model's prediction accuracy and generalization ability for NOx formation concentration and ammonia slip multiple outputs. This provides data support for precise ammonia injection control in SCR systems, ensuring the safe, stable, and environmentally friendly operation of the denitrification system.

[0080] Accurate real-time NOx concentration prediction is key to achieving accurate ammonia injection control. In actual operation, NOx concentration and ammonia slip concentration are correlated. Therefore, developing a multi-output prediction method that can simultaneously predict NOx concentration and ammonia slip concentration is of great significance for optimizing the operation of SCR system, reducing operating costs, and improving environmental benefits.

[0081] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0082] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this invention patent should be determined by the appended claims.

[0083] The embodiments of the multi-output prediction method for NOx formation concentration and ammonia slip in coal-fired boilers based on a deep tree ensemble model provided by the present invention have been described in detail above. Specific examples have been used to illustrate the principles and implementation methods of the present invention, and the descriptions of the embodiments above are only for the purpose of helping to understand the core ideas of the present invention. It should be noted that those skilled in the art can make several improvements and modifications to the present invention without departing from the principles of the present invention, and these improvements and modifications also fall within the protection scope of the claims of the present invention.

Claims

1. A method for predicting NOx formation concentration and ammonia slip multiple outputs in coal-fired boilers based on a deep tree ensemble model, characterized in that, Includes the following steps, S100, Data Processing; Acquire data, process the data, and then output the data; S200: Using the output of step S100 as model input and NOx concentration and ammonia slip concentration as output, construct a deep tree ensemble multi-output prediction model for NOx concentration and ammonia slip concentration. The specific processing steps are as follows: S201; The original input feature variables are first subjected to multi-granularity scanning: The first scanning window has n-dimensional features. After sliding the window, a new matrix is ​​obtained, which is then processed through a random forest and a fully random forest model to obtain the enhanced feature matrix A; The second scanning window has 2n-dimensional features. After sliding the window, a new matrix is ​​obtained, which is then processed through a random forest and a fully random forest model to obtain the enhanced feature matrix B; The third scanning window has 3n-dimensional features. After sliding the window, a new matrix is ​​obtained, which is then processed through a random forest and a fully random forest model to obtain the enhanced feature matrix C. Finally, three sets of enhanced feature matrices A, B, and C with different granularities are obtained. S202； The eight predictions obtained after passing the enhanced feature matrix A through two random forests and two fully random forests, along with the enhanced feature matrix A itself, are input into layer 1. A , S203; Passing through layer 1 A The eight predicted results, along with the enhanced feature matrix B, are input into layer 1. B , S204; Passing through layer 1 B The eight predicted results, along with the enhanced feature matrix C, are input into layer 1. C , S205; Repeat steps S2, S3, and S4 until layer N is reached. C Complete the forest cascade. S206; Prediction result calculation: Layer N C The results were then processed through two random forests and two completely random forests to obtain eight prediction results. The average of the four prediction results for NOx generation concentration is the final predicted NOx generation concentration result, and the average of the four prediction results for ammonia escape concentration is the final predicted ammonia escape concentration result. The process of acquiring data, processing the data, and then outputting the data includes the following steps. S101, data for model training is obtained through data monitored by the continuous emission monitoring system and ammonia slip instrument installed in the flue; S102, Process the outliers in the acquired data, including removing purging conditions and removing outliers using the 3σ rule, that is, identifying data points with values ​​outside the range of [μ-3σ,μ+3σ] as outliers and removing them, where μ is the mean of the data and σ is the standard deviation of the data; S103, after outlier processing, feature variables were selected. A set of feature variables was selected as the first feature variable group (Group1) based on NOx concentration as the target and combined with the maximum information coefficient and mechanism analysis. A set of feature variables was selected as the second feature variable group (Group2) based on ammonia escape concentration as the target and combined with the maximum information coefficient and mechanism analysis. S104, the first feature variable group Group1 and the second feature variable group Group2 obtained in step S103 are fused to form a new feature variable group as the third feature variable group Group3; S105, perform time lag analysis on the characteristic variable parameters of the selected third characteristic variable group Group3, and use the MIC method to calculate the correlation between each characteristic variable parameter in the third characteristic variable group Group3 after time shift and the NOx emission concentration, obtain the time shift value at the maximum correlation, and use it as the lag time of the variable. Use the data after time lag alignment as the data at time t. S106, construct time series data containing the previous k time steps from the new data at time t after time delay alignment; S107, after standardizing the constructed time series input and output data, output the results.

2. The method for predicting NOx formation concentration and ammonia slip output from a coal-fired boiler according to claim 1, characterized in that, The scanning window size n, 2n, 3n and the number of layers N used in the deep tree-integrated NOx concentration and ammonia escape concentration multi-output prediction model are hyperparameters used in training the model.

3. The method for predicting NOx formation concentration and ammonia slip output from a coal-fired boiler according to claim 1, characterized in that, In step S201, after sliding the window with different granularities, a dataset with a different number of features than the original input will be obtained. The original dataset has X features. After sliding the window with Y features as the window size, (X-Y+1) datasets containing Y features can be created, where Y is less than X.

4. The method for predicting NOx formation concentration and ammonia slip output from a coal-fired boiler according to claim 1, characterized in that, Each layer of the deep forest in step S200 includes two models: random forest and fully random forest.

5. The method for predicting NOx formation concentration and ammonia slip output from a coal-fired boiler according to claim 1, characterized in that, The following formula is used to determine the purging condition: ； In the formula, The measurement at the inlet of the SCR denitrification system at time i concentration; The measurement at the inlet of the SCR denitrification system at time i+j Concentration, where j = 1, 2, 3... 20; To determine the inlet of the SCR denitrification system during purging Minimum change in concentration.

6. The method for predicting NOx formation concentration and ammonia slip output from a coal-fired boiler according to claim 1, characterized in that, In step S107, Z-Score normalization is used, and its calculation formula is as follows: ； In the formula, These are the normalized data values. For the input data, The mean of the data. denoted as the standard deviation of the data.

7. The method for predicting NOx formation concentration and ammonia slip output from a coal-fired boiler according to claim 1, characterized in that, In step S102, the inlet of the SCR denitrification system being purged is determined. Minimum change in concentration The value is determined by empirical historical data measured in actual power plants under actual purging conditions.

8. The method for predicting NOx formation concentration and ammonia slip output from a coal-fired boiler according to claim 1, characterized in that, The characteristic variable selection analysis targeting NOx concentration in step S103 first uses the maximum information coefficient to analyze the correlation between operating variables and NOx emission concentration, and selects variables with higher correlation coefficients as data features. Then, the denitrification reaction mechanism is analyzed as mechanistic features. Finally, the determined mechanistic features and data features are merged together as the first characteristic variable group Group1 targeting NOx concentration. The characteristic variable selection analysis targeting ammonia slip concentration in step S103 first uses the maximum information coefficient to analyze the correlation between operating variables and ammonia slip concentration, and selects variables with higher correlation coefficients as data features. Then, the characteristic variables affecting ammonia slip concentration in the denitrification reaction mechanism of the SCR system are analyzed as mechanistic features. Finally, the determined mechanistic features and data features are merged together as the second characteristic variable group Group2 targeting ammonia slip concentration.

9. The method for predicting NOx formation concentration and ammonia slip output from a coal-fired boiler according to claim 1, characterized in that, In the time-delay analysis of step S105, the process time m from fuel input to final combustion and flue gas discharge is first calculated based on actual boiler operation experience. Then, for each variable in the third characteristic variable group Group3 after feature fusion... Take time t, time t-1, time t-2 up to time tm to form a new variable. Time series matrix Calculate the maximum mutual information coefficient (MIC) between the matrix and the predicted target NOx formation concentration, and take the value at the moment when the MIC is the largest. As the time-delay aligned input of this variable at time t, where .

Images