CN116933926A - AE-GAN-based SCR inlet NOx emission prediction method - Google Patents

Abstract

The invention discloses an SCR inlet NOx emission prediction method based on AE-GAN, which comprises the following steps: collecting an SCR inlet NOx value of a power plant and corresponding sensitive parameters related to the NOx value, and constructing a data set; preprocessing a data set in a mutual information sliding time step mode; (3) constructing an AE-GAN network: taking sensitive parameters related to the NOx value as input, performing dimension reduction processing through an AE network, wherein the output of the AE network is taken as the input of a GAN network, and the output of the GAN network is the NOx value; training the AE-GAN network by using the preprocessed data set to obtain an AE-GAN network prediction model; (4) Sensitive parameters related to the NOx value in the SIS database are collected, preprocessed in a mutual information sliding time step mode, input into an AE-GAN network prediction model, and real-time prediction is carried out on the NOx emission amount to output the NOx value. The method provided by the invention predicts the NOx emission in real time, and can improve the prediction accuracy of the SCR inlet NOx emission.

Description

AE-GAN-based SCR inlet NOx emission prediction method

Technical Field

The invention belongs to the technical field of SCR flue gas denitration of coal-fired power plants, and particularly relates to an SCR inlet NOx emission prediction method based on an AE-GAN network.

Background

According to the statistical data of the national electric power industry published by the national energy bureau in 2022 years, the total capacity of the thermal power installation of China accounts for about 52.03 percent by the end of 2022 years, and the energy structure is mainly characterized by rich coal, gas deficiency and less oil, so that in a quite long time in the future, the coal-fired power generation is still one of the main energy sources of China. Currently, there are three main methods for obtaining NOx concentration: (1) The continuous monitoring system for the exhaust gas emission monitors gases such as NOx, but has the advantages of high price, short service life and great delay of NOx concentration signals. (2) Based on computational fluid mechanics and chemical mechanism, building corresponding physical and chemical models. However, the generation mechanism of combustion pollutants in a boiler is very complex, and a relatively accurate model is difficult to generate. (3) The real-time concentration of NOx is predicted based on methods such as artificial intelligence, data mining and the like by analyzing thermal parameters affecting the generation of NOx. The soft measurement mode can better ensure real-time performance, is free from maintenance and has higher precision.

Many researchers solve the problem of complex NOx generation mechanism by adopting a machine learning black box algorithm, the traditional machine learning algorithm is mainly an SVM, common PLS can be adopted to perform feature extraction pretreatment on data on the boiler side, the problem of multiple linearity of independent variables is solved, and an LS-SVM model is established to obtain good generalization effect, such as the Chinese patent with publication number of CN111860923A, and a boiler flue gas NOx emission prediction algorithm based on multi-model clustering integration is disclosed. Dividing a data space according to the output NOx emission amount; determining the variables participating in clustering through variable weights based on correlation analysis and hierarchical clustering based on information entropy; clustering by using a proposed multi-model cluster integration (VMSC) algorithm to obtain a membership matrix of each subspace; and integrating a least square support vector machine (LS-SVM) model of each subspace by adopting a least square method integrating membership degrees, so as to obtain the NOx emission of the boiler flue gas. But the parameter optimization of the SVM is often difficult to determine. Along with the popular application of the XGBoost algorithm in the known algorithm competition Kagle, the XGBoost combined model also forms a big hot spot for predicting the concentration of NOx, and the XGBoost combined model is combined with PLS to perform characteristic dimension reduction. However, the concentration of NOx measured by the method has a certain delay with the process parameters of the boiler side, and the acquisition and calibration of the delay relation are difficult. And the selected data set is usually under a steady-state working condition, so that the prediction effect on the condition of abrupt change of the working condition is poor.

With the digital development of power plants in recent years, more and more data centers of the power plants are perfected, and the data volume is continuously increased. In the big data age, the deep learning method can better mine potential relations among data, and can further improve the prediction performance by combining CNN layers, channel segmentation operators and channel shuffling operation, but the influence effect of data quantity operated under different working conditions on the model is larger. In addition, based on the time sequence characteristic of the LSTM network learning NOx concentration, the state parameters of the preamble can be combined for adjustment, so that the dynamic prediction of NOx emission is realized. But because of the few parameters of the network model, it is difficult to fully exploit the potential between data. The Chinese patent publication CN109165789A discloses a modeling method of a boiler NOx emission prediction model based on LSTM, which comprises the following steps: s1: acquiring boiler operation data corresponding to the boiler operation input variable according to a preset boiler operation input variable; s2: sequencing boiler operation data according to the acquisition time of the data to obtain a boiler parameter matrix, then carrying out data recombination on the boiler parameter matrix according to the data format requirement of the LSTM model to obtain a model parameter matrix, and constructing an LSTM boiler emission initial prediction model by taking the model parameter matrix as a variable; s3: and optimizing the superparameter in the initial LSTM boiler emission prediction model, taking the optimized superparameter as a control quantity, and training the initial LSTM boiler emission prediction model to obtain a dynamic LSTM boiler NOx emission prediction model.

In summary, how to construct a model with better dynamic following performance and higher precision by using a large amount of historical data of a power plant is a problem to be solved in the existing coal-fired power plant out-of-sale control process.

Disclosure of Invention

The invention aims to provide an AE-GAN-based SCR inlet NOx emission prediction method, which predicts the NOx emission in real time and can improve the prediction accuracy of SCR inlet NOx emission.

The technical scheme provided by the invention is as follows:

an AE-GAN based SCR inlet NOx emission prediction method, the method comprising the steps of:

(1) Collecting an SCR inlet NOx value of a power plant and corresponding sensitive parameters related to the NOx value, and constructing a data set;

(2) Preprocessing a data set in a mutual information sliding time step mode;

(3) Building an AE-GAN network: taking sensitive parameters related to the NOx value as input, performing dimension reduction processing through an AE network, taking output of the AE network as input of a GAN network, and finally outputting the NOx value by the GAN network; training the AE-GAN network by using the preprocessed data set to obtain an AE-GAN network prediction model;

(4) Sensitive parameters related to the NOx value in the SIS database are collected, preprocessed in a mutual information sliding time step mode, input into an AE-GAN network prediction model, and real-time prediction is carried out on the NOx emission amount to output the NOx value.

In the step (1), the sensitive parameters related to the NOx value comprise a primary air nozzle temperature, a primary air box temperature, a flue gas oxygen content, a secondary air door opening degree, an SOFA air door, a hearth outlet flue gas temperature, a tertiary air nozzle temperature and a tertiary air perimeter air door.

In step (2), the pretreatment is: and calculating a mutual information value by sliding the sensitive parameter related to the NOx value and the NOx value in different time steps on a time axis, taking the time step with the maximum mutual information value for sliding, and then removing abnormal data by utilizing a 3 sigma criterion.

The Min-max is further used to convert the raw data to the [0,1] interval.

According to the invention, the data set is preprocessed, the sliding time step length is used for calculating the correlation coefficient between each sensitive parameter and NOx to obtain the maximum correlation, and the time delay of the NOx is reduced.

Specifically, the method for calculating the mutual information value by sliding the sensitive parameter related to the NOx value and the NOx value by different time steps on the time axis is as follows:

the entropy of the random variable X can be expressed as:

wherein: h (X) represents the information entropy of X, and P (X) is the probability distribution of X under different values;

for P (x, y) subject to a joint distribution probability, its joint entropy can be defined as:

the mutual information is the relative entropy between the joint distribution P (x, y) and P (x) P (y).

Specifically, the time delay range can be set to be 0-720 s according to engineering experience, the time interval for collecting samples is 60s, the corresponding sliding time step T epsilon [0, 12], and the sensitive parameters are respectively slid for 12 time steps to calculate the mutual information value with the NOx measuring point.

In step (3), the AE network includes an encoding layer and a decoding layer, in training the AE network: setting the number of hidden neurons of a coding layer and a decoding layer, setting a learning rate and the minimum batch number, adopting an RMSprop optimization algorithm to carry out gradient descent, and carrying out dimension descent on sensitive parameters related to NOx values.

In step (3), the error function adopted in the training process of the AE network is expressed as:

wherein: lambda is a super parameter, W is a weight matrix, b is a bias vector, x _i And y _i The i-th input and output are denoted respectively, n being the number of samples.

In step (3), the GAN network includes a generator and a arbiter, the AE network has N feature variables after dimension reduction, N consecutive time points are selected as data samples, and a matrix which can be expressed as n×n for input XCopying N copies of the predicted target data at N time-measuring points, Y being represented as a matrix of N +.>After training data pairs are prepared, setting network parameters of a generator and a discriminator, including convolution kernel size and an activation function, training a GAN network, and forming an AE-GAN network prediction model after verification of a test set in a data set.

In step (3), the loss function in the GAN network training process is:

L _gan ＝argmin _G max _D {E _x,y [logD(x,y)]+E _x [log(1-D(x,G(x)))]}

wherein: g represents a generator, D represents a discriminator, E _x () Representing the desired function of the sample x-distribution, E _x,y () A joint expectation function representing the distribution of x and y, D (x, G (x)) representing the discrimination probability of the discriminator for the generated data, D (x, y) representing the discrimination probability for the real data, argmin _G max _D Representing a process of solving the minimum and maximum values of the binary function;

to prevent model overfitting, L is added ₁ The loss function is expressed as follows:

L ₁ (G(x))＝λE _x,y (||y-G(x)|| ₁ )

wherein: I.I ₁ Is a norm calculation formula, lambda is L ₁ Weight coefficient of the loss function.

Compared with the prior art, the invention has the following beneficial effects:

(1) Compared with other prediction models, the AE-GAN network prediction model provided by the invention has higher SCR inlet NOx concentration prediction precision in a variable working condition, and has good generalization performance;

(2) According to the invention, the sliding time step is processed through mutual information, so that the time delay of NOx is effectively reduced, and the concentration of NOx at the SCR inlet can be predicted in advance.

(3) The invention uses AE network to reduce the sensitive parameter, to reduce the noise interference and improve the training speed of model; by adopting a training mode of game of a generator and a discriminator in the GAN model, potential nonlinear relations between sensitive parameters and NOx are deeply mined, and prediction accuracy is improved.

(4) The AE-GAN network prediction model built by the invention can be embedded into an online service system of a power plant, and the higher-precision prediction effect can provide an auxiliary decision for a power plant deamination system to control the denitrated ammonia injection amount, so that the method has a certain economic significance for the operation of the power plant.

Drawings

FIG. 1 is a schematic diagram of a pulverized coal boiler applicable to the embodiment;

fig. 2 is a schematic structural diagram of an AE network;

FIG. 3 is a schematic diagram of a GAN model network;

fig. 4 is a training flowchart of the AE-GAN network prediction model.

Detailed Description

The present invention will be described in further detail with reference to the drawings and examples, in order to make the objects, technical solutions and advantages of the present invention more apparent. It should be understood that the specific embodiments described herein are for purposes of illustration only and are not intended to limit the scope of the invention.

The AE-GAN-based SCR inlet NOx emission prediction method provided in this embodiment:

step one, collecting NOx emission data/NOx values of an SCR inlet of a power plant and constructing a data set corresponding to DCS parameters/sensitive parameters related to the NOx values.

The pulverized coal boiler is shown in fig. 1, pulverized coal is fed into the boiler by primary air for combustion, and after waste heat recovery equipment such as an air preheater and the like recover waste heat of flue gas generated after combustion is completed, the flue gas enters a flue gas ultra-clean system. The whole adopts a layered combustion mode. The SOFA wind is divided into an upper layer and a lower layer, and the secondary wind is divided into four layers of an upper layer, an upper middle layer, a middle lower layer and a lower layer. The horizontal section is divided into a primary air jet and a secondary air jet. The boiler adopts a bottom nitrogen burner which is distributed at four corners of the boiler.

The NOx emission sensitive parameters comprise primary air nozzle temperature, primary air box temperature, flue gas oxygen content, secondary air door opening, SOFA air door, hearth outlet flue gas temperature, tertiary air nozzle temperature and tertiary air perimeter air door.

And step two, preprocessing the data set by adopting a mutual information sliding time step.

The specific method for pretreatment comprises the following steps:

and calculating a mutual information value by sliding the sensitive parameter related to the NOx value and the NOx value in different time steps on a time axis, taking the time step with the maximum mutual information value for sliding, and then removing abnormal data by utilizing a 3 sigma criterion. And preprocessing the data set, calculating the correlation coefficient between each sensitive parameter and NOx by the sliding time step to obtain the maximum correlation, and reducing the time delay of the NOx.

The Min-max is further used to convert the raw data to the [0,1] interval.

The mutual information calculation method of the sliding time step is as follows:

the entropy of the random variable X can be expressed as:

wherein: h (X) represents the information entropy of X, and P (X) is the probability distribution of X under different values. For P (x, y) subject to a joint distribution probability, its joint entropy can be defined as:

the mutual information is the relative entropy between the joint distribution P (x, y) and the P (x) P (y), the delay range is set to be 0-720 s according to engineering experience, the time interval of sampling samples at sampling time intervals is 60s, the corresponding sliding time step T epsilon [0, 12] is adopted, and the sensitive parameters are respectively slid for 12 time steps and the NOx measuring point to calculate the mutual information value.

Step three, constructing an AE-GAN network: taking sensitive parameters related to the NOx value as input, performing dimension reduction processing through an AE network, splicing the output with a GAN network, taking the output of the AE network as the input of the GAN network, and taking the output of the GAN network as the NOx value; and training the AE-GAN network by using the preprocessed data set to obtain an AE-GAN network prediction model. The method comprises the following steps:

(1) Building and training an AE network:

as shown in FIG. 2, the AE network structure mainly comprises an encoder and a decoder, wherein the encoder mainly comprises an n-dimensional sample set(n is the number of samples and d is the dimension of the sample) to convert into a new sample set +.>The decoder will reconstruct the new sample set back to the original dimension, the decoded sample set should be +.>The training goal is to minimize the deviation of the decoded sample set from the original sample set. The error function can be expressed as:

wherein: lambda is a super parameter, W is a weight matrix, b is a bias vector, x _i And y _i The i-th input and output are denoted respectively, n being the number of samples. It can be found that there is no tag data in the above process to inputThe data itself is trained. The cross entropy is mainly adopted as a loss function, and a regularization term lambda W is added ² ₂ The reduction degree of the weight is controlled, the influence of static noise on irrelevant components in the target and the weight vector is effectively restrained, the occurrence of the overfitting phenomenon can be relieved, lambda is used for controlling the regularized intensity, and the value range is 0-1.

During the training process:

dividing the preprocessed data set into a training set and a testing set;

setting the number of hidden neurons of a coding layer and a decoding layer, giving a learning rate and the minimum batch number, training a network model, adopting an RMSprop optimization algorithm to carry out gradient descent, and carrying out dimension reduction on NOx sensitive parameters, wherein the training parameters are shown in a table 1.

Table 1 training parameter settings

Sequence number	Parameters (parameters)	Value of
			1	Loss function	BCEWithLogitsLoss
2	Optimizer	Adam
			3	Learning rate	0.0001
4	Training rounds	300
			5	Batch size	1

(2) Construction and training of the GAN network:

the GAN network is shown in fig. 3, and mainly comprises a generator and a discriminator, wherein the generator is optimized to generate a forward sample for input, and the discriminator is optimized to accurately discriminate the authenticity of the sample generated by the generator and the true sample. The optimization targets of the two networks are opposite, games are generated in the training process, and Nash equilibrium is finally achieved. The input data and the output data of the generator have a certain corresponding relation. The training samples are paired data { x, y }, the structure of the generator network is that the U-Net is a classical convolution network, the output and input data dimensions of two modules after downsampling and upsampling are consistent, and the U-Net network can well mine potential relations among the data. The optimization objective of the generator is to keep the G (x) of the output consistent with the true y-tag so that the arbiter output is 1. The arbiter outputs a predictive probability value for each region of the input image using the PatchGAN, and cuts a matrix into patches of different N sizes. The discriminator discriminates each patch, and the average result of all areas of one picture is taken as the final discriminator output. The training goal of the arbiter is to resolve the (x, G (x)) and (x, y) data pairs such that the probability value of the output is as small as possible when G (x) and x are generated as inputs to the arbiter D, and the probability of the output is as large as possible when y and x are input. The two networks continuously learn potential data relations between x and y in the process of countermeasure training, so that accurate target data can be generated, and the training loss function is as follows:

L _gan ＝argmin _G max _D {E _x,y [logD(x,y)]+E _x [log(1-D(x,G(x)))]} (1-1)

wherein: g represents a generator, D represents a discriminator, ex () represents a desired function of a sample x distribution, ex, y () represents a joint desired function of an x, y distribution, D (x, G (x)) represents a discrimination probability of the discriminator for generating data, D (x, y) represents a discrimination probability for real data, argmin _G max _D Representing the process of solving the binary function minima maxima.

To prevent model overfitting, L is added ₁ The loss function is expressed as follows:

L ₁ (G(x))＝λE _x,y (||y-G(x)|| ₁ )

wherein: I.I ₁ Is a norm calculation formula, lambda is L ₁ Weight coefficient of the loss function.

The specific training process of the GAN network is as follows:

the number of the characteristic variables after self-coding dimension reduction is N, in order to ensure the characteristic time sequence, N continuous time points are selected as data samples, and the input X can be expressed as a matrix of N multiplied by NIn order to maintain the consistency of the prediction target data matrix and the characteristic variable matrix, N copies of the prediction target data are copied at N time measuring points, wherein Y is represented as N multiplied by N matrix->After the training data pair is prepared, setting network parameters of a generator and a discriminator, including convolution kernel size and an activation function, training the model, and verifying by a test set to form a final AE-GAN prediction model.

As shown in fig. 4, in a specific training process, feature variables affecting a prediction target are subjected to feature screening, sliding step length and normalization (i.e., preprocessing in the second step), feature variables are subjected to dimension reduction processing by using self-coding (i.e., construction and training of an AE network in the third step), a data set is divided (a training set and a verification set), and construction and training of a GAN network in the third step are performed:

firstly, a generator generates a predicted NOx matrix, the training strategy of the discriminator is that the label of the predicted matrix is 0, the real label is 1, and therefore the loss function of the discriminator is calculated, and parameters are updated. The training strategy of the generator is that the label of the prediction matrix is 1, the real label is 1, the loss function is calculated, and the generator parameters are updated. After the trained model is obtained, the generator G is used for testing the data of the verification set, and a prediction result is obtained. And compared to some classical models, the predicted effects are shown in table 2. The accuracy of the GAN model is relatively high, closer to the actual value, and the SVR (linear kernel) model has the lowest accuracy, and the result shows that the GAN model has the optimal learning ability on NOx and related features.

Table 2 model effect comparison

Model	RMSE(mg/Nm 3 )	Regression coefficient R 2	MAPE(％)	Acc(％)
					SVR (Linear core)	14.35	0.864	1.68	73.94
LR	13.28	0.883	1.58	77.50
					BPNN	12.97	0.899	1.57	76.39
GAN	12.28	0.914	1.47	78.36

And step four, collecting sensitive parameters related to NOx in an SIS database, inputting the sensitive parameters into a final AE-GAN network prediction model after time sliding step length, and predicting the NOx emission in real time.

The foregoing is merely exemplary of the present invention and is not intended to limit the present invention. Various modifications and variations of the present invention will be apparent to those skilled in the art. Any modification, equivalent replacement, improvement, etc. which come within the spirit and principles of the invention are to be included in the scope of the claims of the present invention.

Claims (8)

1. An AE-GAN based SCR inlet NOx emission prediction method, the method comprising the steps of:

(1) Collecting an SCR inlet NOx value of a power plant and corresponding sensitive parameters related to the NOx value, and constructing a data set;

(2) Preprocessing a data set in a mutual information sliding time step mode;

(3) Building an AE-GAN network: taking sensitive parameters related to the NOx value as input, performing dimension reduction processing through an AE network, wherein the output of the AE network is taken as the input of a GAN network, and the output of the GAN network is the NOx value; training the AE-GAN network by using the preprocessed data set to obtain an AE-GAN network prediction model;

(4) Sensitive parameters related to the NOx value in the SIS database are collected, preprocessed in a mutual information sliding time step mode, input into an AE-GAN network prediction model, and real-time prediction is carried out on the NOx emission amount to output the NOx value.

2. The AE-GAN based SCR inlet NOx emission prediction method of claim 1, wherein in step (1), the sensitive parameters related to NOx values include primary air nozzle temperature, primary air box temperature, flue gas oxygen content, secondary air door opening, SOFA damper, furnace outlet flue gas temperature, tertiary air nozzle temperature, tertiary air perimeter damper.

3. The AE-GAN based SCR inlet NOx emission prediction method of claim 1, wherein in step (2), the preprocessing is: and calculating a mutual information value by sliding the sensitive parameter related to the NOx value and the NOx value in different time steps on a time axis, taking the time step with the maximum mutual information value for sliding, and then removing abnormal data by utilizing a 3 sigma criterion.

4. The AE-GAN based SCR inlet NOx emission prediction method of claim 3, wherein the method of calculating the mutual information value by sliding the sensitive parameter related to the NOx value and the NOx value by different time steps on the time axis is as follows:

the entropy of the random variable X can be expressed as:

wherein: h (X) represents the information entropy of X, and P (X) is the probability distribution of X under different values;

for P (x, y) subject to a joint distribution probability, its joint entropy can be defined as:

mutual information is the relative entropy between the joint distribution P (x, y) and P (x) P (y), x is the investigation variable, and y is the target variable.

5. The AE-GAN based SCR inlet NOx emission prediction method of claim 1, wherein in step (3), the AE network comprises an encoding layer and a decoding layer, in which the AE network is trained: setting the number of hidden neurons of a coding layer and a decoding layer, setting a learning rate and the minimum batch number, adopting an RMSprop optimization algorithm to carry out gradient descent, and carrying out dimension descent on sensitive parameters related to NOx values.

6. The AE-GAN based SCR inlet NOx emission prediction method of claim 5, wherein in step (3), the error function employed in the training process of the AE network is expressed as:

wherein: lambda is a super parameter, W is a weight matrix, b is a bias vector, x _i And y _i The i-th input and output are denoted respectively, n being the number of samples.

7. The AE-GAN based SCR inlet NOx emission prediction method of claim 1, wherein in step (3), the GAN network comprises a generator and a discriminator, the AE network has N feature variables after dimension reduction, N consecutive time points are selected as data samples, and a matrix representable as N X N for input XCopying N copies of the predicted target data at N time-measuring points, Y being represented as a matrix of N +.>After training data pairs are prepared, setting network parameters of a generator and a discriminator, including convolution kernel size and an activation function, training a GAN network, and forming an AE-GAN network prediction model after verification of a test set in a data set.

8. The AE-GAN based SCR inlet NOx emission prediction method of claim 7, wherein in step (3), the loss function during GAN network training is:

L _gan ＝argmin _G max _D {E _x,y [logD(x,y)]+E _x [log(1-D(x,G(x)))]}

wherein: g represents a generator, D represents a discriminator, E _x () Representing the desired function of the sample x-distribution, E _x,y () A joint expectation function representing the distribution of x and y, D (x, G (x)) representing the discrimination probability of the discriminator for the generated data, D (x, y) representing the discrimination probability for the real data, argmin _G max _D Representing a process of solving the minimum and maximum values of the binary function;

to prevent model overfitting, L is added ₁ The loss function is expressed as follows:

L ₁ (G(x))＝λE _x,y (||y-G(x)|| ₁ )

wherein: I.I ₁ Is a norm calculation formula, lambda is L ₁ Weight coefficient of the loss function.