CN113435584A - SCR (Selective catalytic reduction) outlet nitrogen oxide concentration prediction method based on LSTM (localized surface plasmon resonance) - Google Patents

Abstract

The invention discloses an LSTM-based SCR outlet nitrogen oxide concentration prediction method, which adopts an improved mutual information characteristic selection algorithm (BMIFS), considers the correlation and the redundancy to carry out auxiliary variable screening, and establishes an LSTM prediction model based on the finally determined auxiliary variables. SCR outlet NO based on BMIFS-LSTM provided by the invention_xThe concentration estimation model has better estimation precision and fitting effect, and solves the problems of larger estimation result deviation and poorer effect caused by a plurality of redundant or irrelevant variables in the input variables during modeling of the conventional LSTM algorithm.

Description

SCR (Selective catalytic reduction) outlet nitrogen oxide concentration prediction method based on LSTM (localized surface plasmon resonance)

Technical Field

The invention relates to an SCR outlet NO based on LSTM (Long-short term memory network)_xA concentration prediction method of NO at the outlet of the SCR denitration system_xThe concentration is taken as a research object to accurately reflect the NO at the outlet of the SCR_xThe real-time change of the concentration further guides the action of the reactor in time.

Background

In the electric power market of China, coal-fired thermal power generation always dominates, and NO generated by burning of coal-fired boilers of thermal power plants_xIs one of the main sources of atmospheric pollutants. Coal, which is used as a main energy source for power generation in a thermal power plant, generates a large amount of nitrogen oxides during incineration in a boiler, and is generally referred to as NO_x. In nature, during the formation and landing of rain and snow, NO in the air is absorbed_xAnd the like, and further form severe results such as building corrosion caused by acid rain, crop death and the like. At the same time, NO_xAnd the pollutants can also generate photochemical reaction with other pollutants under the action of sunlight (ultraviolet rays) to generate secondary mixed pollutants, namely photochemical smog pollution. With the increasing of domestic environmental protection consciousness day by day, the flue gas denitration optimization is promoted.

In recent years, long-short term memory neural networks (LSTM) have achieved dramatic results in processing large data. The method can not only achieve the purposes that the traditional neural network is subjected to learning training and feature extraction, then the high-level features are constructed by organizing the bottom-level features, and finally the distribution characteristics under the data are obtained, but also more importantly, the LSTM adds a state gate in the neurons thereof, can screen and process massive data, effectively solves the problems of gradient disappearance and gradient explosion, and improves the processing capacity of big data.

Aiming at denitration, most of thermal power plants adopt Selective Catalytic Reduction (SCR) technology to realize denitration so as to reduce NO_xAnd (4) discharging. For the content of nitrogen oxides in flue gas, each factory generally utilizes an automatic flue gas monitoring system to measure the concentration of the flue gas in real time, but the system has larger delay in measurement and cannot accurately reflect an SCR outletNO_xThe real-time change of the concentration can not guide the action of the reactor in time.

Disclosure of Invention

The invention aims to solve the defects in the prior art and provides an LSTM-based SCR outlet NO_xThe concentration prediction method has good prediction precision and can accurately reflect the NO at the outlet of the SCR_xReal-time changes in concentration.

In order to achieve the purpose, the technical scheme provided by the invention is as follows:

an LSTM-based SCR outlet nitrogen oxide concentration prediction method adopts an improved mutual information characteristic selection algorithm (BMIFS), auxiliary variable screening is carried out by considering correlation and redundancy, and an LSTM prediction model is established based on finally determined auxiliary variables.

The invention adopts an improved mutual information characteristic selection algorithm to ensure the relevance and redundancy of auxiliary variable screening, so as to establish an LSTM estimation model, and have better estimation precision and fitting effect.

Further, in the process of circularly selecting the auxiliary variables, the improved mutual information feature selection algorithm considers the number | S | of the selected variables, takes 1/| S | as weight, and simultaneously adds the relevance between the variables to be selected and the dominant variable into the relevance between the variables to be selected. The defects of the MIFS algorithm are overcome.

Further, the expression of the improved mutual information characteristic selection algorithm is as follows:

in the above formula, | S | represents the number of the selected characteristic variables; MR is in a selected set of variables S, f_iRelative to S_jThe formula is as follows:

if I (f)_i；c)＝0, then the characteristic variable f_iWill be rejected; if f is_iAnd S_jThere is a large correlation between all co-dominant variables, but f_iAnd S_jThere is also a high degree of redundancy between, then f_iAnd also removed. Therefore, the threshold TH is set to 0 and G in advance_MIIn comparison, if G is present_MIIf the current variable f is less than or equal to 0_iAnd the dominant variable has no great correlation, so the dominant variable is eliminated; in case of G_MIIf f is greater than or equal to 0, f is retained_iAnd entering a candidate variable set.

The auxiliary variables finally confirmed by the improved mutual information characteristic selection algorithm are as follows: boiler load, total coal quantity, total primary air quantity, total secondary air quantity, AB layer secondary air door baffle opening degree, flue gas oxygen content and flue gas temperature.

The invention uses a Keras framework aiming at the establishment of the LSTM estimation model and adopts an Adam optimization algorithm.

In some embodiments, preferably, the number of LSTM prediction model network layers is 2, and the number of neurons in hidden layers is 100. The learning rate η of the Adam optimization algorithm is 0.003, the training times are 2000 times, and the number b of batch samples is 50.

Furthermore, in order to shorten the training time of the model and optimize parameter variables, input sample data for establishing the LSTM prediction model needs to be standardized, that is, the vector space of the data is reduced according to a certain proportion and put into the space of positive-phase distribution, so that the larger difference of the data under different dimensions can be eliminated, and the convergence speed is improved. And the processed data, 80% as training data of the model and 20% as test data.

Compared with the prior art, the invention has the following advantages:

1) the invention utilizes an intelligent control algorithm to establish an SCR denitration outlet nitrogen oxide concentration model through a historical data drive-based modeling method, and accurately reflects SCR outlet NO_xThe real-time change of the concentration further realizes the optimized control of the ammonia spraying amount.

2) The invention adopts an improved mutual information characteristic selection algorithm (BMIFS) to add the relevance between the variables to be selected and the dominant variables into the relevance between the variables to be selected, thereby solving the defects of the MIFS algorithm.

3) SCR outlet NO based on BMIFS-LSTM provided by the invention_xThe concentration estimation model has better estimation precision and fitting effect, and solves the problems of larger estimation result deviation and poorer effect caused by a plurality of redundant or irrelevant variables in the input variables during modeling of the conventional LSTM algorithm.

Drawings

FIG. 1 is a schematic diagram of an input/output structure of an LSTM network;

FIG. 2 is an algorithmic flow chart of a model;

FIG. 3 is a comparison graph of the fitting effect of training data;

FIG. 4 is a graph of a comparison of estimated error of training data;

FIG. 5 is a comparison graph of predicted effect of test data;

FIG. 6 is a comparison graph of the estimated error of test data.

Detailed Description

The following detailed description is made with reference to the principles and specific embodiments of the invention.

First, principle introduction

1.1 SCR deNOx systems

The SCR flue gas denitration system mainly comprises NH₃Storage and preparation system, NH₃Flue gas mixing system, NH₃The system comprises an injection system, an SCR reactor, a reactor bypass, an economizer bypass, a detection control system and other auxiliary systems. For the arrangement of the SCR reactor, a high ash arrangement is generally used, i.e. in the flue between the economizer and the air preheater.

The SCR denitration system mainly comprises the following structures:

(1)NH₃the high-purity liquid ammonia of the storage and preparation system is stored in a liquid ammonia tank, and the liquid ammonia is converted into ammonia gas after being heated by liquefied gas in an electric evaporator. Ammonia and dilution air in NH₃Mixing in an air mixer.

(2)NH₃The flue gas mixing system performs NH by a flue structural member, such as a guide plate, a static mixer and the like₃And (4) mixing the flue gas.

(3) Ammonia injector for ammonia injection system is required to be installed upstream of SCR reactor and at a distance to ensure injected NH₃And is fully mixed with the flue gas through the flue. The diluted ammonia gas is delivered to the ammonia spraying busbar through the ammonia supply pipeline and is uniformly sprayed through the ammonia spraying grid.

(4) The SCR reactor completes the SCR reaction process and typically employs a 2+1 or 3+1 arrangement.

(5) The economizer bypass controls the reaction temperature by adjusting the ratio of the economizer flue gas to the bypass flue gas.

(6) The SCR reactor bypass is used for isolating the SCR reactor during the starting and stopping of the boiler or under the emergency condition, so that the catalyst is not damaged, and the power consumption of the induced draft fan is saved.

(7) The main controlled objects of the detection control system comprise: ammonia spraying amount, denitration ventilation, water temperature of an ammonia evaporator, pressure of a buffer tank, flue gas temperature at an inlet of the SCR reactor, a soot blowing system and a flue gas baffle system. The ammonia injection amount is controlled by controlling NO at the inlet and outlet of the SCR system_xConcentration and O₂Content, dilution fan flow, flue gas temperature and flue gas flow. The detecting part mainly uses NH₃The monitoring analyzer monitors the ammonia slip.

(8) The auxiliary system mainly comprises a grid spray nozzle for purging the ammonia spray, a compressed air pipeline for preventing the spray nozzle from being blocked and a soot blower for purging the SCR reactor.

1.2 LSTM estimation model structure

In order to shorten the training time of the model and optimize the parameter variables, the input samples need to be normalized, that is, the vector space of the data is reduced according to a certain proportion and put into the positive space. Therefore, the larger difference of data under different dimensions can be eliminated, the convergence rate is improved, and the formula is as follows:

dividing data into training set and testing set and standardizing, and then establishing basis according to the following stepsSCR outlet NO at LSTM_xAnd (6) estimating the model.

1) Input layer

Training sample data x epsilon R^b×tWhere b is the number of samples used for each model training, and t is the sample data dimension. Since the LSTM input layer requires that the sample data must be three-dimensional, the three dimensions are:

sample: a sequence is a sample and may contain a plurality of samples.

Time step II: one time step represents one observation point in the sample.

Characteristics: one feature is obtained in one time step.

The expression of the converted three-dimensional matrix is x ∈ R^b×s×iS denotes the time dimension of the sample and i denotes the feature. In Keras, we can use the reshape () function in the Numpy array for three-dimensional reconstruction. By letting x be R^b×s×iBy mapping at the input layer, we can obtain the input after changing the sample dimension, as shown in equation 2:

y⁽ⁱ⁾＝x·W⁽ⁱ⁾+b⁽ⁱ⁾ (2)

in the above formula, W⁽ⁱ⁾∈R^i×i1，b⁽ⁱ⁾∈Rⁱ¹，y⁽ⁱ⁾∈R^b×s×i1。

2) LSTM network layer

LSTM input is y⁽ⁱ⁾∈R^b×s×i1Assuming that the network has n neurons, the hidden layer output at the last moment of each sample is taken as the output y of the LSTM^(h)Then y is^(h)∈R^b×d. The input/output flow structure is shown in fig. 1.

3) Output layer

The network adopts a softmax layer for output, and the output formula is as follows:

y′＝softmax(y^h·W^(o)) (3)

in the above formula, W^(o)∈R^d×nN is the number of classifications, y 'is the network output, y' is e.g. R^b×n。

4) Loss function

The difference between the output of the training model and the actual data output, called the loss, can be obtained by comparing the two. The smaller the loss value is, the better the training effect of the model is, and if the predicted value is consistent with the actual value, no loss is caused. The function used to calculate the magnitude of the Loss is called the Loss function, which gives an objective measure of the predicted effect. The formula is shown as (4):

H(y)＝-∑_by′log(y) (4)

through repeated tests, the LSTM network model is finally determined to have 2 LSTM layers, each layer has 100 nodes, Adam is selected as an optimization algorithm, the batch size is set to be 20, and the epoch is set to be 2000. The algorithm flow of the model is shown in fig. 2.

1.3 improved mutual information characteristic selection Algorithm (BMIFS)

For the reason that the MIFS cannot guarantee the relevance and the redundancy during screening of the auxiliary variables, an improved mutual information feature selection algorithm, namely a BMIFS algorithm, is adopted in the invention. The improvement of the algorithm is that in the process of circularly selecting the auxiliary variables, the influence of the number | S | of the selected variables is taken into consideration, 1/| S | is taken as weight, and meanwhile, the relevance between the variables to be selected and the dominant variables is added into the relevance between the variables to be selected, so that the defects of the MIFS algorithm are overcome. The expression of the algorithm is as follows (5):

in the above formula, | S | represents the number of selected characteristic variables, MR is in the selected variable set S, f_iRelative to S_jThe formula (6):

if I (f)_i(ii) a c) If 0, the characteristic variable f_iWill be rejected; if f is_iAnd S_jThere is a large correlation between all co-dominant variables, but f_iAnd S_jThere is also a high degree of redundancy between, then f_iAnd also removed. Therefore, here, the threshold TH is set to 0 and G in advance_MIIn comparison, if G is present_MIIf the current variable f is less than or equal to 0_iAnd the dominant variable is not related to the other variables, so that the dominant variable is eliminated. In case of G_MIIf f is greater than or equal to 0, f is retained_iAnd entering a candidate variable set.

Example 1

Auxiliary variable screening is carried out by adopting the BMIFS algorithm, and SCR outlet NO is established based on LSTM on the basis_xAnd (5) a concentration estimation model. The results are as follows:

1、NO_xsecondary variable screening results

The research object of the invention is the outlet NO of the SCR denitration system_xAnd (3) through the analysis of the NOx generation mechanism, after a certain amount of reliable site historical operation data of a certain 300MW thermal power generating unit is collected, preprocessing is carried out, and auxiliary variable dimensionality reduction is selected by combining with improved BMIFS, wherein beta is set to be 0.7.

The raw input variables include: boiler load, total coal amount, total primary air amount, total secondary air amount, SOFA3, SOFA2, SOFA1, OFA2, OFA1, EF, E, DE, D, CD2, CD1, C, BC, B, AB, A and AA, and 17 layers of air door baffle opening degrees, flue gas temperature, flue gas oxygen content and denitration reactor outlet NOx concentration. Since the OF1 and CD2 shutter openings are always zero, they are rejected in the preprocessing.

After calculation through the BMIFS algorithm, the variable with the maximum relevance with the leading variable, namely the maximum mutual information value, can be obtained as the total secondary air volume, and based on the total secondary air volume, the residual quantity is further obtained to enable the evaluation function G to be evaluated_M6 auxiliary variables > 0, as shown in Table 1.

TABLE 1 auxiliary variable evaluation function values

After screening, the model finally determines 7 auxiliary variables, which are respectively: load, total coal quantity, total primary air quantity, total secondary air quantity, AB layer secondary air door baffle opening degree, flue gas oxygen content and flue gas temperature.

2. Modeling results

When the Keras framework is used for LSTM network building, necessary parameters need to be set, the parameters are set differently, and finally the characteristics of the model are different. Parameter optimization is the process of choosing an optimal set of parameters for a learning algorithm. Under the Keras framework, parameters to be adjusted mainly include the number of neural network layers, the number of hidden layer neurons, the total training times, the size of batch samples, the learning rate and the like. The invention mainly aims at an LSTM estimation model and mainly optimizes two parameters of total training times and learning rate.

Experiments show that the accuracy rate gradually converges when the training times reach about 1000 times, and the accuracy rate obviously decreases after the training times reach about 9000 times, and the main reason is gradient explosion caused by excessive training. Therefore, the training times are set to 2000 times, so that the network model training can be realized in less time, and the gradient explosion is avoided. The learning rate is 0.001,0.003 and 0.006, the model training times are within 5000 times, and the accuracy rate changes basically in a consistent manner, but when the training times exceed 5000 times, the curve with the learning rate of 0.001 obviously slips down, and when the training times reach 6800 times, the curve with the learning rate of 0.006 also obviously slips down. The learning rate determines the speed of updating the parameters to the optimal value to a certain extent, and experimental result analysis can show that when the learning rate is too large, the gradient descending step length of each training of the model is too large, so that the optimal solution is easy to miss.

After adjusting parameters through multiple experiments, the learning rate of the Adam algorithm is finally determined to be eta equal to 0.003, the number of batch samples is determined to be b equal to 50, the number of LSTM network layers is 2, the number of neurons in hidden layers is 100, and the training frequency is determined to be T equal to 2000. And respectively establishing an LSTM network model without auxiliary variable extraction and an LSTM network model subjected to auxiliary variable screening by using the preprocessed 1600 groups of data, wherein 80% of the 1600 groups of data are used as training data of the model and 20% of the data are used as test data of the model, wherein the fitting effect of the final model training data is shown in figure 3, and the relative error is shown in figure 4.

According to fig. 3, it can be seen that both the two types of model LSTM and the BMIFS-LSTM of the present invention can better fit training data, but as can be seen from the relative error of fig. 4, the model established based on the BMIFS-LSTM of the present invention has a smaller error than the model of LSTM, and therefore, it can be obtained that a sample selected in advance through the BMIFS auxiliary variables can better train the model, thereby achieving a better fitting effect.

And then, estimating the test data by using the trained model to obtain final estimation results and errors of the two models, as shown in fig. 5 and 6.

According to fig. 5, it can be analyzed and found that the prediction accuracy of the prediction model (LSTM) without auxiliary variable extraction is not as high as that of the prediction model (BMIFS-LSTM) with auxiliary variable extraction, and in combination with fig. 6, it can be further found that the relative error of the prediction model based on BMIFS-LSTM is smaller than that of the prediction model based on BMIFS-LSTM, which also further shows that the prediction accuracy of the model of the present invention is higher.

Claims (10)

1. The SCR outlet nitrogen oxide concentration prediction method is characterized by adopting an improved mutual information characteristic selection algorithm, carrying out auxiliary variable screening by considering correlation and redundancy, and establishing an LSTM estimation model based on the finally determined auxiliary variables.

2. The method for predicting the concentration of nitrogen oxides at an SCR outlet according to claim 1, wherein the improved mutual information characteristic selection algorithm considers the number | S | of the selected variables in the process of circularly selecting the auxiliary variables, takes 1/| S | as weight, and adds the correlation between the selected variables and the dominant variables into the correlation between the selected variables.

3. The SCR outlet nitrogen oxide concentration prediction method of claim 2, wherein the expression of the improved mutual information feature selection algorithm is as follows:

in the above formula, | S | represents the number of the selected characteristic variables; MR is in a selected set of variables S, f_iRelative to S_jThe formula is as follows:

if I (f)_i(ii) a c) If 0, the characteristic variable f_iWill be rejected; if f is_iAnd S_jThere is a large correlation between all co-dominant variables, but f_iAnd S_jThere is also a high degree of redundancy between, then f_iAnd also removed.

4. The SCR outlet NOx concentration prediction method of claim 3, wherein the improved mutual information feature selection algorithm presets thresholds TH-0 and G during the loop selection of the auxiliary variable_MIIn comparison, if G is present_MIIf the current variable f is less than or equal to 0_iAnd the dominant variable has no great correlation, so the dominant variable is eliminated; in case of G_MIIf f is greater than or equal to 0, f is retained_iAnd entering a candidate variable set.

5. The SCR outlet NOx concentration prediction method of claim 4, wherein the improved mutual information feature selection algorithm ultimately validates the auxiliary variables as: boiler load, total coal quantity, total primary air quantity, total secondary air quantity, AB layer secondary air door baffle opening degree, flue gas oxygen content and flue gas temperature.

6. The SCR outlet nitrogen oxide concentration prediction method of any one of claims 1 to 5, wherein the LSTM prediction model is established by using a Keras framework and adopting an Adam optimization algorithm.

7. The method for predicting the concentration of nitrogen oxides at the outlet of SCR of claim 6, wherein the number of LSTM predictive model network layers is 2, and the number of neurons in hidden layers is 100.

8. The method of predicting SCR outlet nox concentration of claim 7, wherein the learning rate η of the Adam optimization algorithm is 0.003 and the number of training is 2000.

9. The method of predicting SCR outlet nox concentration according to claim 8, wherein the number of batch samples of the LSTM prediction model, b, is 50.

10. The method of predicting SCR outlet nox concentration according to claim 9, wherein the LSTM prediction model normalizes input sample data and compares the normalized input sample data with a reference value of 8: 2 into training and test sets.

Images