CN114177747A - Flue gas desulfurization sulfur dioxide concentration prediction method based on machine learning algorithm - Google Patents

Flue gas desulfurization sulfur dioxide concentration prediction method based on machine learning algorithm Download PDF

Info

Publication number
CN114177747A
CN114177747A CN202111473270.0A CN202111473270A CN114177747A CN 114177747 A CN114177747 A CN 114177747A CN 202111473270 A CN202111473270 A CN 202111473270A CN 114177747 A CN114177747 A CN 114177747A
Authority
CN
China
Prior art keywords
flue gas
concentration
data
gas desulfurization
machine learning
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN202111473270.0A
Other languages
Chinese (zh)
Inventor
谭琨
贾义
李承泉
张方醒
徐鑫荣
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Kunyue Internet Environmental Technology Jiangsu Co ltd
Original Assignee
Kunyue Internet Environmental Technology Jiangsu Co ltd
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Kunyue Internet Environmental Technology Jiangsu Co ltd filed Critical Kunyue Internet Environmental Technology Jiangsu Co ltd
Priority to CN202111473270.0A priority Critical patent/CN114177747A/en
Publication of CN114177747A publication Critical patent/CN114177747A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • B—PERFORMING OPERATIONS; TRANSPORTING
    • B01—PHYSICAL OR CHEMICAL PROCESSES OR APPARATUS IN GENERAL
    • B01D—SEPARATION
    • B01D53/00—Separation of gases or vapours; Recovering vapours of volatile solvents from gases; Chemical or biological purification of waste gases, e.g. engine exhaust gases, smoke, fumes, flue gases, aerosols
    • B01D53/34—Chemical or biological purification of waste gases
    • B01D53/74—General processes for purification of waste gases; Apparatus or devices specially adapted therefor
    • B01D53/77—Liquid phase processes
    • B01D53/78—Liquid phase processes with gas-liquid contact
    • B—PERFORMING OPERATIONS; TRANSPORTING
    • B01—PHYSICAL OR CHEMICAL PROCESSES OR APPARATUS IN GENERAL
    • B01D—SEPARATION
    • B01D53/00—Separation of gases or vapours; Recovering vapours of volatile solvents from gases; Chemical or biological purification of waste gases, e.g. engine exhaust gases, smoke, fumes, flue gases, aerosols
    • B01D53/34—Chemical or biological purification of waste gases
    • B01D53/346—Controlling the process
    • B—PERFORMING OPERATIONS; TRANSPORTING
    • B01—PHYSICAL OR CHEMICAL PROCESSES OR APPARATUS IN GENERAL
    • B01D—SEPARATION
    • B01D53/00—Separation of gases or vapours; Recovering vapours of volatile solvents from gases; Chemical or biological purification of waste gases, e.g. engine exhaust gases, smoke, fumes, flue gases, aerosols
    • B01D53/34—Chemical or biological purification of waste gases
    • B01D53/46—Removing components of defined structure
    • B01D53/48—Sulfur compounds
    • B01D53/50—Sulfur oxides
    • B01D53/501—Sulfur oxides by treating the gases with a solution or a suspension of an alkali or earth-alkali or ammonium compound
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • B—PERFORMING OPERATIONS; TRANSPORTING
    • B01—PHYSICAL OR CHEMICAL PROCESSES OR APPARATUS IN GENERAL
    • B01D—SEPARATION
    • B01D2251/00—Reactants
    • B01D2251/40—Alkaline earth metal or magnesium compounds
    • B01D2251/404—Alkaline earth metal or magnesium compounds of calcium
    • B—PERFORMING OPERATIONS; TRANSPORTING
    • B01—PHYSICAL OR CHEMICAL PROCESSES OR APPARATUS IN GENERAL
    • B01D—SEPARATION
    • B01D2258/00—Sources of waste gases
    • B01D2258/02—Other waste gases
    • B01D2258/0283—Flue gases

Abstract

The invention provides a method for predicting the concentration of sulfur dioxide in flue gas desulfurization based on a machine learning algorithm. Because the outlet SO of the flue gas desulfurization device2Concentration and inlet flue gas volume, SO2The concentration, the temperature, the humidity, the oxygen content, the spraying amount of the absorption liquid, the PH value, the density, the oxidation rate and other factors are related, and the invention applies a machine learning algorithm model to the SO at the outlet of the wet flue gas desulfurization device2And predicting and early warning the concentration. The invention effectively solves the problem that the traditional hysteresis effect delays the increase of the desulfurizer, and the SO at the outlet of the device2Excessive concentrationProblems that cannot be completely avoided; and the excessive desulfurizer entering the by-products is reduced, and the operation cost is reduced. Realize wet flue gas desulfurization device export SO2The concentration prediction early warning, the algorithm model participates in the operation control, and the SO at the outlet of the sulfur device can be effectively reduced by increasing the spraying amount of the absorption liquid and the adding amount of the desulfurizer in time2And the concentration exceeds the standard risk, and the stable standard emission of the wet flue gas desulfurization device is realized.

Description

Flue gas desulfurization sulfur dioxide concentration prediction method based on machine learning algorithm
Technical Field
The invention belongs to the field of industrial flue gas treatment, and particularly relates to a flue gas desulfurization sulfur dioxide concentration prediction method based on a machine learning algorithm.
Background
China is a large coal country, coal is still the main fuel for industrial use at present, and when the coal releases heat in the combustion process, a large amount of pollutants such as particulate matters, sulfur dioxide, greenhouse gases and the like can be generated, so that the pollution to the ecological environment is caused. The wet flue gas desulfurization technology is one of the worldwide commercialized desulfurization methods, can efficiently remove sulfur oxides in flue gas, is easy to recycle by-products, and is the most effective flue gas desulfurization technology for controlling the pollution of atmospheric sulfur dioxide. The flue gas containing sulfur dioxide enters a desulfurization device, and is contacted with a spray liquid containing an absorbent in a desulfurization tower, and the sulfur dioxide in the flue gas is absorbed, removed, washed and purified, and then the flue gas is discharged through a chimney.
The desulfurization efficiency of the desulfurization unit is related to the liquid-gas ratio, the chemical dosage of the desulfurizing agent (such as calcium-sulfur ratio) and the operation parameters (such as flue gas temperature, pressure, flow, humidity, oxygen content, sulfur dioxide concentration and other flue gas components, the temperature, pressure, flow, density, pH value, oxidation rate, components and the like of the spray liquid, the flow, concentration, density, components and the like of the absorbent). The higher the efficiency of the desulfurization unit, the lower the concentration of sulfur dioxide in the outlet flue gas. For the same inlet flue gas condition, the higher desulfurization efficiency can be obtained by increasing the spraying amount of the absorption liquid and increasing the flow of the desulfurizer to improve the pH value of the absorption liquid, and the concentration of sulfur dioxide in the flue gas of the device is reduced. In actual production, the spraying amount of the absorption liquid is provided by a circulating pump, and the flow rate of the circulating pump is not regulated generally. The amount of the desulfurizing agent to be added is usually adjusted by the concentration of sulfur dioxide at the outlet of the desulfurizing device. The control has certain hysteresis, and when the production load fluctuates greatly, the flow rate of flue gas at the inlet of the desulfurization device and the concentration of sulfur dioxide change greatly, so that the concentration of sulfur dioxide at the outlet of the desulfurization device fluctuates and even exceeds the standard. The control method commonly used at present is to increase the amount of the desulfurizing agent to improve the desulfurization efficiency of the apparatus. This measure has two disadvantages: firstly, the effect of increasing the desulfurizer is delayed by a hysteresis effect, and the excessive concentration of sulfur dioxide at the outlet of the device cannot be completely avoided; and excessive desulfurizing agent enters the by-product, so that the operation cost is increased.
Disclosure of Invention
The invention aims to solve the defects in the prior art and provides a method for predicting the concentration of sulfur dioxide in flue gas desulfurization based on a machine learning algorithm.
In order to achieve the purpose, the invention adopts the following technical scheme: a method for predicting the concentration of sulfur dioxide in flue gas desulfurization based on a machine learning algorithm comprises the following steps:
s1: collecting historical operating data of the wet flue gas desulfurization device as sample set data;
s2: arranging the sample set data acquired in the S1 into a type and a format required by machine learning, and cleaning the data to form sample data of the machine learning;
s2-1: deleting missing values in the sample set, and sequencing from small to large to obtain sequences { X1, X2, X3, … … Xn } of each parameter;
s2-2: let QL＝X(n/4),QU＝X(3n/4),IQR＝QU-QL(ii) a n/4 and 3n/4 are rounded off to obtain integers;
QLis the lower quartile; qUIs the upper quartile; IQR is a four-bit distance;
QL＝X(n/4),QU＝X(3n/4)all values are corresponding to the number series { X1, X2, X3, … … Xn };
when n is 10, QL＝X(10/4)＝X(2.5)，QLTaking X2A value of (d);
QU＝X(3*10/4)＝X(7.5)，QUtaking X7A value of (d);
then IQR is equal to QU-QL＝X7-X2；
S2-3: removal of greater than Q in the S2-1 sequenceU+1.5IQR and less than QL-a value of 1.5IQR, the rusty data constituting a new data sample for machine learning;
s3: according to the data of the flue gas of the wet desulphurization device, the data of the absorption spray liquid and the outlet SO2Establishing a machine learning algorithm model according to the relevance of the concentration;
s3-1: taking 20 time sequences as a group of input sequences, wherein each time node hi is obtained by calculation through a tree network structure, and the algorithm formula is as follows:
fi＝σ(wixt+bi)
ft＝σ(wfhi-1+bf)
pi＝tanh(wp[hi-1；xt]+bp)
ot＝tanh(wo[hi-1；xt]+bo)
hi＝tanh(ft*ot+pi*fi)
wherein f isiThe method is a mapping relation and is used for recording the information at the current moment;
σ is a sigmoid function, which is formulated as follows:
Figure BDA0003389359200000031
xtcharacteristic information input at the moment;
ftis a mapping relationship;
hi-1is the result of the information matrix operation at the previous moment;
[hi-1；xt]represents the calculation result of the previous time andsplicing the feature information at the present moment;
the tanh function is as follows:
Figure BDA0003389359200000041
represents a matrix dot product operation;
wt, Wf, Wo, Wp, bt, bf, bo, bp are all parameter matrixes;
s3-2: an Attention mechanism is introduced, when a time sequence is input, more Attention is paid to information of different time points in the previous time period, and information with longer recording time is recorded; wherein the Attention is a selection mechanism, and selects the most relevant position information of the current node;
the score value of each node is the product of the hidden state hi (h1, h2 … hk) (k is the number of information input in each iteration) of the input sequence and the last node hi-1, and is subjected to softmax weighted calculation: the formula is as follows:
Figure BDA0003389359200000042
the formula is a normalized exponential function and means gradient logarithm normalization of finite term discrete probability distribution;
Figure BDA0003389359200000043
the value of k is from 1 to T, and T is the length of the input time sequence;
Figure BDA0003389359200000044
represents a summation;
exp(et) Represents the power of an exponential function e;
wherein ek＝hi
S3-3: the correlation is maximal when the mean of all input features is extracted for the first 5-30 minutes (preferably 10-15 minutes), which is expressed as follows:
a(hj)＝tanh(WX+B)
w and B are parameter matrixes, the running state of the desulfurizing tower in the first ten minutes is extracted,
where hj is the average input of the first 5-30 minutes (preferably 10-15 minutes) feature;
x is the average value of the characteristic parameters of the first 10 minutes;
s3-4: by the formula:
Figure BDA0003389359200000051
calculating to obtain a final result, namely outlet sulfur dioxide concentration;
wherein m is a selection information parameter matrix;
s4: training the algorithm model established in the step 3 by using the data of S2; the accuracy and precision of the result are improved by testing and adjusting the parameters of the algorithm model;
s5: inputting real-time operation data, and predicting the SO at the outlet of the wet flue gas desulfurization device by using the algorithm model obtained in S42Concentration; the obtained prediction result is the outlet SO of the wet flue gas desulfurization device2Concentration;
s6: judging the prediction result of S5, if the prediction result exceeds the set value, outputting an alarm event, and storing the algorithm model to enter the next prediction; if the set value is not exceeded, comparing the measured value with the actual value;
the set value is SO at the outlet of the desulfurizer2Concentration control value, currently implemented ultra-low emission standard is 35mg/m3；
S7: when the measured value is consistent with the predicted value, the algorithm model is stored to enter the next prediction; if the measured value is not consistent with the predicted value, the group of running data enters a historical database, and the algorithm model is carried out again; the algorithm model after training, testing and iteration is used for the outlet SO of the S5 wet flue gas desulfurization device2Predicting the concentration;
training and testing refer to repeating the working contents of the steps S3 and S4;
iteration is a process of forming a historical database by using an S7 rule and repeating the steps S2-S4 to obtain a new algorithm model to replace the original algorithm model.
Preferably, the historical operating data of the wet flue gas desulfurization device collected in S1 includes the temperature, pressure, flow rate, humidity, oxygen content, SO2 concentration and other flue gas components of the flue gas entering and exiting the desulfurization device; the temperature, pressure, flow, density, PH value, oxidation rate and components of the spray liquid for washing and purifying the flue gas; flow rate, concentration, density, composition of the absorbent required for flue gas desulfurization.
Preferably, the method for data cleansing in S2 includes two methods, a data-based method and a rule-based method; the data-based mode comprises a classification method, a neighbor method, a clustering method and a statistical method; the rule-based approaches include data being unchanged for long periods of time, data being too variable, measurement values being over-range, and data being over-threshold.
Preferably, the method for establishing the machine learning algorithm model in S3 adopts a recurrent neural network RNN machine learning algorithm model.
Preferably, the S4 further includes the following steps:
s4-1: dividing a desulphurization device operation parameter matrix into a training set matrix and a test set matrix according to time;
s4-2: inputting the selected parameter matrix of the desulfurization device of the training set into the training input of a machine learning algorithm model, and constructing an outlet SO of the wet flue gas desulfurization device2An algorithm model with the concentration as a prediction target, and an outlet SO corresponding to the operation parameter matrix of the desulfurization device of the training set2And (4) concentration.
Preferably, the alarm threshold and the alarm form in S6 may be set and modified according to actual use requirements.
Compared with the prior art, the invention has the beneficial effects that:
(1) analyzing historical operation data of the wet desulphurization device by machine learning and adopting a Recurrent Neural Network (RNN) machine learning algorithm model, and excavating each operation characteristic parameter and an outlet SO of the desulphurization device2The internal logic of concentration makes the principle model of desulfurization reaction explicit for the outlet SO of wet desulfurization device2Of concentrationAnd the operation state of the desulfurization device can be conveniently mastered by operators in real time through prediction.
(2) Realize wet flue gas desulfurization device export SO2The concentration prediction early warning, the algorithm model participates in the operation control, and the SO at the outlet of the sulfur device can be effectively reduced by increasing the spraying amount of the absorption liquid and the adding amount of the desulfurizer in time2And the concentration exceeds the standard risk, and the stable standard emission of the wet flue gas desulfurization device is realized.
(3) Realize wet flue gas desulfurization device export SO2And (3) predicting and early warning the concentration, wherein the algorithm model participates in operation control, and the spraying amount of the absorption liquid and the adding amount of the desulfurizer are timely reduced. The power consumption of the circulating pump can be reduced by reducing the spraying amount of the absorption liquid. The risk of excessive feeding of the desulfurizer can be reduced by reducing the feeding amount of the desulfurizer, the utilization efficiency of the desulfurizer is improved, and the consumption of the desulfurizer is reduced. Therefore, the operation cost of the desulfurization device can be reduced by reducing the spraying amount of the absorption liquid and the adding amount of the desulfurizer.
Drawings
FIG. 1 is a schematic flow diagram of the present invention;
FIG. 2 is a diagram of an RNN network architecture;
FIG. 3 is an example of a characteristic parameter operating data history;
fig. 4 is an example of a time series analysis of the operation history data of the wet desulfurization apparatus.
Detailed Description
In order to further understand the objects, structures, features and functions of the present invention, the following embodiments are described in detail.
Referring to fig. 1-4, the invention provides a method for predicting the concentration of sulfur dioxide in flue gas desulfurization based on a machine learning algorithm.
A method for predicting the concentration of sulfur dioxide in flue gas desulfurization based on a machine learning algorithm comprises the following steps:
s1: collecting historical operating data of the wet flue gas desulfurization device as sample set data;
s2: arranging the sample set data acquired in the S1 into a type and a format required by machine learning, and cleaning the data to form sample data of the machine learning;
s2-1: deleting missing values in the sample set, and sequencing from small to large to obtain sequences { X1, X2, X3, … … Xn } of each parameter;
s2-2: let QL＝X(n/4),QU＝X(3n/4),IQR＝QU-QL(ii) a n/4 and 3n/4 are rounded off to obtain integers;
QLis the lower quartile; qUIs the upper quartile; IQR is a four-bit distance;
QL＝X(n/4),QU＝X(3n/4)is a series of numbers { X1,X2,X3,……XnThe corresponding value;
when n is 10, QL＝X(10/4)＝X(2.5)，QLTaking X2A value of (d);
QU＝X(3*10/4)＝X(7.5)，QUtaking X7A value of (d);
then IQR is equal to QU-QL＝X7-X2；
S2-3: removal of greater than Q in the S2-1 sequenceU+1.5IQR and less than QL-a value of 1.5IQR, the rusty data constituting a new data sample for machine learning;
s3: according to the data of the flue gas of the wet desulphurization device, the data of the absorption spray liquid and the outlet SO2Establishing a machine learning algorithm model according to the relevance of the concentration;
s3-1: taking 20 time sequences as a group of input sequences, wherein each time node hi is obtained by calculation through a tree network structure, and the algorithm formula is as follows:
fi＝σ(wixt+bi)
ft＝σ(wfhi-1+bf)
pi＝tanh(wp[hi-1；xt]+bp)
ot＝tanh(wo[hi-1；xt]+bo)
hi＝tanh(ft*ot+pi*fi)
wherein f isiThe method is a mapping relation and is used for recording the information at the current moment;
σ is a sigmoid function, which is formulated as follows:
Figure BDA0003389359200000091
xtcharacteristic information input at the moment;
ftis a mapping relationship;
hi-1is the result of the information matrix operation at the previous moment;
[hi-1；xt]splicing the calculation result representing the previous moment and the characteristic information of the current moment;
the tanh function is as follows:
Figure BDA0003389359200000101
represents a matrix dot product operation;
wt, Wf, Wo, Wp, bt, bf, bo, bp are all parameter matrixes;
s3-2: an Attention mechanism is introduced, when a time sequence is input, more Attention is paid to information of different time points in the previous time period, and information with longer recording time is recorded; wherein the Attention is a selection mechanism, and selects the most relevant position information of the current node;
the score value of each node is the product of the hidden state hi (h1, h2 … hk) (k is the number of information input in each iteration) of the input sequence and the last node hi-1, and is subjected to softmax weighted calculation: the formula is as follows:
Figure BDA0003389359200000102
the formula is a normalized exponential function and means gradient logarithm normalization of finite term discrete probability distribution;
Figure BDA0003389359200000103
the value of k is from 1 to T, and T is the length of the input time sequence;
Figure BDA0003389359200000104
represents a summation;
exp(et) Represents the power of an exponential function e;
wherein ek＝hi
S3-3: the correlation is maximal when the mean of all input features is extracted for the first 5-30 minutes (preferably 10-15 minutes), which is expressed as follows:
a(hj)＝tanh(WX+B)
w and B are parameter matrixes, the running state of the desulfurizing tower in the first ten minutes is extracted,
where hj is the average input of the first 5-30 minutes (preferably 10-15 minutes) feature;
x is the average value of the characteristic parameters of the first 10 minutes;
s3-4: by the formula:
Figure BDA0003389359200000111
calculating to obtain a final result, namely outlet sulfur dioxide concentration;
wherein m is a selection information parameter matrix;
s4: training the algorithm model established in the step 3 by using the data of S2; the accuracy and precision of the result are improved by testing and adjusting the parameters of the algorithm model;
s5: inputting real-time operation data, and predicting the SO at the outlet of the wet flue gas desulfurization device by using the algorithm model obtained in S42Concentration; the obtained prediction result is the outlet SO of the wet flue gas desulfurization device2Concentration;
s6: judging the prediction result of S5, if the prediction result exceeds the set value, outputting an alarm event, and storing the algorithm model to enter the next prediction; if the set value is not exceeded, comparing the measured value with the actual value;
the set value is SO at the outlet of the desulfurizer2Concentration control value, currently implemented ultra-low emission standard is 35mg/m3；
S7: when the measured value is consistent with the predicted value, the algorithm model is stored to enter the next prediction; if the measured value is not consistent with the predicted value, the group of running data enters a historical database, and the algorithm model is carried out again; the algorithm model after training, testing and iteration is used for the outlet SO of the S5 wet flue gas desulfurization device2Predicting the concentration;
training and testing refer to repeating the working contents of the steps S3 and S4;
iteration is a process of forming a historical database by using an S7 rule and repeating the steps S2-S4 to obtain a new algorithm model to replace the original algorithm model.
Preferably, the historical operating data of the wet flue gas desulfurization device collected in S1 includes the temperature, pressure, flow rate, humidity, oxygen content, SO2 concentration and other flue gas components of the flue gas entering and exiting the desulfurization device; the temperature, pressure, flow, density, PH value, oxidation rate and components of the spray liquid for washing and purifying the flue gas; flow rate, concentration, density, composition of the absorbent required for flue gas desulfurization.
Preferably, the method for data cleansing in S2 includes two methods, a data-based method and a rule-based method; the data-based mode comprises a classification method, a neighbor method, a clustering method and a statistical method; the rule-based approaches include data being unchanged for long periods of time, data being too variable, measurement values being over-range, and data being over-threshold.
Preferably, the method for establishing the machine learning algorithm model in S3 adopts a recurrent neural network RNN machine learning algorithm model.
Analyzing historical operation data of the wet desulphurization device by machine learning and adopting a Recurrent Neural Network (RNN) machine learning algorithm model, and excavating each operation characteristic parameter and an outlet SO of the desulphurization device2Inherent logic of concentration, making the principle model of desulfurization reaction explicitFor wet desulfurization device outlet SO2And the concentration is predicted, so that operators can conveniently master the operation state of the desulfurization device in real time.
Realize wet flue gas desulfurization device export SO2The concentration prediction early warning, the algorithm model participates in the operation control, and the SO at the outlet of the sulfur device can be effectively reduced by increasing the spraying amount of the absorption liquid and the adding amount of the desulfurizer in time2And the concentration exceeds the standard risk, and the stable standard emission of the wet flue gas desulfurization device is realized.
Preferably, the S4 further includes the following steps:
s4-1: dividing a desulphurization device operation parameter matrix into a training set matrix and a test set matrix according to time;
s4-2: inputting the selected parameter matrix of the desulfurization device of the training set into the training input of a machine learning algorithm model, and constructing an outlet SO of the wet flue gas desulfurization device2An algorithm model with the concentration as a prediction target, and an outlet SO corresponding to the operation parameter matrix of the desulfurization device of the training set2And (4) concentration.
Preferably, the alarm threshold and the alarm form in S6 may be set and modified according to actual use requirements.
Realize wet flue gas desulfurization device export SO2And (3) predicting and early warning the concentration, wherein the algorithm model participates in operation control, and the spraying amount of the absorption liquid and the adding amount of the desulfurizer are timely reduced. The power consumption of the circulating pump can be reduced by reducing the spraying amount of the absorption liquid. The risk of excessive feeding of the desulfurizer can be reduced by reducing the feeding amount of the desulfurizer, the utilization efficiency of the desulfurizer is improved, and the consumption of the desulfurizer is reduced. Therefore, the operation cost of the desulfurization device can be reduced by reducing the spraying amount of the absorption liquid and the adding amount of the desulfurizer.
The present invention has been described in relation to the above embodiments, which are only exemplary of the implementation of the present invention. It should be noted that the disclosed embodiments do not limit the scope of the invention. Rather, it is intended that all such modifications and variations be included within the spirit and scope of this invention.

Claims (6)

1. A flue gas desulfurization sulfur dioxide concentration prediction method based on a machine learning algorithm is characterized by comprising the following steps: the method comprises the following steps:
s1: collecting historical operating data of the wet flue gas desulfurization device as sample set data;
s2: arranging the sample set data acquired in the S1 into a type and a format required by machine learning, and cleaning the data to form sample data of the machine learning;
s2-1: deleting missing values in the sample set, and sequencing from small to large to obtain a sequence { X ] of each parameter1,X2,X3,……Xn}；
S2-2: let QL＝X(n/4),QU＝X(3n/4),IQR＝QU-QL(ii) a n/4 and 3n/4 are rounded off to obtain integers;
QLis the lower quartile; qUIs the upper quartile; IQR is a four-bit distance;
X(n/4)、X(3n/4)is a series of numbers { X1,X2,X3,……XnThe corresponding value;
s2-3: removal of greater than Q in the S2-1 sequenceU+1.5IQR and less than QL-a value of 1.5IQR, the rusty data constituting a new data sample for machine learning;
s3: according to the data of the flue gas of the wet desulphurization device, the data of the absorption spray liquid and the outlet SO2Establishing a machine learning algorithm model according to the relevance of the concentration;
s3-1: taking 20 time sequences as a group of input sequences, wherein each time node hi is obtained by calculation through a tree network structure, and the algorithm formula is as follows:
fi＝σ(wixt+bi)
ft＝σ(wfhi-1+bf)
pi＝tanh(wp[hi-1；xt]+bp)
ot＝tanh(wo[hi-1；xt]+bo)
hi＝tanh(ft*ot+pi*fi)
wherein f isiThe method is a mapping relation and is used for recording the information at the current moment;
σ is a sigmoid function, which is formulated as follows:
Figure FDA0003389359190000021
xtcharacteristic information input at the moment;
ftis a mapping relationship;
hi-1is the result of the information matrix operation at the previous moment;
[hi-1；xt]splicing the calculation result representing the previous moment and the characteristic information of the current moment;
the tanh function is as follows:
Figure FDA0003389359190000022
represents a matrix dot product operation;
wt, Wf, Wo, Wp, bt, bf, bo, bp are all parameter matrixes;
s3-2: an Attention mechanism is introduced, when a time sequence is input, more Attention is paid to information of different time points in the previous time period, and information with longer recording time is recorded; wherein the Attention is a selection mechanism, and selects the most relevant position information of the current node;
the score value of each node is the hidden state hi (h) of the input sequence1,h2…hk) (k is the number of information input per iteration) and the last node hi-1The score value is subjected to softmax weighted calculation: the formula is as follows:
Figure FDA0003389359190000031
the formula is a normalized exponential function and means gradient logarithm normalization of finite term discrete probability distribution;
Figure FDA0003389359190000032
the value of k is from 1 to T, and T is the length of the input time sequence;
Figure FDA0003389359190000033
represents a summation;
exp(et) Represents the power of an exponential function e;
wherein ek＝hi
S3-3: the correlation is maximal when the mean of all input features is extracted for the first 5-30 minutes (preferably 10-15 minutes), which is expressed as follows:
a(hj)＝tanh(WX+B)
w and B are parameter matrixes, the running state of the desulfurizing tower in the first ten minutes is extracted,
a(hj) Is the result after the front mapping;
wherein h isjInput as an average of the first 5-30 minutes (preferably 10-15 minutes) characteristics;
x is the average value of the characteristic parameters of the first 10 minutes;
s3-4: by the formula:
Figure FDA0003389359190000041
calculating to obtain a final result, namely outlet sulfur dioxide concentration;
wherein m is a selection information parameter matrix;
s4: training the algorithm model established in the step 3 by using the data of S2; the accuracy and precision of the result are improved by testing and adjusting the parameters of the algorithm model;
s5: inputting real-time operation data, and predicting the SO at the outlet of the wet flue gas desulfurization device by using the algorithm model obtained in S42Concentration; obtained preThe measured result is the concentration of SO2 at the outlet of the wet flue gas desulfurization device;
s6: judging the prediction result of S5, if the prediction result exceeds the set value, outputting an alarm event, and storing the algorithm model to enter the next prediction; if the set value is not exceeded, comparing the measured value with the actual value;
the set value is SO at the outlet of the desulfurizer2Concentration control value, currently implemented ultra-low emission standard is 35mg/m3；
S7: when the measured value is consistent with the predicted value, the algorithm model is stored to enter the next prediction; if the measured value is not consistent with the predicted value, the group of running data enters a historical database, and the algorithm model is carried out again; the algorithm model after training, testing and iteration is used for the outlet SO of the S5 wet flue gas desulfurization device2Predicting the concentration;
training and testing refer to repeating the working contents of the steps S3 and S4;
iteration is a process of forming a historical database by using an S7 rule and repeating the steps S2-S4 to obtain a new algorithm model to replace the original algorithm model.
2. The method for predicting the concentration of sulfur dioxide in flue gas desulfurization based on machine learning algorithm as claimed in claim 1, wherein: the historical operating data of the wet flue gas desulfurization device collected in the S1 comprises the temperature, pressure, flow, humidity, oxygen content and SO of flue gas entering and exiting the desulfurization device2Concentration and other smoke constituents; the temperature, pressure, flow, density, PH value and oxidation rate of the spraying liquid for washing and purifying the flue gas; flow rate, concentration, density, composition of the absorbent required for flue gas desulfurization.
3. The method for predicting the concentration of sulfur dioxide in flue gas desulfurization based on machine learning algorithm as claimed in claim 1, wherein: the method for cleaning the data in the S2 comprises a data-based method and a rule-based method; the data-based mode comprises a classification method, a neighbor method, a clustering method and a statistical method; the rule-based approaches include data being unchanged for long periods of time, data being too variable, measurement values being over-range, and data being over-threshold.
4. The method for predicting the concentration of sulfur dioxide in flue gas desulfurization based on machine learning algorithm as claimed in claim 1, wherein: the method for establishing the machine learning algorithm model in the step S3 adopts a Recurrent Neural Network (RNN) machine learning algorithm model.
5. The method for predicting the concentration of sulfur dioxide in flue gas desulfurization based on machine learning algorithm as claimed in claim 1, wherein: the S4 further includes the following steps:
s4-1: dividing a desulphurization device operation parameter matrix into a training set matrix and a test set matrix according to time;
s4-2: inputting the selected parameter matrix of the desulfurization device of the training set into the training input of a machine learning algorithm model, and constructing an outlet SO of the wet flue gas desulfurization device2An algorithm model with the concentration as a prediction target, and an outlet SO corresponding to the operation parameter matrix of the desulfurization device of the training set2And (4) concentration.
6. The method for predicting the concentration of sulfur dioxide in flue gas desulfurization based on machine learning algorithm as claimed in claim 1, wherein: the alarm threshold and the alarm form in the S6 can be set and modified according to actual use requirements.
CN202111473270.0A 2021-12-02 2021-12-02 Flue gas desulfurization sulfur dioxide concentration prediction method based on machine learning algorithm Pending CN114177747A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202111473270.0A CN114177747A (en) 2021-12-02 2021-12-02 Flue gas desulfurization sulfur dioxide concentration prediction method based on machine learning algorithm

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202111473270.0A CN114177747A (en) 2021-12-02 2021-12-02 Flue gas desulfurization sulfur dioxide concentration prediction method based on machine learning algorithm

Publications (1)

Publication Number Publication Date
CN114177747A true CN114177747A (en) 2022-03-15

Family

ID=80603392

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202111473270.0A Pending CN114177747A (en) 2021-12-02 2021-12-02 Flue gas desulfurization sulfur dioxide concentration prediction method based on machine learning algorithm

Country Status (1)

Country Link
CN (1) CN114177747A (en)

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115228273A (en) * 2022-08-12 2022-10-25 中国华电科工集团有限公司 Control method of seawater desulfurization system
CN115414770A (en) * 2022-09-15 2022-12-02 杭州楚环科技股份有限公司 Garbage waste gas treatment method
CN116832588A (en) * 2023-08-25 2023-10-03 湖北鼎信成套设备有限公司 Acid regeneration flue gas purifying device and method thereof

Citations (11)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20180060665A1 (en) * 2016-08-29 2018-03-01 Nec Laboratories America, Inc. Dual Stage Attention Based Recurrent Neural Network for Time Series Prediction
CN109343367A (en) * 2018-10-26 2019-02-15 齐鲁工业大学 A method of based on network response surface flue gas desulfurization
CN109948869A (en) * 2019-04-22 2019-06-28 东南大学 Desulphurization system SO based on orderly cluster discretization2Exit concentration prediction technique
CN110162553A (en) * 2019-05-21 2019-08-23 南京邮电大学 Users' Interests Mining method based on attention-RNN
CN110570044A (en) * 2019-09-16 2019-12-13 重庆大学 next-place prediction method based on recurrent neural network and attention mechanism
CN111753145A (en) * 2020-06-10 2020-10-09 西北工业大学 Mobile application use prediction method based on time sequence mode
WO2020258611A1 (en) * 2019-06-28 2020-12-30 山东科技大学 Lymph node ct detection system employing recurrent spatio-temporal attention mechanism
CN112288137A (en) * 2020-10-09 2021-01-29 国网电力科学研究院有限公司 LSTM short-term load prediction method and device considering electricity price and Attention mechanism
CN112423863A (en) * 2018-07-13 2021-02-26 三菱动力株式会社 Method and device for controlling wet flue gas desulfurization device, and remote monitoring system provided with device for controlling wet flue gas desulfurization device
CN113082954A (en) * 2021-04-07 2021-07-09 浙江大学 Whole-process intelligent operation regulation and control system of wet desulphurization device
CN113469449A (en) * 2021-07-13 2021-10-01 浙江菲达环保科技股份有限公司 Optimizing control method and system for desulfurization system

Patent Citations (11)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20180060665A1 (en) * 2016-08-29 2018-03-01 Nec Laboratories America, Inc. Dual Stage Attention Based Recurrent Neural Network for Time Series Prediction
CN112423863A (en) * 2018-07-13 2021-02-26 三菱动力株式会社 Method and device for controlling wet flue gas desulfurization device, and remote monitoring system provided with device for controlling wet flue gas desulfurization device
CN109343367A (en) * 2018-10-26 2019-02-15 齐鲁工业大学 A method of based on network response surface flue gas desulfurization
CN109948869A (en) * 2019-04-22 2019-06-28 东南大学 Desulphurization system SO based on orderly cluster discretization2Exit concentration prediction technique
CN110162553A (en) * 2019-05-21 2019-08-23 南京邮电大学 Users' Interests Mining method based on attention-RNN
WO2020258611A1 (en) * 2019-06-28 2020-12-30 山东科技大学 Lymph node ct detection system employing recurrent spatio-temporal attention mechanism
CN110570044A (en) * 2019-09-16 2019-12-13 重庆大学 next-place prediction method based on recurrent neural network and attention mechanism
CN111753145A (en) * 2020-06-10 2020-10-09 西北工业大学 Mobile application use prediction method based on time sequence mode
CN112288137A (en) * 2020-10-09 2021-01-29 国网电力科学研究院有限公司 LSTM short-term load prediction method and device considering electricity price and Attention mechanism
CN113082954A (en) * 2021-04-07 2021-07-09 浙江大学 Whole-process intelligent operation regulation and control system of wet desulphurization device
CN113469449A (en) * 2021-07-13 2021-10-01 浙江菲达环保科技股份有限公司 Optimizing control method and system for desulfurization system

Cited By (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115228273A (en) * 2022-08-12 2022-10-25 中国华电科工集团有限公司 Control method of seawater desulfurization system
CN115414770A (en) * 2022-09-15 2022-12-02 杭州楚环科技股份有限公司 Garbage waste gas treatment method
CN115414770B (en) * 2022-09-15 2023-12-01 杭州楚环科技股份有限公司 Garbage waste gas treatment method
CN116832588A (en) * 2023-08-25 2023-10-03 湖北鼎信成套设备有限公司 Acid regeneration flue gas purifying device and method thereof
CN116832588B (en) * 2023-08-25 2024-02-02 湖北鼎信成套设备有限公司 Acid regeneration flue gas purifying device and method thereof

Similar Documents

Publication Publication Date Title
CN114177747A (en) Flue gas desulfurization sulfur dioxide concentration prediction method based on machine learning algorithm
CN108549792B (en) Soft measurement method for dioxin emission concentration in solid waste incineration process based on latent structure mapping algorithm
CN111860701B (en) Denitration system working condition discrimination preprocessing method based on clustering method
CN111240282B (en) Process optimization method, device, equipment and computer readable storage medium
CN111921377B (en) SCR denitration ammonia injection control method and system based on mechanism and data driving
Nkulikiyinka et al. Prediction of sorption enhanced steam methane reforming products from machine learning based soft-sensor models
CN111260149B (en) Dioxin emission concentration prediction method
CN110135057B (en) Soft measurement method for dioxin emission concentration in solid waste incineration process based on multilayer characteristic selection
CN114191953A (en) Flue gas desulfurization and denitrification control method based on convolutional neural network and XGboost
CN113175678B (en) Garbage incineration monitoring method and device
CN114225662B (en) Hysteresis model-based flue gas desulfurization and denitrification optimal control method
CN112650063B (en) Self-adaptive soft measurement method based on semi-supervised incremental Gaussian mixture regression
Sun et al. Prediction of oxygen content using weighted PCA and improved LSTM network in MSWI process
CN106680428A (en) Soft measuring method for denitration control system
Lee et al. Optimizing the specific surface area of fly ash-based sorbents for flue gas desulfurization
CN112733876A (en) Soft measurement method for nitrogen oxides in urban solid waste incineration process based on modular neural network
CN111589301A (en) Method, device, equipment and storage medium for predicting SCR denitration performance of coal-fired power plant
CN116688754A (en) Ship flue gas desulfurization automatic control system and method thereof
CN114330845A (en) MSWI process dioxin emission prediction method based on multi-window concept drift detection
Hao et al. Prediction of nitrogen oxide emission concentration in cement production process: a method of deep belief network with clustering and time series
CN114266461A (en) MSWI process dioxin emission risk early warning method based on visual distribution GAN
CN114398836A (en) MSWI process dioxin emission soft measurement method based on width mixed forest regression
CN113780383A (en) Dioxin emission concentration prediction method based on semi-supervised random forest and deep forest regression integration
CN114169235A (en) Machine learning algorithm-based flue gas desulfurization oxidation system fault prediction method
CN115775042A (en) Crop heavy metal enrichment risk prediction method and system based on Bayesian theory

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
CB02 Change of applicant information

Address after: 224051 innovation center, 42 environmental protection Avenue, environmental protection science and Technology City, Tinghu District, Yancheng City, Jiangsu Province

Applicant after: Jiangsu Kunlun Internet Technology Co.,Ltd.

Address before: 224051 innovation center, 42 environmental protection Avenue, environmental protection science and Technology City, Tinghu District, Yancheng City, Jiangsu Province

Applicant before: Kunyue Internet Environmental Technology (Jiangsu) Co.,Ltd.

CB02 Change of applicant information