CN114881355A - Extreme learning machine-based multi-condition prediction method for desulfurization system - Google Patents

Abstract

The invention provides a desulfurization system multi-working-condition prediction method based on an extreme learning machine, which comprises the steps of obtaining historical data of operating parameters of a boiler and a desulfurization system of a thermal power plant, classifying working conditions according to six mark position characteristic parameters, and determining corresponding working condition training data sets; establishing classification models under different working conditions based on a Bayesian algorithm aiming at a training data set, and selecting a proper modeling sample set; estimating delay time of the variable by utilizing the Pearson coefficient, and recombining the modeling data through the delay time; establishing different marking working conditions SO based on extreme learning machine algorithm by using recombined data set ₂ A prediction model; inputting the real-time data of the system into a classification model, judging the real-time working condition class of the system, and adoptingPredicting SO with corresponding prediction model ₂ The concentration of the emission. According to the method, on the premise of reducing the influence of delay time, the current working condition of the system is judged based on the boiler operation data, different working condition prediction models are switched, and the prediction efficiency and accuracy of the outlet SO2 are improved.

Description

Desulfurization system multi-working-condition prediction method based on extreme learning machine

Technical Field

The invention relates to the technical field of computer modeling prediction, in particular but not limited to a desulfurization system multi-condition prediction method based on an extreme learning machine.

Background

With the increasing prominence of environmental problems, under the environmental protection policy of strict sulfur dioxide emission control in China, Flue Gas Desulfurization (FGD) is widely applied to thermal power plants, so that the emission of atmospheric pollutants in power plants in China is greatly reduced, and the environment is improved.

Because the desulfurization process presents the characteristics of multiple working conditions and large hysteresis, the parameter values recorded at the current moment of a Distributed Control System (DCS) cannot accurately predict the concentration of SO2 in the clean flue gas, SO that the delay time of key parameters in the system and the working conditions of the desulfurization process become important factors influencing the establishment of an SO2 concentration model. With the increasing automation degree of power plants, the states and parameters of boiler units and other flue gas treatment equipment can be detected by sensors and recorded in a database, but the data are not well utilized. These data therefore provide new ideas for us to optimize the SO2 prediction model.

In view of the above, there is a need to provide a new prediction method to solve at least some of the above problems.

Disclosure of Invention

Aiming at one or more problems in the prior art, the invention provides a desulfurization system multi-working-condition prediction method based on an extreme learning machine, which is used for solving the variable working condition and large delay of a wet desulfurization system in a thermal power plantIn case of failure to accurately predict the export SO ₂ The problem of concentration.

The technical solution for realizing the purpose of the invention is as follows:

a desulfurization system multi-condition prediction method based on an extreme learning machine comprises the following steps:

step 1: acquiring historical data of operating parameters of a boiler unit and a desulfurization system, selecting marking position characteristic parameters for determining different operating conditions of the boiler, and classifying the operating conditions of the boiler and the desulfurization system of the thermal power plant according to the marking position characteristic parameters;

step 2: training historical data under different working conditions based on a Bayesian classification algorithm, and establishing naive Bayesian classification models under different working conditions;

and step 3: estimating delay time of each parameter of the flue gas, the slurry density of the absorption tower and the slurry circulating pump flow by using a Pearson coefficient method, and performing recombination pretreatment on the historical data sample according to the delay time to obtain a reconstructed data sample set;

and 4, step 4: based on the reconstructed data sample set, respectively establishing SO of different working conditions by adopting extreme learning machine algorithm ₂ A prediction model;

and 5: obtaining real-time operation data of the system, inputting the real-time operation data into a naive Bayesian classification model to obtain real-time working condition classes of the desulfurization system, and switching to SO of corresponding working conditions according to different working condition classes ₂ A prediction model;

step 6: inputting the real-time operation data of the system into a corresponding prediction model to obtain SO ₂ A predicted value of the emission concentration.

Optionally, the flag bit feature parameters in step 1 include: the method comprises the following steps of boiler load W, gas turbine rotating speed r, boiler outlet flue gas oxygen content Q, boiler outlet flue gas flow F, boiler outlet flue gas temperature T, boiler outlet flue gas flow velocity V, primary fan current I, fuel input quantity M, steam temperature T and pressure P, wherein the classification of working conditions comprises 6 working conditions such as starting, grid connection/energy supply, cracking, blowing out, stopping, accidents and the like.

Optionally, the state characteristics of the 6 operating conditions are as follows:

state characteristics of the starting conditions: starting the boiler, gradually reducing the oxygen content Q of the flue gas at the outlet of the boiler, gradually increasing the flow speed V of the flue gas at the outlet of the boiler to be more than 2m/s, and gradually increasing the temperature T of the flue gas at the outlet of the boiler to be more than 40 ℃;

state characteristics of grid connection/energy supply working conditions: the rotating speed r of the gas turbine is approximately equal to 3000r/min, the flue gas temperature T of the outlet of the boiler is approximately equal to 540 ℃, the current I of a primary fan is greater than 0, and the fan is started;

state characteristics of the cracking condition: the fuel input amount M is gradually reduced, and the boiler load W is gradually reduced to 50%;

the state characteristics of the furnace shutdown working condition: the flow velocity V of the boiler outlet flue gas, the flow F of the boiler outlet flue gas and the temperature T of the boiler outlet flue gas are gradually reduced, and the oxygen content Q of the boiler outlet flue gas is gradually increased;

state characteristics of the shutdown condition: the oxygen content Q of the flue gas at the outlet of the boiler is more than 18 percent, the flow velocity V of the flue gas at the outlet of the boiler is less than 1.8m/s, the flow F of the flue gas at the outlet of the boiler is less than 15 percent of the rated working condition, and the temperature T of the flue gas at the outlet of the boiler is less than 30 ℃;

state characteristics of accident conditions: the exhaust gas outlet automatically monitors data exceeding standard and abnormal fluctuation of partial equipment parameters.

Optionally, the step 2 of establishing the naive bayes classification model under different working conditions specifically includes:

step 2-1: selecting 10 key variables such as boiler load W, gas turbine rotating speed r, boiler outlet flue gas oxygen content Q, boiler outlet flue gas flow F, boiler outlet flue gas temperature T, boiler outlet flue gas flow velocity V, primary fan current I, fuel input quantity M, steam temperature T, pressure P and the like as characteristic attributes of a classification model, taking historical data corresponding to the characteristic attributes as training samples of the classification model, and marking the working condition types of the training samples according to the state characteristics of the working conditions;

step 2-2: respectively calculating the frequency P (working condition) of the occurrence of the samples of each working condition in the training samples, dividing the samples of each characteristic attribute according to the critical value of the characteristic attribute when the working condition changes, calculating the frequency P (variable) of the divided samples of each characteristic attribute in the training samples, and then calculating the conditional probability P (variable | working condition) of the divided samples of each characteristic attribute in the occurrence of each working condition;

step 2-3: based on a naive Bayes theory, calculating the conditional probability of the sample of each working condition appearing under each divided characteristic attribute:

step 2-4: calculating the probability sum of the training samples under each working condition:

s (working condition) ═ Σ P (working condition | variable)

And the working condition type corresponding to the maximum value of the S (working condition) is the working condition type to which the sample belongs, and the working condition type number of the sample is output.

Optionally, the critical value of the characteristic attribute in step 2-2 when the operating condition changes includes:

boiler load W: { W is less than or equal to 20%, W is less than or equal to 20% < 50%, W is more than or equal to 50% }

Gas turbine rotational speed r: { r is less than or equal to 2000, 2000< r <3500, r is more than or equal to 3500}

Oxygen content Q of boiler outlet flue gas: { Q is less than or equal to 5%, Q is more than 5% and less than 10%, and Q is more than or equal to 10% }

Boiler outlet flue gas flow F: { F <0.15 × F, F ≧ 0.15 × F }

Boiler outlet flue gas temperature T: { T is less than or equal to 40 ℃, T is more than 40 ℃ and less than 540 ℃, and T is more than or equal to 540 }

Boiler outlet flue gas flow velocity V: { V <1.8m/s, V ≧ 1.8m/s }

Primary fan current I: { I < 0.15I, I ≧ 0.15I }

Fuel charge amount M: { M < 0.15M, M ≧ 0.15M }

Steam temperature T': { T ' is less than or equal to 400 ℃, T ' is less than or equal to 600 ℃ at the temperature of 400 ℃, and T ' is more than or equal to 600 }

Steam pressure P: { P is less than or equal to 10kPa, P is less than 25kPa and is more than or equal to 25kPa when 10kPa is less than or equal to 10 kPa.

Optionally, the specific step of obtaining the reconstructed data sample set in step 3 includes:

step 3-1: defining the working condition parameters in the historical data as input variable X and SO at the outlet of the desulfurization system ₂ Concentration is an output variable Y, where X ═ X ₁ (t),…，x _m (t)]M is the dimension of the input variable, Y (t), t is 1 … N, and N is the number of samples of the historical data;

step 3-2: determining the maximum delay time k of the ith working condition parameter, and inputting the ith variable x in the maximum delay time range _i (t) carrying out reconstruction segmentation according to the acquisition time interval to obtain k groups of reconstruction working condition data sequences, wherein x is respectively _i (t-1)、…、x _i (t-k)；

Step 3-3: respectively calculating k groups of reconstruction working condition data sequences and SO ₂ Pearson's coefficient between concentrations y (t);

step 3-4: selecting delay time corresponding to a group of reconstruction working condition data sequences with the largest Pearson coefficient as a prediction variable x _i (t) and using the set of reconstructed condition data sequences as an input variable x _i (t) reconstructed data sequence, input variable x _i (t) the reconstructed dimension is 1;

step 3-5: and repeating the steps 3-2 to 3-4, calculating the optimal delay time of all the input variables X and the reconstruction data sequences thereof, combining the reconstruction data sequences of all the input variables X to obtain a reconstruction data sample, and using the reconstruction data sample as an input sample of the prediction model.

Optionally, the formula for calculating the pearson coefficient in step 3-3 is:

wherein, X _i For reconstructing parameters of the sequence of operating condition data, Y _i Is an outlet SO ₂ The concentration of the active ingredients in the mixture is,

to reconstruct the parameter average of the operating condition data sequence,

is an outlet SO ₂ Average value of concentration.

Optional, step(s)4 establishing SO under different working conditions ₂ The prediction model comprises the following specific steps:

step 4-1: acquiring a reconstructed data sample set under a working condition, and taking historical data of characteristic variables in the reconstructed data sample set as an extreme learning machine SO ₂ An input sample matrix of the prediction model;

step 4-2: establishing a single-layer feedforward neural network for the sample of the characteristic variable related to the working condition, wherein the single-layer feedforward neural network prediction model comprises 6 input layers, L hidden layer nodes and 1 output layer node;

step 4-3: randomly generating weight matrix W of input layer and hidden layer of single-layer feedforward neural network _i Obtaining a hidden layer output matrix H with 6 rows and L columns and an output weight matrix beta with L rows and 1 columns;

step 4-4: to SO ₂ An L2 regularization term is introduced into an error function of the prediction model, and the minimum error function of the model is obtained as follows:

wherein C is a regularization coefficient, H is a hidden layer output matrix, T is a training target, β is an output weight matrix, and | is a ferobenius norm of a matrix element;

and 4-5: calculating to obtain a weight matrix beta of the prediction model of the extreme learning machine ^* ：

SO under that condition ₂ The prediction model is:

Y＝XW _i β ^*

wherein X is the input matrix and Y is SO ₂ A predicted value of emission concentration;

and 4-6: repeating the steps 4-1 to 4-5 to respectively establish SO under different working conditions ₂ And (4) predicting the model.

Optionally, in step 4-1, the input sample matrices and the corresponding characteristic variables under different working conditions are respectively:

the input sample matrix under the starting condition is as follows: x ═ Q, V, T, ρ, F', C;

the input sample matrix under the grid-connected/functional working condition is as follows: x ═ r, T, I, ρ, F', C;

the input sample matrix under the splitting condition is: x ═ Q, M, W, ρ, F', C;

the input sample matrix under the blowing-out condition is as follows: x ═ Q, V, T, ρ, F', C;

wherein, p is the slurry density of the absorption tower, F' is the slurry circulating pump flow, C is the concentration of sulfur dioxide at the outlet of the boiler, W is the load of the boiler, r is the rotating speed of the gas turbine, Q is the oxygen content of the flue gas at the outlet of the boiler, T is the temperature of the flue gas at the outlet of the boiler, V is the flow velocity of the flue gas at the outlet of the boiler, I is the current of a primary fan, and M is the input amount of fuel.

Compared with the prior art, the invention adopting the technical scheme has the following technical effects:

the desulfurization system multi-working-condition prediction method based on the extreme learning machine recombines data and establishes a prediction model by adopting historical data to establish a working condition classification model and estimating the delay time of a variable, thereby realizing that the influence of the delay time is reduced, simultaneously judging the current working condition of the system based on the real-time operation data of the boiler and switching the prediction models under different working conditions, and effectively improving the SO outlet ₂ Efficiency and accuracy of the prediction.

Drawings

The accompanying drawings are included to provide a further understanding of the invention, and are incorporated in and constitute a part of this specification, illustrate embodiments of the invention and together with the description serve to explain the principles of the invention. In the drawings:

FIG. 1 is a process flow diagram of a thermal power plant desulfurization system to which the extreme learning machine-based desulfurization system multi-condition prediction method of the present invention is applied.

FIG. 2 is a flow chart of the extreme learning machine-based desulfurization system multi-condition prediction method for establishing classification models under different conditions by using a naive Bayes classification algorithm.

FIG. 3 is a flow chart of data preprocessing by using a Pearson coefficient method according to the extreme learning machine-based desulfurization system multi-condition prediction method of the present invention.

FIG. 4 is a flow chart of the modeling of the index position multi-condition of the extreme learning machine-based desulfurization system multi-condition prediction method of the present invention.

FIG. 5 is a flow chart of extreme learning machine algorithm modeling of the extreme learning machine-based desulfurization system multi-condition prediction method of the present invention.

FIG. 6 shows SO of the extreme learning machine-based desulfurization system under the start-up condition ₂ And (5) a prediction result graph.

FIG. 7 is a diagram of the SO2 prediction result of the extreme learning machine-based desulfurization system multi-condition prediction method applied under the grid-connected/energy-supplied condition.

FIG. 8 is a diagram of the prediction results of the extreme learning machine-based desulfurization system multi-condition prediction method applied to the disaggregated conditions.

FIG. 9 is a diagram of the prediction results of the extreme learning machine-based desulfurization system multi-condition prediction method applied under the blowing-out condition.

FIG. 10 is a general flowchart of the extreme learning machine-based desulfurization system multi-condition prediction method of the present invention.

The reference numerals have the meanings given below: 1: a boiler; 2: a superheater; 3: a steam drum; 4: a gas turbine; 5: a desulfurizer preparation tank; 6: a pulping box; 7: a desulfurizing tower; 8: and a slurry circulating pump.

Detailed Description

For a further understanding of the invention, reference will now be made to the preferred embodiments of the invention by way of example, and it is to be understood that the description is intended to further illustrate features and advantages of the invention, and not to limit the scope of the claims.

The description in this section is for several exemplary embodiments only, and the present invention is not limited only to the scope of the embodiments described. Combinations of different embodiments, and substitutions of features from different embodiments, or similar prior art means may be substituted for or substituted for features of the embodiments shown and described.

Fig. 1 shows a flow chart of a wet desulfurization process of a thermal power plant to which the present invention is applied. The states of the incineration boiler of the thermal power plant are divided into: starting, grid connection/energy supply, cracking, furnace shutdown, accidents and other six states, and different working conditions can be distinguished according to the parameter characteristics of a boiler, a steam drum, a desulfurization facility and the like and the flue gas parameter characteristics.

According to one aspect of the invention, a method for predicting multiple working conditions of a desulfurization system based on an extreme learning machine is shown in fig. 10, and comprises the following steps:

step 1: acquiring historical data of operating parameters of a boiler unit and a desulfurization system, wherein the historical data can be acquired from a Distributed Control System (DCS) system of an enterprise central control room, and a historical data set is used as a training set of a model;

selecting and determining marking position characteristic parameters of different operating conditions of the boiler, wherein the marking position characteristic parameters comprise: boiler load W, gas turbine rotating speed r, boiler outlet flue gas oxygen content Q, boiler outlet flue gas flow F, boiler outlet flue gas temperature T, boiler outlet flue gas flow velocity V, primary fan current I, fuel input amount M, steam temperature T and pressure P.

And classifying the working conditions of the boiler and the desulfurization system of the thermal power plant according to the characteristic parameters of the marker bits, wherein the classification of the working conditions comprises 6 working condition states of starting, grid connection/energy supply, cracking, furnace shutdown, accidents and the like, and the working conditions of the boiler and the desulfurization system of the thermal power plant are respectively judged according to eight operation variables of boiler load W, gas turbine rotating speed r, boiler outlet flue gas oxygen content Q, boiler outlet flue gas flow F, boiler outlet flue gas temperature T, boiler outlet flue gas flow speed V, primary fan current I and fuel input quantity M, and the state characteristics of the 6 working conditions are respectively as follows:

state characteristics of the starting conditions: starting the boiler, gradually reducing the oxygen content Q of the flue gas at the outlet of the boiler, gradually increasing the flow speed V of the flue gas at the outlet of the boiler to be more than 2m/s, and gradually increasing the temperature T of the flue gas at the outlet of the boiler to be more than 40 ℃;

state characteristics of grid connection/energy supply working conditions: the rotating speed r of the gas turbine is approximately equal to 3000r/min, the flue gas temperature T of the outlet of the boiler is approximately equal to 540 ℃, the current I of a primary fan is greater than 0, and the fan is started;

state characteristics of the cracking condition: the fuel input amount M is gradually reduced, and the boiler load W is gradually reduced to 50%;

the state characteristics of the furnace shutdown working condition: the flow velocity V of the boiler outlet flue gas, the flow F of the boiler outlet flue gas and the temperature T of the boiler outlet flue gas are gradually reduced, and the oxygen content Q of the boiler outlet flue gas is gradually increased;

state characteristics of the shutdown condition: the oxygen content Q of the flue gas at the outlet of the boiler is more than 18 percent, the flow velocity V of the flue gas at the outlet of the boiler is less than 1.8m/s, the flow F of the flue gas at the outlet of the boiler is less than 15 percent of the rated working condition, and the temperature T of the flue gas at the outlet of the boiler is less than 30 ℃;

state characteristics of accident conditions: the exhaust gas outlet automatically monitors data exceeding standard and abnormal fluctuation of partial equipment parameters.

Step 2: training historical data under different working conditions based on a Bayesian classification algorithm, and establishing naive Bayes classification models under different working conditions; as shown in fig. 2, the specific steps include:

step 2-1: selecting 10 key variables such as boiler load W, gas turbine rotating speed r, boiler outlet flue gas oxygen content Q, boiler outlet flue gas flow F, boiler outlet flue gas temperature T, boiler outlet flue gas flow velocity V, primary fan current I, fuel input quantity M, steam temperature T, pressure P and the like as characteristic attributes of a classification model, taking historical data corresponding to the characteristic attributes as training samples of the classification model, and marking the working condition types of the training samples according to the state characteristics of the working conditions in the step 1;

step 2-2: respectively calculating the occurrence frequencies P (starting), P (grid connection), P (disconnection) and P (furnace shutdown) of samples marked as starting, grid connection/energy supply, disconnection and furnace shutdown working conditions in training samples, dividing the samples of each characteristic attribute according to the critical value of the characteristic attribute when the working conditions change, calculating the occurrence frequency P (variable) of the divided samples of each characteristic attribute in the training samples, and calculating the occurrence conditional probability P (variable | working conditions) of the divided samples of each characteristic attribute under each working condition; the critical values of the characteristic attributes during the working condition change are specifically divided as follows:

boiler load W: { W is less than or equal to 20%, W is less than or equal to 20% < 50%, W is more than or equal to 50% }

Gas turbine rotational speed r: { r is less than or equal to 2000, 2000< r <3500, r is more than or equal to 3500}

Oxygen content Q of boiler outlet flue gas: { Q is less than or equal to 5%, Q is more than 5% and less than 10%, Q is more than or equal to 10% }

Boiler outlet flue gas flow F: { F <0.15 × F, F ≧ 0.15 × F }

Boiler outlet flue gas temperature T: { T is less than or equal to 40 ℃, T is more than 40 ℃ and less than 540 ℃, and T is more than or equal to 540 }

Boiler outlet flue gas flow velocity V: { V <1.8m/s, V ≧ 1.8m/s }

Primary fan current I: { I < 0.15I, I ≧ 0.15I }

Fuel charge amount M: { M < 0.15M, M ≧ 0.15M }

Steam temperature T': { T ' is less than or equal to 400 ℃, T ' is less than or equal to 600 ℃ at the temperature of 400 ℃, and T ' is more than or equal to 600 }

Steam pressure P: { P is less than or equal to 10kPa, P is less than 25kPa and is more than or equal to 25kPa when 10kPa is less than or equal to 10kPa };

step 2-3: based on a naive Bayes theory, calculating the conditional probability of the sample of each working condition appearing under each divided characteristic attribute:

obtaining the probability P (starting | variable), P (grid-connected | variable), P (splitting | variable) and P (blowing-out | variable) of each characteristic variable belonging to each working condition;

step 2-4: calculating the probability sum of the training samples under each working condition:

s (working condition) ═ Σ P (working condition | variable)

The working condition type corresponding to the maximum value of the S (working condition) is the working condition type to which the sample belongs, the numbers of the four working condition types of starting, grid connection/energy supply, splitting and blowing out are 1, 2, 3 and 4, and the working condition type numbers of the sample are output. In one embodiment, for 6-month operation data of a 500MW unit flue gas desulfurization system of a certain thermal power plant, 400 groups of training samples of the model are determined after screening, wherein 100 groups of four working conditions of starting, grid connection/energy supply, splitting and furnace shutdown are respectively adopted. The data of P (variable | condition) obtained by dividing and calculating the characteristic attributes are shown in the following table:

TABLE 1P statistical table (variable | Condition)

And step 3: estimating delay time of each parameter of the flue gas, the slurry density of the absorption tower and the slurry circulating pump flow by using a Pearson coefficient method, and performing recombination pretreatment on the historical data sample according to the delay time to obtain a reconstructed data sample set; as shown in fig. 3, the specific steps include:

step 3-1: defining the working condition parameters in the historical data as input variable X and SO at the outlet of the desulfurization system ₂ Concentration is an output variable Y, where X ═ X ₁ (t),…，x _m (t)]M is the dimension of the input variable, Y (t), t is 1 … N, and N is the number of samples of the historical data;

step 3-2: determining the maximum delay time k of the ith working condition parameter, and inputting the ith variable x in the maximum delay time range _i (t) carrying out reconstruction segmentation according to the acquisition time interval to obtain k groups of reconstruction working condition data sequences, wherein x is respectively _i (t-1)、…、x _i (t-k)；

Step 3-3: respectively calculating k groups of reconstruction working condition data sequences and SO ₂ Pearson's coefficient between concentrations y (t); the formula for calculating the pearson coefficient is:

wherein, X _i For reconstructing parameters of the sequence of operating condition data, Y _i Is an outlet SO ₂ The concentration of the active ingredients in the mixture is,

to reconstruct the parameter average of the operating condition data sequence,

is an outlet SO ₂ The average value of the concentration;

step 3-4: the correlation degree between the group of data sequences with the maximum Pearson coefficient and the output y (t) is maximum, and the corresponding delay time is the prediction variable x _i (t) selecting the delay time corresponding to a group of reconstruction working condition data sequences with the largest Pearson coefficient as a prediction variable x _i (t) and using the set of reconstructed condition data sequences as an input variable x _i (t) reconstructed data sequence, input variable x _i (t) the reconstructed dimension is 1;

step 3-5: and repeating the steps 3-2 to 3-4, calculating the optimal delay time of all the input variables X and the reconstruction data sequences thereof, combining the reconstruction data sequences of all the input variables X to obtain a reconstruction data sample, and using the reconstruction data sample as an input sample of the prediction model.

In one embodiment, the delay time analysis is performed on the condition parameters required for modeling, and an optimal delay time statistical table of each input variable is obtained as follows:

TABLE 2 statistical table of delay times of variables

Variable names

Q

V

T

ρ

F′

C

I

M

W

Delay time

440s

120s

110s

50s

20s

170s

60s

55s

200s

And 4, step 4: based on the reconstructed data sample set, respectively establishing SO of different working conditions by adopting extreme learning machine algorithm ₂ A prediction model; fig. 5 shows a modeling flow of the extreme learning machine algorithm, which includes the following specific steps:

step 4-1: acquiring reconstructed data sample sets under different working conditions, and taking historical data of characteristic variables such as absorption tower slurry density rho, slurry circulating pump flow F', boiler outlet sulfur dioxide concentration C, boiler load W, gas turbine rotating speed r, boiler outlet flue gas oxygen content Q, boiler outlet flue gas temperature T, boiler outlet flue gas flow speed V, primary fan current I and fuel input M in the reconstructed data sample sets as a limit learning machine SO ₂ Inputting a sample matrix of the prediction model;

the input sample matrix and the corresponding characteristic variables under different working conditions are respectively as follows:

the input sample matrix under the starting condition is as follows: x ═ Q, V, T, ρ, F', C;

the input sample matrix under the grid-connected/functional working condition is as follows: x ═ r, T, I, ρ, F', C;

the input sample matrix under the splitting condition is: x ═ Q, M, W, ρ, F', C;

the input sample matrix under the blowing-out working condition is as follows: x ═ Q, V, T, ρ, F', C;

as shown in fig. 4, a flow chart of a flag bit multi-condition modeling is shown, and in an embodiment, before historical data of a desulfurization system and a corresponding boiler unit is obtained, by understanding contents of fuel, flue gas components and generation mechanisms used by the boiler unit, a process flow of the desulfurization system, and the like, possible conditions of the boiler unit are determined to be starting, grid connection/energy supply, cracking, furnace shutdown, accidents, and the like. The historical data may be obtained from a Distributed Control System (DCS) system of a central control room of the enterprise, and may be used as a training set of the model. And preprocessing historical data of the characteristic variables obtained by analysis to be used as input of a prediction model of the extreme learning machine.

Step 4-2: as shown in fig. 5, a single-layer feedforward neural network is established for samples of characteristic variables related to different working conditions, and the single-layer feedforward neural network prediction model includes 6 input layers, L hidden layer nodes, and 1 output layer node;

step 4-3: randomly generating a weight matrix W for a single-layer feedforward neural network _i Obtaining a hidden layer output matrix H with 6 rows and L columns and an output weight matrix beta with L rows and 1 columns; solving the output weights minimizes the error function of the model, which is as follows:

wherein, H is a hidden layer output matrix, T is a training target, beta is an output weight matrix, and | | is a Ferobenius norm of matrix elements;

step 4-4: to SO ₂ An L2 regularization term is introduced into an error function of the prediction model, and the minimum error function of the model is obtained as follows:

c is a regularization coefficient, H is a hidden layer output matrix, T is a training target, beta is an output weight matrix, and | | is a Ferobenius norm of a matrix element, and the problem that the minimum error function is equivalent to the ridge regression problem is solved;

and 4-5: calculating to obtain a weight matrix beta of the prediction model of the extreme learning machine ^* ：

SO under that condition ₂ The prediction model is:

Y＝XW _i β ^*

wherein X is the input matrix and Y is SO ₂ A predicted value of the emission concentration;

and 4-6: repeating the steps 4-1 to 4-5 to respectively establish SO under different working conditions ₂ And (4) predicting the model.

And 5: obtaining real-time operation data of the system, inputting the real-time operation data into a naive Bayesian classification model to obtain real-time working condition classes of the desulfurization system, and switching to SO of corresponding working conditions according to different working condition classes ₂ A prediction model;

step 6: inputting the real-time operation data of the system into the corresponding prediction model to obtain SO ₂ A predicted value of the emission concentration.

FIG. 6 shows SO in the start-up condition ₂ The result of emission concentration prediction, FIG. 7 is SO under grid-connected/power-supplied conditions ₂ FIG. 8 shows the result of emission concentration prediction for SO under the split condition ₂ FIG. 9 shows the result of prediction of emission concentration, and SO in blowing-out condition ₂ And (5) predicting the emission concentration.

The prediction method of the technical scheme is compared with the prediction structure error of the BP neural network under different working conditions as follows:

TABLE 2 accuracy comparison table

	Starting up	Grid connection/energy supply	Splitting column	Blowing out
					Prediction method of the present disclosure	94.5	91.2	94.3	93.2
BP neural network	83.9	85.5	87.2	89.9

As shown in the table, the prediction method of the technical scheme can effectively improve SO ₂ Accuracy of prediction of emission concentration.

In one embodiment, real-time data is input into an established working condition classification model to obtain a working condition of a boiler unit corresponding to a desulfurization system, and the predicted working condition of the boiler unit is compared with a manually marked working condition in a Distributed Control System (DCS) to identify whether a false marking condition exists. In another embodiment, real-time data is input into the established classification model and the discharge SO ₂ The concentration prediction model is used for obtaining SO at the outlet of the desulfurization system ₂ After the concentration prediction value, the prediction can be referred toThe preparation amount and the dosage of the lime slurry of the desulfurization system are set according to the value, so that the cost of enterprise flue gas treatment is reduced.

The description and applications of the invention herein are illustrative and are not intended to limit the scope of the invention to the embodiments described above. The descriptions related to the effects or advantages in the specification may not be reflected in practical experimental examples due to uncertainty of specific condition parameters or influence of other factors, and the descriptions related to the effects or advantages are not used for limiting the scope of the invention. Variations and modifications of the embodiments disclosed herein are possible, and alternative and equivalent various components of the embodiments will be apparent to those of ordinary skill in the art. It will be clear to those skilled in the art that the present invention may be embodied in other forms, structures, arrangements, proportions, and with other components, materials, and parts, without departing from the spirit or essential characteristics thereof. Other variations and modifications of the embodiments disclosed herein may be made without departing from the scope and spirit of the invention.

Claims (9)

1. A desulfurization system multi-condition prediction method based on an extreme learning machine is characterized by comprising the following steps:

step 1: acquiring historical data of operating parameters of a boiler unit and a desulfurization system, selecting marking position characteristic parameters for determining different operating conditions of the boiler, and classifying the operating conditions of the boiler and the desulfurization system of the thermal power plant according to the marking position characteristic parameters;

step 2: training historical data under different working conditions based on a Bayesian classification algorithm, and establishing naive Bayes classification models under different working conditions;

and step 3: estimating delay time of each parameter of the flue gas, the slurry density of the absorption tower and the slurry circulating pump flow by using a Pearson coefficient method, and performing recombination pretreatment on the historical data sample according to the delay time to obtain a reconstructed data sample set;

and 4, step 4: based on the reconstructed data sample set, respectively establishing SO of different working conditions by adopting extreme learning machine algorithm ₂ A prediction model;

step (ii) of5: obtaining real-time operation data of the system, inputting the real-time operation data into a naive Bayesian classification model to obtain real-time working condition classes of the desulfurization system, and switching to SO of corresponding working conditions according to different working condition classes ₂ A prediction model;

step 6: inputting the real-time operation data of the system into a corresponding prediction model to obtain SO ₂ A predicted value of the emission concentration.

2. The extreme learning machine-based desulfurization system multi-condition prediction method according to claim 1, wherein the marking bit characteristic parameters in the step 1 comprise: the method comprises the following steps of boiler load W, gas turbine rotating speed r, boiler outlet flue gas oxygen content Q, boiler outlet flue gas flow F, boiler outlet flue gas temperature T, boiler outlet flue gas flow velocity V, primary fan current I, fuel input quantity M, steam temperature T and pressure P, wherein the classification of working conditions comprises 6 working conditions such as starting, grid connection/energy supply, cracking, blowing out, stopping, accidents and the like.

3. The extreme learning machine-based desulfurization system multi-condition prediction method according to claim 2, characterized in that the state characteristics of the 6 conditions are as follows:

state characteristics of the starting conditions: starting the boiler, gradually reducing the oxygen content Q of the flue gas at the outlet of the boiler, gradually increasing the flow speed V of the flue gas at the outlet of the boiler to be more than 2m/s, and gradually increasing the temperature T of the flue gas at the outlet of the boiler to be more than 40 ℃;

state characteristics of grid connection/energy supply working conditions: the rotating speed r of the gas turbine is approximately equal to 3000r/min, the flue gas temperature T of the outlet of the boiler is approximately equal to 540 ℃, the current I of a primary fan is greater than 0, and the fan is started;

state characteristics of the cracking condition: the fuel input amount M is gradually reduced, and the boiler load W is gradually reduced to 50%;

the state characteristics of the furnace shutdown working condition: the flow velocity V of the boiler outlet flue gas, the flow F of the boiler outlet flue gas and the temperature T of the boiler outlet flue gas are gradually reduced, and the oxygen content Q of the boiler outlet flue gas is gradually increased;

state characteristics of the shutdown condition: the oxygen content Q of the flue gas at the outlet of the boiler is more than 18 percent, the flow velocity V of the flue gas at the outlet of the boiler is less than 1.8m/s, the flow F of the flue gas at the outlet of the boiler is less than 15 percent of the rated working condition, and the temperature T of the flue gas at the outlet of the boiler is less than 30 ℃;

state characteristics of accident conditions: the exhaust gas outlet automatically monitors data exceeding standard and abnormal fluctuation of partial equipment parameters.

4. The extreme learning machine-based desulfurization system multi-condition prediction method according to claim 1, wherein the specific steps of establishing naive Bayes classification models under different conditions in the step 2 comprise:

step 2-1: selecting 10 key variables such as boiler load W, gas turbine rotating speed r, boiler outlet flue gas oxygen content Q, boiler outlet flue gas flow F, boiler outlet flue gas temperature T, boiler outlet flue gas flow velocity V, primary fan current I, fuel input quantity M, steam temperature T, pressure P and the like as characteristic attributes of a classification model, taking historical data corresponding to the characteristic attributes as training samples of the classification model, and marking the working condition types of the training samples according to the state characteristics of the working conditions;

step 2-2: respectively calculating the frequency P (working condition) of the occurrence of the samples of each working condition in the training samples, dividing the samples of each characteristic attribute according to the critical value of the characteristic attribute when the working condition changes, calculating the frequency P (variable) of the divided samples of each characteristic attribute in the training samples, and then calculating the conditional probability P (variable | working condition) of the divided samples of each characteristic attribute in the occurrence of each working condition;

step 2-3: based on a naive Bayes theory, calculating the conditional probability of the sample of each working condition appearing under each divided characteristic attribute:

step 2-4: calculating the probability sum of the training samples under each working condition:

s (working condition) ═ Σ P (working condition | variable)

And the working condition type corresponding to the maximum value of the S (working condition) is the working condition type to which the sample belongs, and the working condition type number of the sample is output.

5. The extreme learning machine-based desulfurization system multi-condition prediction method according to claim 4, wherein the critical value of the characteristic attribute in step 2-2 when the condition changes comprises:

boiler load W: { W is less than or equal to 20%, W is less than or equal to 20% < 50%, W is more than or equal to 50% }

Gas turbine rotational speed r: { r is less than or equal to 2000, 2000< r <3500, r is more than or equal to 3500}

Oxygen content Q of boiler outlet flue gas: { Q is less than or equal to 5%, Q is more than 5% and less than 10%, and Q is more than or equal to 10% }

Boiler outlet flue gas flow F: { F <0.15 × F, F ≧ 0.15 × F }

Boiler outlet flue gas temperature T: { T is less than or equal to 40 ℃, T is more than 40 ℃ and less than 540 ℃, and T is more than or equal to 540 }

Boiler outlet flue gas flow velocity V: { V <1.8m/s, V ≧ 1.8m/s }

Primary fan current I: { I < 0.15I, I ≧ 0.15I }

Fuel charge amount M: { M < 0.15M, M ≧ 0.15M }

Steam temperature T': { T ' is less than or equal to 400 ℃, T ' is less than or equal to 600 ℃ at the temperature of 400 ℃, and T ' is more than or equal to 600 }

Steam pressure P: { P is less than or equal to 10kPa, P is less than 25kPa and is more than or equal to 25kPa when 10kPa is less than or equal to 10 kPa.

6. The extreme learning machine-based desulfurization system multi-condition prediction method according to claim 1, wherein the specific step of obtaining the reconstructed data sample set in the step 3 comprises:

step 3-1: defining the working condition parameters in the historical data as input variable X and SO at the outlet of the desulfurization system ₂ Concentration is an output variable Y, where X ═ X ₁ (t),…，x _m (t)]M is the dimension of the input variable, Y (t), t is 1 … N, and N is the number of samples of the historical data;

step 3-2: determining the maximum delay time k of the ith working condition parameter, and inputting the ith variable x in the maximum delay time range _i (t) carrying out reconstruction segmentation according to the acquisition time interval to obtain k groups of reconstruction working condition data sequences, wherein x is respectively _i (t-1)、…、x _i (t-k)；

Step 3-3: respectively calculating k groups of reconstruction working condition data sequences and SO ₂ Pearson's coefficient between concentrations y (t);

step 3-4: selecting delay time corresponding to a group of reconstruction working condition data sequences with the largest Pearson coefficient as a prediction variable x _i (t) and using the set of reconstructed condition data sequences as an input variable x _i (t) reconstructed data sequence, input variable x _i (t) the reconstructed dimension is 1;

step 3-5: and repeating the steps 3-2 to 3-4, calculating the optimal delay time of all the input variables X and the reconstruction data sequences thereof, combining the reconstruction data sequences of all the input variables X to obtain a reconstruction data sample, and using the reconstruction data sample as an input sample of the prediction model.

7. The extreme learning machine-based desulfurization system multi-condition prediction method according to claim 6, characterized in that the formula for calculating the Pearson coefficient in step 3-3 is as follows:

wherein, X _i For reconstructing parameters of the sequence of operating condition data, Y _i Is an outlet SO ₂ The concentration of the active ingredients in the mixture is,

to reconstruct the parameter average of the operating condition data sequence,

is an outlet SO ₂ Average value of concentration.

8. The extreme learning machine-based desulfurization system multi-condition prediction method according to claim 1, characterized in that SO of different conditions is established in step 4 ₂ The prediction model comprises the following specific steps:

step 4-1: obtain oneReconstructing the data sample set under the working condition, and taking the historical data of the characteristic variables in the reconstructed data sample set under the working condition as the extreme learning machine SO ₂ An input sample matrix of the prediction model;

step 4-2: establishing a single-layer feedforward neural network for the sample of the characteristic variable related to the working condition, wherein the single-layer feedforward neural network prediction model comprises 6 input layers, L hidden layer nodes and 1 output layer node;

step 4-3: randomly generating weight matrix W of input layer and hidden layer of single-layer feedforward neural network _i Obtaining a hidden layer output matrix H with 6 rows and L columns and an output weight matrix beta with L rows and 1 columns;

step 4-4: to SO ₂ An L2 regularization term is introduced into an error function of the prediction model, and the minimum error function of the model is obtained as follows:

wherein C is a regularization coefficient, T is a training target, and | is a ferobenius norm of a matrix element;

and 4-5: calculating to obtain a weight matrix beta of the prediction model of the extreme learning machine ^* ：

SO under that condition ₂ The prediction model is:

Y＝XW _i β ^*

wherein X is the input matrix and Y is SO ₂ A predicted value of emission concentration;

and 4-6: repeating the steps 4-1 to 4-5 to respectively establish SO under different working conditions ₂ And (4) predicting the model.

9. The extreme learning machine-based desulfurization system multi-condition prediction method according to claim 8, wherein in step 4-1, the input sample matrix and the corresponding characteristic variables under different conditions are respectively:

the input sample matrix under the starting condition is as follows: x ═ Q, V, T, ρ, F', C;

the input sample matrix under the grid-connected/functional working condition is as follows: x ═ r, T, I, ρ, F', C;

the input sample matrix under the splitting condition is: x ═ Q, M, W, ρ, F', C;

the input sample matrix under the blowing-out condition is as follows: x ═ Q, V, T, ρ, F', C;

wherein, p is the slurry density of the absorption tower, F' is the slurry circulating pump flow, C is the concentration of sulfur dioxide at the outlet of the boiler, W is the load of the boiler, r is the rotating speed of the gas turbine, Q is the oxygen content of the flue gas at the outlet of the boiler, T is the temperature of the flue gas at the outlet of the boiler, V is the flow velocity of the flue gas at the outlet of the boiler, I is the current of a primary fan, and M is the input amount of fuel.

Images