Prediction method for primary air volume of coal mill

Abstract

The invention discloses a coal mill primary air volume prediction method which comprises the following steps: 1) determining an input variable of a prediction model, and selecting a plurality of variables as auxiliary variables by performing mechanism analysis on a coal mill primary air supply system; 2) performing basic measurement unit normalization on training sample data by using a normalization method, removing data samples with high similarity by using a similarity function method, and performing redundancy processing on the data samples; the method comprises the steps of (1) extracting features among data through kernel principal component analysis to reduce the data dimensionality, (2) establishing a primary air volume prediction model based on the long short-term memory network, (3) extracting the features among the data through kernel principal component analysis to reduce the data dimensionality, and (3) establishing the primary air volume prediction model based on the long short-term memory network to improve the prediction precision.

Description

Technical Field

[0001] The invention relates to a prediction method, in particular to a method for predicting the primary air volume of a coal mill. Background Art

[0002] The current coal mill inlet air duct is very compact, with closely arranged dampers and the mixing of cold and hot air, which makes the fluid velocity distribution in the mixing section extremely uneven, causing the air volume measurement value to be opposite to the hot air door opening, making it impossible for the coal mill to be put into automatic operation, seriously affecting the safe operation of the coal-fired unit.

[0003] Coal-fired boilers in large thermal power plants generally use a direct-blowing pulverizing system. The stability and accuracy of the primary air volume measurement of the coal mill is crucial. It not only affects the accuracy of the air-coal ratio, but is also one of the important factors affecting the economy and stability of the combustion system. However, the primary air duct of a large coal mill generally has the problem of uneven flow velocity in the pipe, resulting in a large deviation between the air volume measurement and the actual value. For the measurement of primary air volume, the contact measurement method is currently generally used. Due to the complex on-site environment and many influencing factors, the measurement results often have serious drift. Summary of the invention

[0004] The purpose of the present invention is to overcome the defects of the prior art and provide a method for predicting the primary air volume of a coal mill, which extracts the features between data through kernel principal component analysis and reduces the data dimension; establishes a primary air volume prediction model based on a long short-term memory network to improve the prediction accuracy.

[0005] The object of the present invention is achieved in this way: a method for predicting the primary air volume of a coal mill comprises the following steps:

[0006] 1) Determine the input variables of the prediction model, and select multiple variables as auxiliary variables by analyzing the mechanism of the primary air supply system of the coal mill;

[0007] 2) Use the normalization method to unify the basic measurement units of the training sample data, use the similarity function method to remove data samples with high similarity, and perform redundant processing on the data samples;

[0008] 3) Use kernel principal component analysis to extract features between data and reduce data dimensions;

[0009] 4) Establish a primary air volume prediction model based on long short-term memory network.

[0010] As a further limitation of the present invention, the step 2) specifically includes: selecting actual operation data of the coal mill unit, recording n auxiliary variable operation data and their corresponding actual air supply volumes at the same time; recording once at a certain time interval, and finally obtaining a data set X;

[0011] Xi =[x i1 ,x i2 ,…,x i(n+1) ];

[0012] Among them, X i is the i-th sample in the data set X; the first n elements are the n auxiliary variables in the i-th sample; x i(n+1) is the actual air supply volume corresponding to the auxiliary variable at the same moment;

[0013] The similarity function method is used to remove redundant data; the similarity between samples is judged by calculating the Euclidean distance between different samples, and then the selection of samples is determined;

[0014] The similarity function is: D i,j =exp(-‖x i -x j ‖2 / δ);

[0015] Among them, x i represents the i-th sample data; x j represents the jth sample data; D i,j is the similarity value to be obtained; δ is the normalization parameter; ‖·‖2 is the vector bi-norm, which is used to obtain the Euclidean distance between samples;

[0016] After similarity calculation, we get new data set X′, X i '=[x i ′1,x i ′2,…,x i ' (n+1) ]; where X i ' is the i-th sample in the data set X'; the first n elements are the n auxiliary variables in the i-th sample; x i ' (n+1) It is the actual air supply volume corresponding to the auxiliary variable at the same moment.

[0017] As a further limitation of the present invention, the new data set X′ obtained by similarity function calculation is subjected to kernel principal component analysis; Map X′ to a high-dimensional space, and then use the PCA method to calculate the principal component in the high-dimensional space;

[0018] Let vector ω i is the eigenvector in the high-dimensional space, λ i For its corresponding eigenvalue, the PCA in high-dimensional space is:

[0019]

[0020] The eigenvector ω i The sample set is represented as:

[0021]

[0022] Substituting formula (2) into formula (1), we get:

[0023] Multiply both sides of equation (3) by have to:

[0024]

[0025] Introduce the kernel function matrix K:

[0026]

[0027] Substituting equation (5) into equation (4) and performing kernel function matrix operation, we can get K 2 α=λ i Kα, substitute α into formula (2) to obtain the eigenvector ω i , using ω i As a set of bases of the high-dimensional space, this set of bases is a subspace of the high-dimensional space, and the test sample x is represented in this subspace to become a vector after dimensionality reduction;

[0028]

[0029] The variance contribution and cumulative variance contribution of each principal component are obtained through KPCA. The actual air supply volume corresponding to the principal elements with the highest cumulative variance contribution constitutes the data set X″. The data set X″ is normalized and the normalized data is used as the input data of the model.

[0030] As a further limitation of the present invention, the long short-term memory network LSTM forms a cell state through a forget gate, an input gate, an output gate and a tanh activation function in each unit to achieve the selection of historical information;

[0031] From the input data x t The output data h of the hidden layer at the previous moment t-1 They work together on the forget gate, which filters the information, memorizes the important feature information in the time series, and discards the irrelevant information;

[0032] The input data x t And the output data h of the hidden layer at the previous moment t-1 As the input information of the input gate, update;

[0033] The memory cell is fed with input data x t , the output data h of the hidden layer at the previous moment t-1 And the memory cell state C at the previous moment t-1Update its own status;

[0034] Recently, the input data x t、 The output data h of the hidden layer at the previous moment t-1 And the current memory cell state C t Acting together on the output gate, outputting the hidden layer information h at the current moment t ;

[0035] Determine the number of hidden layers of the prediction model based on the long short-term memory network; after determining the number of hidden layers, use the sparrow algorithm to optimize the parameters such as the number of iterations epoch, the number of hidden layer output nodes units, the number of memory days days and the learning rate lr; and select the maximum reference error between the predicted value and the true value as the fitness function for training;

[0036] The fitness function is: Where n is the number of predicted values, μ is the population size, and Y i is the predicted value, is the true value.

[0037] The present invention adopts the above technical solution, and has the following beneficial effects compared with the prior art: the present invention uses kernel principal component analysis to extract the features between data to the maximum extent and reduce the data dimension; LSTM is used to establish a primary air volume prediction model. In order to improve the prediction accuracy, the sparrow algorithm is combined to optimize the multiple hyperparameters existing in the model; the KPCA algorithm can effectively extract the data features of the original data set and improve the generalization ability of the model; the similarity function method effectively removes redundant samples and improves the prediction accuracy; the established prediction model has the advantages of higher accuracy and lower volatility. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 Flowchart of the present invention. DETAILED DESCRIPTION

[0039] like Figure 1 A method for predicting the primary air volume of a coal mill is shown, comprising the following steps:

[0040] 1) Determine the input variables of the prediction model, and select multiple variables as auxiliary variables by analyzing the mechanism of the primary air supply system of the coal mill;

[0041] Through the mechanism analysis of the primary air supply system of the coal mill, the flow formula It can be seen that the flow rate is directly related to the density ρ and differential pressure ΔP of the gas. The density of the gas is affected by temperature and pressure, so the temperature difference between the inlet and outlet and the pressure at the inlet of the primary fan are selected as auxiliary variables. The air volume is adjusted by the fan blade opening and the fan current, and the air volume is related to the coal supply. Too much or too little air volume will affect the combustion efficiency. The primary air is a mixture of cold air and hot air, so the cold air opening and hot air opening can be used as auxiliary variables.

[0042] When actually measuring the primary air volume in the duct, the total primary air volume is equal to the sum of the primary hot air volume and the primary cold air volume;

[0043] The primary hot air volume is:

[0044]

[0045] Among them, p i,1 is the air pressure at the hot air duct inlet, p h is the outlet pressure of the hot air duct; ρ h is the hot air density;

[0046] The primary cooling air volume is:

[0047]

[0048] Among them, p i,2 is the air pressure at the inlet of the cold air duct, p c is the outlet air pressure of the cold air duct; ρ c is the cold wind air density;

[0049] The total air volume of the primary air is: Q i =Q c +Q h ;

[0050] From equations (1) and (2), it can be seen that the actual air supply volume of the air supply system is proportional to the square of the differential pressure, so the differential pressure Δp1 between the primary air inlet and outlet of the air preheater and the differential pressure Δp2 between the primary fan outlet air pressure and the coal mill inlet air pressure are selected as auxiliary variables;

[0051] The actual density of primary air is related to the temperature difference. The temperature difference between the inlet and outlet of primary air in the air preheater ΔT1 and the temperature difference between the outlet air temperature of the primary fan and the inlet air temperature of the coal mill ΔT2 are selected as auxiliary variables.

[0052] When the primary air flows in the pipeline, the opening of the regulating valve will produce corresponding resistance. The opening of the coal mill hot air regulating valve and the opening of the coal mill cold air regulating valve are selected as auxiliary variables.

[0053] The dominant variable of primary air flow is the primary fan current, and the current size determines the size of the total air volume; the primary fan current is selected as the auxiliary variable;

[0054] The primary air volume is adjusted by the fan blade opening and the fan current. The air volume is related to the coal feeding amount. Too much or too little air volume will affect the combustion efficiency. In the pulverizing system, belt weighing is generally used to calculate the coal volume, so the coal mill current and the instantaneous coal feeding amount of the coal mill are used as auxiliary variables.

[0055] Therefore, nine variables are selected as auxiliary variables, which are the differential pressure between the primary air inlet and outlet of the air preheater Δp1, the differential pressure between the primary fan outlet air pressure and the pulverizer inlet air pressure Δp2, the opening of the pulverizer hot air regulating valve, the opening of the pulverizer cold air regulating valve, the temperature difference between the primary air inlet and outlet of the air preheater ΔT1, the temperature difference between the primary fan outlet air temperature and the pulverizer inlet air temperature ΔT2, the primary fan current, the pulverizer current and the instantaneous coal feed of the pulverizer.

[0056] 2) Use the normalization method to unify the basic measurement units of the training sample data, use the similarity function method to remove data samples with high similarity, and perform redundant processing on the data samples;

[0057] Select the actual operation data of the coal mill unit, record the operation data of 9 auxiliary variables and their corresponding actual air supply volume at the same time; record once every 5 minutes, and finally get the data set X;

[0058] X i =[x i1 ,x i2 ,…,x i10 ];

[0059] Among them, X i is the i-th sample in the data set X; the first 9 elements are the 9 auxiliary variables in the i-th sample; x i10 is the actual air supply volume corresponding to the auxiliary variable at the same moment;

[0060] The similarity function method is used to remove redundant data; the similarity between samples is judged by calculating the Euclidean distance between different samples, and then the selection of samples is determined;

[0061] The similarity function is: D i,j =exp(-‖x i -x j ‖2 / δ);

[0062] Among them, x i represents the i-th sample data; x j represents the jth sample data; D i,j is the similarity value to be obtained; δ is the normalization parameter; ‖·‖2 is the vector bi-norm, which is used to obtain the Euclidean distance between samples;

[0063] After similarity calculation, we get new data set X′, X i '=[x i ′1,x i ′2,…,x i ' 10 ]; where X i ' is the i-th sample in the data set X'; the first 9 elements are the 9 auxiliary variables in the i-th sample; x i ' 10 It is the actual air supply volume corresponding to the auxiliary variable at the same moment.

[0064] 3) Use kernel principal component analysis to extract features between data and reduce data dimensions;

[0065] The new data set X′ obtained by similarity function calculation is subjected to kernel principal component analysis; Map X′ to a high-dimensional space, and then use the PCA method to calculate the principal component in the high-dimensional space;

[0066] Let vector ω i is the eigenvector in the high-dimensional space, λ i For its corresponding eigenvalue, the PCA in high-dimensional space is:

[0067]

[0068] The eigenvector ω i The sample set is represented as:

[0069]

[0070] Substituting formula (4) into formula (3), we get:

[0071] Multiply both sides of equation (5) by have to:

[0072]

[0073] Introduce the kernel function matrix K:

[0074]

[0075] Substituting equation (7) into equation (6) and performing kernel function matrix operation, we can get K 2 α=λ i Kα, substitute α into formula (4) to obtain the eigenvector v i , using v i As a set of bases of the high-dimensional space, this set of bases is a subspace of the high-dimensional space, and the test sample x is represented in this subspace to become a vector after dimensionality reduction;

[0076]

[0077] The variance contribution and cumulative variance contribution of each principal component are obtained through KPCA. The actual air supply volume corresponding to the principal elements with the highest cumulative variance contribution (the first seven principal elements) constitutes the data set X″. The data set X″ is normalized and the normalized data is used as the input data of the model.

[0078] 4) Establish a primary air volume prediction model based on long short-term memory network;

[0079] The long short-term memory network LSTM forms the cell state through the forget gate, input gate, output gate and tanh activation function in each unit to achieve the selection of historical information;

[0080] From the input data x t The output data h of the hidden layer at the previous moment t-1 They work together on the forget gate, which filters the information, memorizes the important feature information in the time series, and discards irrelevant information;

[0081] The input data x t And the output data h of the hidden layer at the previous moment t-1 As the input information of the input gate, update;

[0082] The memory cell is fed with input data x t , the output data h of the hidden layer at the previous moment t-1 And the memory cell state C at the previous moment t-1 Update its own status;

[0083] Recently, the input data x t、 The output data h of the hidden layer at the previous moment t-1 And the current memory cell state C t Acting together on the output gate, outputting the hidden layer information h at the current moment t ;

[0084] Determine the number of hidden layers of the prediction model based on the long short-term memory network. In this embodiment, the number of hidden layers is 3 layers. After determining the number of hidden layers, use the sparrow algorithm to optimize the parameters such as the number of iterations epoch, the number of hidden layer output nodes units, the number of memory days and the learning rate lr.

[0085] The core of the sparrow algorithm is the position update of the discoverer, joiner, and vigilant. According to the algorithm settings, the proportion of vigilants in the population is 10% to 20%; the ratio between discoverers and joiners changes dynamically, that is, when one individual becomes a discoverer, it means that another individual becomes a joiner.

[0086] The discoverers in the sparrows have higher fitness values, which are used to find food for the entire sparrow population and provide foraging directions for all joiners. Their position update formula is as follows:

[0087]

[0088] where: t is the current iteration number; tmax is the maximum iteration number; X i,j is the position information of the i-th sparrow in the j-th dimension; a is a random number within (0, 1); q is a random number following a normal distribution; h is a matrix with all elements being 1; R2 is the early warning value; ST is the safety value; R2 < ST indicates that there are no predators around, and the discoverers can perform extensive search operations; R2 ≥ ST indicates that predators are found, and all sparrows have to fly to a safe place to forage.

[0089] For the joiners, their position update formula is as follows:

[0090]

[0091] where, X P represents the optimal position of the current discoverers; X w represents the current global worst position; A is a matrix.

[0092] The position update formula for the vigilant ones is as follows:

[0093]

[0094] where, X b is the current global optimal position; β is the step size control parameter; k is a random number within (-1, 1); f i is the fitness value of the current sparrow individual; f b is the current global best fitness value; f w is the current global worst fitness value, and ε is a constant. f i > f b , indicating that the sparrow is at the edge of the population at this time and is extremely vulnerable to attacks by predators. f i = f b , indicating that the sparrows in the middle of the population are aware of the danger and want to get closer to other sparrows to reduce the risk of being preyed upon.

[0095] Through cyclic iteration, continuously update the positions among the discoverers, joiners, and vigilant ones, and finally find the optimal parameters to complete the optimization task.

[0096] The population size is initialized to 100, the number of iterations is 200, the epoch interval is set to [1000, 5000], units is set to [8, 32], the learning rate lr is set to [0.001, 0.01], and days is set to [3, 8]; the parameters are substituted into formulas (8), (9), and (10) for position update.

[0097] The maximum reference error between the predicted value and the true value is selected as the fitness function for training; the fitness function is: Where n is the number of predicted values, μ is the population size, and Y i is the predicted value, is the true value.

[0098] In order to quantify the prediction effect of the method (KPCA-LSTM) of the present invention and evaluate the quality of the model, the maximum absolute error, root mean square error (RMSE) and average error (AE) were used to analyze the prediction accuracy and tracking change ability of the model; the specific results are shown in Table 1.

[0099] Maximum absolute error value: i=1,2,…,n, n is the number of test samples;

[0100] Root Mean Square Error (RMSE):

[0101]

[0102] Average Error (AE):

[0103]

[0104] Table 1 shows the prediction results of different models

[0105]

[0106] From the analysis of the results in Table 1, it can be seen that the prediction results of the prediction model established by KPCA-LSTM are the best in all evaluation indicators, proving that it is superior to other prediction models in terms of prediction accuracy and tracking ability. The method of the present invention extracts features between data through kernel principal component analysis to reduce data dimensions; establishes a primary air volume prediction model based on long short-term memory network to improve prediction accuracy.

[0107] The present invention is not limited to the above-mentioned embodiments. On the basis of the technical solution disclosed in the present invention, technicians in this field can make some substitutions and deformations to some technical features therein according to the disclosed technical content without creative labor, and these substitutions and deformations are all within the protection scope of the present invention.

Claims

1. A method for predicting primary air volume of a coal mill, characterized in that: The following steps are involved: 1) Determine the input variables of the prediction model, and select multiple variables as auxiliary variables by analyzing the mechanism of the primary air supply system of the coal mill; 2) Use the normalization method to unify the basic measurement units of the training sample data, use the similarity function method to remove data samples with high similarity, and perform redundant processing on the data samples; 3) Use kernel principal component analysis to extract features between data and reduce data dimensions; 4) Establish a primary air volume prediction model based on long short-term memory network.

2. A method for predicting primary air volume of a coal mill according to claim 1, characterized in that: The step 2) specifically includes: selecting the actual operation data of the coal mill unit, recording the operation data of n auxiliary variables and their corresponding actual air supply volumes at the same time; recording once at a certain time interval, and finally obtaining a data set X; X i =[x i1 ,x i2 ,…,x i(n+1) ]; Among them, X i is the i-th sample in the data set X; the first n elements are the n auxiliary variables in the i-th sample; x i(n+1) is the actual air supply volume corresponding to the auxiliary variable at the same moment; The similarity function method is used to remove redundant data; the similarity between samples is judged by calculating the Euclidean distance between different samples, and then the selection of samples is determined; The similarity function is: D i,j =exp(-‖x i -x j ‖2 / δ); Among them, x i represents the i-th sample data; x j represents the jth sample data; D i,j is the similarity value to be obtained; δ is the normalization parameter; ‖·‖2 is the vector bi-norm, which is used to obtain the Euclidean distance between samples; After similarity calculation, we get new data set X′, X i '=[x i ′1,x i ′2,…,x i ' (n+1) ]; where X i ' is the i-th sample in the data set X'; the first n elements are the n auxiliary variables in the i-th sample; x i ' (n+1) It is the actual air supply volume corresponding to the auxiliary variable at the same moment.

3. A method for predicting primary air volume of a coal mill according to claim 1, characterized in that: The step 3) specifically includes: performing a kernel principal component analysis on the new data set X′ obtained by calculating the similarity function; Map X′ to a high-dimensional space, and then use the PCA method to calculate the principal component in the high-dimensional space; Let vector ω i is the eigenvector in the high-dimensional space, λ i For its corresponding eigenvalue, the PCA in high-dimensional space is: The eigenvector ω i The sample set is represented as: Substituting formula (2) into formula (1), we get: Multiply both sides of equation (3) by have to: Introduce the kernel function matrix K: Substituting equation (5) into equation (4) and performing kernel function matrix operation, we can get K 2 α=λ i Kα, substitute α into formula (2) to obtain the eigenvector ω i , using ω i As a set of bases of the high-dimensional space, this set of bases is a subspace of the high-dimensional space, and the test sample x is represented in this subspace to become a vector after dimensionality reduction; The variance contribution and cumulative variance contribution of each principal component are obtained through KPCA. The actual air supply volume corresponding to the principal elements with the highest cumulative variance contribution constitutes the data set X″. The data set X″ is normalized and the normalized data is used as the input data of the model.

4. A method for predicting primary air volume of a coal mill according to claim 1, characterized in that: The step 4) specifically includes: the long short-term memory network LSTM forms a cell state through a forget gate, an input gate, an output gate and a tanh activation function in each unit to realize the selection of historical information; From the input data x t The output data h of the hidden layer at the previous moment t-1 They work together on the forget gate, which filters the information, memorizes the important feature information in the time series, and discards irrelevant information; The input data x t And the output data h of the hidden layer at the previous moment t-1 As the input information of the input gate, update; The memory cell is fed with input data x t , the output data h of the hidden layer at the previous moment t-1 And the memory cell state C at the previous moment t-1 Update its own status; Recently, the input data x t、 The output data h of the hidden layer at the previous moment t-1 And the current memory cell state C t Acting together on the output gate, outputting the hidden layer information h at the current moment t ; Determine the number of hidden layers of the prediction model based on the long short-term memory network; after determining the number of hidden layers, use the sparrow algorithm to optimize the parameters such as the number of iterations epoch, the number of hidden layer output nodes units, the number of memory days days and the learning rate lr; and select the maximum reference error between the predicted value and the true value as the fitness function for training; The fitness function is: Where n is the number of predicted values, μ is the population size, and Y i is the predicted value, is the true value.

Images