Method and device for predicting NOx emission concentration of boiler and medium

Abstract

The invention discloses a boiler NOx emission concentration prediction method and device and a medium. The method comprises the steps of obtaining boiler emission NOx concentration historical data and related auxiliary variables, conducting feature selection on multivariable data through LightGBM, conducting KGDA data enhancement on a data set after feature selection, and expanding the data set into a new concentration variable data set; multi-dimensional initialization of a population is realized by introducing an adaptive t distribution variation strategy, a search boundary is adjusted by adopting a dynamic boundary strategy, an energy valley optimizer algorithm is further optimized, and an IEVO algorithm is obtained through improvement; a PatchTST prediction model and an ETSform prediction model are established to predict the new data set, and an improved energy valley optimizer algorithm IEVO is used to optimize model hyper-parameters; and carrying out weighted fusion on prediction results of the two models by utilizing IEVO, and carrying out error correction on weighted results by adopting GRELM to obtain a final prediction result. According to the invention, accurate prediction of the NOx concentration of the boiler combustion outlet can be realized.

Description

Technical Field

[0001] The present invention belongs to the technical field of data processing, and in particular relates to a method, device and medium for predicting boiler NOx emission concentration. Background Art

[0002] As an important source of industrial pollution, accurate prediction of boiler emission concentration is crucial for environmental supervision and enterprise management. With the development of industry and the enhancement of environmental awareness, the task of predicting boiler emission concentration has become challenging and uncertain, and has become a complex and changeable problem. Boiler emission concentration is affected by multiple factors such as combustion process, fuel characteristics, operating conditions and meteorology, which complicates the prediction task.

[0003] The current boiler emission concentration prediction methods are mainly divided into physical models and machine learning models; physical models rely on simplified assumptions and are difficult to adapt to complex combustion and emission changes; although machine learning models can process historical data, they have limited accuracy and generalization capabilities when faced with nonlinear, multivariate emission data. Traditional emission concentration prediction models often rely on simplified physical assumptions and linear statistical methods, which are difficult to effectively capture the nonlinear characteristics and multi-scale variability of emission data, resulting in limited prediction accuracy and reliability under complex and changing environmental conditions. They are also sensitive to problems such as incomplete data, erroneous measurements, and loss, which affects the establishment of the model and the accuracy of the prediction results.

[0004] In response to the above challenges, the present invention proposes a boiler NOx emission concentration prediction method, which aims to improve the accuracy and effectiveness of boiler emission concentration prediction by combining advanced feature selection technology, data enhancement strategy, optimization algorithm and deep learning model. The present invention can not only process multivariate data, but also automatically adjust the model hyperparameters through adaptive optimization algorithm, effectively decompose and analyze the data, and thus provide more reliable boiler emission concentration prediction in complex environments. Summary of the invention

[0005] Purpose of the invention: In view of the problems existing in the prior art, the present invention provides a method, device and medium for predicting the NOx emission concentration of a boiler, which can achieve accurate prediction of the NOx emission concentration at the boiler combustion outlet.

[0006] Technical solution: A method for predicting boiler NOx emission concentration according to the present invention specifically comprises the following steps:

[0007] (1) Obtain historical NOx concentration data and auxiliary variable data of boiler emissions and normalize the data;

[0008] (2) Use LightGBM to perform feature selection on multivariate data and select the top n most important features based on feature importance; perform KGDA data enhancement on the dataset after feature selection and expand it into a new dataset;

[0009] (3) Using the improved energy valley optimizer algorithm IEVO to optimize the hyperparameters of the two models, and using the optimal parameters to predict the output components of the models; the IEVO algorithm is based on the EVO algorithm and adopts the adaptive t-distribution mutation strategy and the dynamic boundary strategy to improve the original population initialization process, and introduces the exponential moving average to improve the adaptive weight;

[0010] (4) Constructing the PatchTST model and the ETSformer model, inputting the data set obtained in step (2) into the two models for training, and optimizing the hyperparameters of the PatchTST and ETSformer models using the IEVO algorithm;

[0011] (5) The IEVO algorithm is used to perform weight fusion on the prediction results of the PatchTST and ETSformer models to obtain the optimal weight combination, and GRELM is used to perform error correction on the weighted results to obtain the final prediction results.

[0012] Furthermore, the auxiliary variable data in step (1) include medium characteristics, air supply volume, oxygen content, lower calorific value, unit load, air supply temperature and exhaust gas temperature.

[0013] Furthermore, the process of normalizing the data in step (1) is as follows:

[0014]

[0015] Among them, X' represents the normalized data value, X represents the original data value, and X max and X min Respectively represent the maximum and minimum values ​​of the column data.

[0016] Furthermore, the implementation process of step (2) is as follows:

[0017] (21) Use LightGBM to perform feature selection on multivariate data, sort them in descending order according to the absolute value of the gradient, select the first a% of the data as the large gradient subset A; randomly sample the remaining b*(1–a)% as the small gradient sample subset B, and add weight to B 1>b>0;

[0018] (22) The gradient of subset B is normalized and merged with A. According to the subset A∪B, the estimated variance gain is used To select the best splitting node for feature j, the specific formula is as follows:

[0019]

[0020] Among them, A l and A r They are the left and right branch sample sets after the large gradient sample subset A is split, and B l and B r are the left and right branch sample sets after the small gradient sample subset B is split, g i Represents sample x i The negative gradient of the loss function with respect to the model output, is the number of samples of the left branch with d as the splitting node according to feature j, is the number of samples of the right branch;

[0021] (23) Calculate the feature importance score Score(i) to evaluate the correlation between each feature and the class label. Score(i) is expressed as:

[0022] Score(i)={s|s=w i x i} (3)

[0023] Among them, w i represents the weight of each feature, x i Represents a feature set;

[0024] (24) The maximum mean difference loss (MMD) adversarial training is introduced to improve the performance of the generative adversarial network (GAN). The maximum mean difference generative adversarial network (MMD-GAN) is constructed, and the traditional GDA data enhancement technology is further improved to KGDA. The parameters of the generator G(z) are fixed, and the noise distribution P z (z) randomly select m samples Z = {Z 1 Z 2 ,…,z n}, and extract i data X={x 1 x 2 ……x i};

[0025] (25)MMD-GAN minimizes P r (x) and P g The MMD distance of (x') is used to generate samples, and the loss function of the discriminator is calculated according to formula (5). The Adam optimizer is used to update the network parameters of the discriminator. The loss functions of the generator and the discriminator are defined as:

[0026]

[0027] Where: k D is the kernel function, xi and x j are the i-th and j-th true samples in X, respectively, x' i and x' j are the i-th and j-th generated samples in X', respectively, and m and n are the number of samples in X and X' respectively;

[0028] (26) Repeat step (24), calculate the loss function of the generator according to the result of step (24), and use the Adam optimizer to update the network parameters of the generator; determine whether the number of iterations has been reached, and if not, continue training;

[0029] (27) The noise data that follows the prior distribution is taken as input, and the generator G(z) is used to create a dataset Q, which is then merged with the original dataset to form a new dataset M.

[0030] Furthermore, the implementation process of step (3) is as follows:

[0031] (31) Set the population size, dimension, number of iterations, and upper and lower limits of the search space of the EVO algorithm;

[0032] (32) An adaptive t-distribution strategy is introduced to initialize the population, and a dynamic boundary strategy is used to adjust the search range when a potential optimal solution area is found to balance global and local searches. The improved formula is as follows:

[0033]

[0034] Where: is the position after disturbance; t(iter) is the t distribution mutation strategy with the iteration number t distribution as the degree of freedom parameter;

[0035] (33) is randomly initialized, and the initial particle swarm can be expressed by a matrix:

[0036]

[0037] Among them, X i represents the i-th particle, g represents the initial particle swarm size, d represents the dimension of the variable to be optimized, represents the jth decision variable of the i-th particle;

[0038] (34) Determine the enrichment boundary of the particles, evaluate the objective function of the particles, and determine the particle enrichment level and stability. The specific calculation formula is as follows:

[0039]

[0040] Where NEL i is the neutron enrichment level of the i-th particle, EB is the enrichment boundary of the particle; SLi is the stability level of the \(i\)-th particle, where \(BI\) and \(WI\) are the particles with the best and worst stability levels in the particle, respectively;

[0041] (35) If \(NEL\) i > \(EB\), then a random number in the range \([0, 1]\) is generated as the stability bound \(SB\); if \(SL\) i > \(SB\), then the position is updated according to the \(\alpha\)-ray principle, generating two random integers as decision variables, and replaced by the particle or the candidate particle with the best stability level; two random integers are generated in the range \([1, d]\), representing the number of \(\alpha\)-rays emitted \(Alpha IndexⅠ\) and the specific number defining the ray \(Alpha IndexⅡ\), and the calculation formula is as follows:

[0042]

[0043] In the formula, is the newly generated particle in the universe, \(X\) i is the current position vector of the \(i\)-th particle in the search space, \(X\) BI is the position vector of the particle with the best stability;

[0044] (36) According to the \(r\)-ray principle, the particle position is randomly adjusted, and the photons in the particle are used as decision variables, and these variables will be replaced by neighboring particles or candidate solutions. The position of the second candidate solution is updated as:

[0045]

[0046] In the formula, is the total distance between the \(i\)-th particle and the \(k\)-th adjacent particle, \((x\) i , \(y\) i ) and \((x\) k , \(y\) k ) represent the coordinates of two particles in the search space, is the newly generated candidate solution in the search space, \(X\) Ng is the position vector of the neighboring particles around the \(i\)-th particle;

[0047] (37) If \(SL\) i < \(SB\), then the positions of the two particles are updated according to the \(\beta\)-decay principle, and the calculation formula is:

[0048]

[0049] In the formula, \(X\) CP is the position vector of the particle center, \(r\) 1 , \(r\) 2 , \(r\) 3 and \(r\) 4 are all random numbers, ranging from \([0, 1]\)

[0050] (38) If NEL i <EB, determine a random motion in the search space, and the expression is:

[0051]

[0052] In the formula, is the i-th candidate solution in the search space, r is a random number, and the range is between [0, 1];

[0053] (39) Adjust the weight by introducing the Exponential Moving Average (EMA), and search for the global optimal solution according to the position update formula; if the maximum number of iterations is not reached, return to step (3.3), and after reaching, output the optimal parameters and make a prediction. The calculation formula is as follows:

[0054] w t = α · w t-1 + (1 - α) · x t (17)

[0055] where, w t is the weight at time t, w t-1 is the weight at time t - 1; x t is the new data point at time t, and α is the exponential smoothing parameter, which controls the influence of historical data.

[0056] Furthermore, the implementation process of step (4) is as follows:

[0057] (41) Construct a PatchTST model. Given a multivariate time series dataset, input a sequence of length L (x 1 , x 2 … x n ) into the model. After the data processing layer receives the data, it divides the sequence into N slices. The calculation formula for N is:

[0058]

[0059] In the formula, L is the length of the input sequence, p is the patch length, and s is the step size;

[0060] (42) The Transformer layer maps the patch to a space with dimension D by multiplying with w p ∈ R D×P to obtain a feature projection and add it to the position encoding W Pox ∈ R D×N to get a sequence containing the time order of the patch That is:

[0061]

[0062] Where W Pox is the position encoding matrix, W p is the mapping matrix, is the i-th slice, is the feature projection result of the i-th slice;

[0063] (43) Each head h = 1, 2, ..., H in the multi-head attention is used to convert it into a Query vector Q, a Key vector K, and a Value vector V. The calculation formula is:

[0064]

[0065] In the formula, is the h-th Q-transform matrix, is the hth K transformation matrix, is the hth V transformation matrix;

[0066] (44) Perform Softmax normalization to obtain a weight matrix, where each value of the matrix is ​​greater than 0 and less than 1, and the sum is 1; multiply the weight matrix by V, and calculate the weighted sum to obtain the output result The calculation formula is as follows:

[0067]

[0068] Where, d k is the scaling factor;

[0069] (45) Construct the ETSformer model and transform the time series Q t-L:t Decomposed into seasonal and trend components, and the trend component is further divided into horizontal component and growth component; then converted to latent space through the Embedding module;

[0070] (46) The encoder is used to extract the growth and seasonal components from the time series of the concentration variable and pass them to the next layer after nonlinear transformation. The calculation formula is as follows:

[0071]

[0072] Among them, t is the current time, L is the length of the historical sequence, n is the number of layers; LN is layer normalization, and FF is the position feedforward network; is the residual of the nth encoder; is the residual of the previous encoder layer, NH-ESA is the head exponential smoothing attention mechanism, FN is the frequency attention mechanism, For the historical growth part, For the historical seasonal part;

[0073] (47) Define the horizontal part at the current moment

[0074]

[0075] Among them, β is a smoothing parameter, is the growth part of the previous moment, is the seasonal part of the current moment;

[0076] (48) The prediction of the K-step growth part and the seasonal part are obtained through the decoder. The specific calculation formula is as follows:

[0077]

[0078] in, is the growth part at the current moment, and They represent the forecast of the K-step growth part and the seasonal part respectively; E t:t+K represents the prediction of the K-step horizontal part; I represents the number of stacks, and LE is the growth damping;

[0079] (49) Combining the level, growth and seasonal forecasts, we can obtain the K-step forecast of boiler NOx gas emission concentration: The calculation formula is as follows:

[0080]

[0081] Furthermore, in step (5), the IEVO algorithm is used to perform weight fusion on the prediction results of the PatchTST and ETSformer models. The implementation process is as follows:

[0082] According to the prediction results obtained in step (4), an input matrix is ​​established; the objective function of weight optimization is constructed, and the formula is as follows:

[0083]

[0084] Among them, F is the weighted result of the components, K i is the component sequence, W i is the weight of each component;

[0085] The IEVO algorithm is used to adaptively optimize the weights of each component to obtain the final NOx emission concentration prediction result.

[0086] Furthermore, the implementation process of using GRELM to perform error correction on the weighted result in step (5) is as follows:

[0087] Subtract the actual NOx concentration data from the weighted prediction result to obtain the error time series. The specific formula is as follows:

[0088] Error=Svalue-Kvalue

[0089] In the formula, Error represents the error value of boiler NOx concentration, Svalue represents the measured value, and Kvalue represents the initial predicted value;

[0090] The obtained error time series and NOx auxiliary variables are reorganized into the model input matrix, and then the GRELM model is used for error correction prediction to obtain the boiler NOx emission concentration error prediction result GError;

[0091] The final prediction result is obtained by adding the initial prediction result and the error prediction result. The formula is as follows:

[0092] ZError=Kvalue+GError

[0093] Where ZError represents the final boiler NOx concentration prediction value.

[0094] The storage medium of the present invention stores a computer program, and when the computer program is executed by at least one processor, the steps of the boiler NOx emission concentration prediction method as described above are implemented.

[0095] An electronic device according to the present invention comprises a memory and a processor, wherein:

[0096] A memory for storing computer programs that can be run on the processor;

[0097] The processor is used to execute the steps of the boiler NOx emission concentration prediction method as described above when running the computer program.

[0098] Beneficial effects: Compared with the prior art, the beneficial effects of the present invention are as follows: 1. The present invention adopts LightGBM for feature selection, eliminates noise features, retains key information, reduces the risk of overfitting, and improves prediction accuracy; LightGBM is used to select the top n most important features according to feature importance, which helps the model to concentrate on learning the features that have the greatest impact on the prediction results; the present invention adopts the improved KGDA data enhancement technology to generate diversified training samples, so that the model can better adapt to different data distributions and improve the robustness of the model; 2. The present invention improves the energy valley optimizer algorithm EVO, improves population initialization through adaptive t distribution mutation strategy and dynamic boundary strategy, enhances global search capability, helps to jump out of local optimal solutions, balances global search and local search, and improves convergence speed and search efficiency; then, by introducing exponential moving average to improve adaptive weights, the algorithm can more flexibly adjust the weights between individuals, and finally obtains the improved energy valley optimizer algorithm IE VO; 3. Construct the PatchTST model. The PatchTST model uses the Transformer structure, which can process multivariate time series data, effectively capture long-distance dependencies in time series, segment the time series through the "Patch" mechanism, automatically learn the appropriate data block size, and enhance the learning of local features of time series data; construct the ETSformer model, which maps the time series to the latent space through the Embedding module, can adapt to different trends and seasonal components, and improve the prediction ability of long time series; use the improved IEVO algorithm to optimize the hyperparameters of the PatchTST and ETSformer models, and improve the generalization ability of the model; 4. Through the IEVO algorithm, the prediction results of the weighted fusion output components are used, and GRELM is used for error correction, which can output a more accurate NOx concentration of boiler combustion emissions. Through the calculation of observed values ​​and predicted values, the prediction error is greatly reduced. BRIEF DESCRIPTION OF THE DRAWINGS

[0099] Figure 1 This is a flow chart of the boiler NOx emission concentration prediction method proposed by the present invention;

[0100] Figure 2 This is a schematic diagram of the PatchTST model structure proposed in the present invention;

[0101] Figure 3 This is a schematic diagram of the ETSformer model structure proposed in the present invention. DETAILED DESCRIPTION

[0102] The present invention is further described in detail below based on the accompanying drawings.

[0103] like Figure 1As shown, the present invention provides a method for predicting boiler NOx emission concentration. First, historical NOx concentration data and auxiliary variable data of boiler emissions are obtained; LightGBM is used to perform feature selection on multivariate data, and data enhancement (K Generative Data Augmentation, KGDA) is performed on the data set after feature selection to expand it into a new data set; an adaptive t-distribution mutation strategy is adopted to realize multi-dimensional initialization of the population, and a dynamic boundary strategy is adopted to adjust the search boundary, thereby optimizing the Energy Valley Optimizer algorithm (Energy Valley Optimizer, EVO), and the IEVO algorithm is obtained through improvement; a PatchTST prediction model and an ETSformer prediction model are established to predict the new data set, and the improved Energy Valley Optimizer algorithm IEVO is used to optimize the model hyperparameters; finally, IEVO is used to weight the prediction results of the two models, and GRELM is used to perform error correction on the weighted results to obtain the final prediction results. Compared with the prior art, the present invention can achieve accurate prediction of the NOx concentration at the boiler combustion outlet. The specific steps are as follows:

[0104] Step 1: Obtain historical NOx concentration data and auxiliary variable data of boiler emissions, and preprocess the data set;

[0105] A NOx concentration variable data set is established, and its auxiliary variable data include medium characteristics, air supply volume, oxygen content, lower calorific value, unit load, air supply temperature and exhaust temperature.

[0106] Since the dimensions of the variables are inconsistent, in order to unify the dimensions and thus improve the prediction fit of the model, the data is normalized. The formula is as follows:

[0107]

[0108] Among them, X represents the original data value, X max and X min They represent the maximum and minimum values ​​of the column data respectively, and X' represents the normalized data value.

[0109] Step 2: Use LightGBM to perform feature selection on multivariate data, select the top N most important features in the initial state, and perform data enhancement through KGDA.

[0110] First, LightGBM is used to perform feature selection on multivariate data. The data is sorted in descending order according to the absolute value of the gradient, and the first a% of the data is selected as the large gradient subset A. The remaining b*(1–a)% is randomly sampled as the small gradient sample subset B, and B is weighted. 1>b>0.

[0111] Normalize the gradient of subset B and merge it with A. According to the subset A∪B, use the estimated variance gain To select the best splitting node for feature j, the specific formula is as follows:

[0112]

[0113] Among them, A l and A r They are the left and right branch sample sets after the large gradient sample subset A is split, and B l and B r are the left and right branch sample sets after the small gradient sample subset B is split, g i Represents sample x i The negative gradient of the loss function with respect to the model output, is the number of samples of the left branch with d as the splitting node according to feature j, is the number of samples in the right branch.

[0114] Calculate the feature importance score Score(i) to evaluate the correlation between each feature and the class label. Score(i) is expressed as:

[0115] Score(i)={s|s=w i x i}, (3)

[0116] Among them, w i represents the weight of each feature, x i Represents a feature set.

[0117] The maximum mean discrepancy (MMD) loss adversarial training is introduced to improve the performance of the generative adversarial network (GAN), and the maximum mean discrepancy generative adversarial network (MMD-GAN) is constructed, and the traditional GDA data enhancement technology is further improved to KGDA. The parameters of the generator G(z) are fixed, and the noise distribution P z (z) randomly select m samples Z = {z 1 z 2 ,…,z n}, and extract i data X={x 1 x 2 ……x i}.

[0118] MMD-GAN minimizes P r (x) and P gThe MMD distance of (x') is used to generate samples, and the loss function of the discriminator is calculated according to formula (5). The Adam optimizer is used to update the network parameters of the discriminator. The loss functions of the generator and the discriminator are defined as:

[0119]

[0120] Where: k D is the kernel function, x i and x j are the i-th and j-th true samples in X, respectively, x' i and x' j are the i-th and j-th generated samples in X', respectively. m and n are the number of samples in X and X', respectively.

[0121] Calculate the loss function of the generator and use the Adam optimizer to update the network parameters of the generator. Determine whether the number of iterations has been reached. If not, continue training. Take the noise data that follows the prior distribution as input, use the generator G(z) to create a dataset Q, and merge it with the original dataset to form a new dataset M.

[0122] Step 3: Improve the Energy Valley Optimizer (EVO) algorithm to obtain the IEVO algorithm.

[0123] 3.1: First, set the population size, dimension, number of iterations, and upper and lower limits of the search space of the EVO algorithm.

[0124] 3.2: The algorithm introduces an adaptive t-distribution strategy to initialize the population to adapt to the search requirements at different stages, mutate and perturb the optimal point update, and enhance the global search capability. The dynamic boundary strategy is used to adjust the search range when the potential optimal solution area is found, balancing the global and local search. The improved formula is as follows:

[0125]

[0126] Where: is the position after disturbance; t(iter) is the t distribution mutation strategy with the iteration number t distribution as the degree of freedom parameter.

[0127] 3.3: Perform random initialization. The initial particle swarm can be expressed by a matrix:

[0128]

[0129] Among them, X i represents the i-th particle, g represents the initial particle swarm size, d represents the dimension of the variable to be optimized, represents the jth decision variable of the ith particle.

[0130] 3.4: Determine the particle enrichment boundary, evaluate the particle objective function, and determine the particle enrichment level and stability. The specific calculation formula is as follows:

[0131]

[0132] Where NEL i is the neutron enrichment level of the i-th particle, EB is the enrichment boundary of the particle; SL i is the stability level of the ith particle, BI and WI are the particles with the best and worst stability levels, respectively.

[0133] 3.5: If NEL i >EB, then generate a random number in the range [0,1] as the stability bound SB. i >SB, the position is updated according to the α-ray principle, and two random integers are generated as decision variables, which are replaced by particles or candidate particles with the best stability level. Two random integers are generated in the range [1, d], representing the number of emitted α-rays Alpha IndexⅠ and the specific number of the defined ray (Alpha IndexⅡ), respectively. The calculation formula is as follows:

[0134]

[0135] In the formula, is a newly generated particle in the universe, X i is the current position vector of the i-th particle in the search space, X BI is the position vector of the particle with the best stability.

[0136] 3.6: According to the r-ray principle, the particle position is randomly adjusted, and the photons in the particles are used as decision variables. These variables will be replaced by neighboring particles or candidate solutions, and the position of the second candidate solution is updated as:

[0137]

[0138] In the formula, is the total distance between the ith particle and the kth adjacent particle, (x i ,y i ) and (x k ,y k ) represents the coordinates of the two particles in the search space, is the candidate solution newly generated in the search space, X Ng is the position vector of the neighboring particles around the ith particle.

[0139] 3.7: If SL i<SB, then the positions of the two particles are updated according to the principle of beta decay. During this process, the particles will move towards the particle with the best stability in a controlled manner. The calculation formula is as follows:

[0140]

[0141] In the formula, X CP is the position vector of the particle center, r 1 , r 2 , r 3 and r 4 are all random numbers, and the range is between [0, 1].

[0142] 3.8: If NEL i <EB, determine a random movement in the search space. The expression is:

[0143]

[0144] In the formula, is the i-th candidate solution in the search space, and r is a random number with a range between [0, 1].

[0145] 3.9: Adjust the weights by introducing the Exponential Moving Average (EMA), and search for the global optimal solution according to the position update formula. If the maximum number of iterations is not reached, return to step (3.3). After reaching it, output the optimal parameters and make predictions. The calculation formula is as follows:

[0146] w t = α·w t-1 + (1 - α)·x t (17)

[0147] Among them, w t is the weight at time t, and w t-1 is the weight at time t - 1. x t is the new data point at time t, and α is the exponential smoothing parameter (usually between 0 and 1), which controls the influence of historical data.

[0148] Step 4: Build a prediction model based on PatchTST and ETSformer, predict the new dataset of concentration variables, and use the IEVO algorithm to optimize the model hyperparameters to improve the accuracy of emission gas concentration prediction.

[0149] 4.1: Construct a PatchTST model as Figure 2 shown. Given a multivariate time series dataset, for a sequence of length L (x 1 , x 2 … x n) is input into the model. After receiving the data, the data processing layer divides the sequence into N patches. The calculation formula of N is:

[0150]

[0151] Where L is the input sequence length, p is the patch length, and s is the step size.

[0152] 4.2: Transformer layer passes the patch through w p ∈R D×P Multiply and map to a space of dimension D to obtain feature projection and position encoding W Pox ∈R D×N Add together to get a sequence containing the time order of the patches Right now

[0153]

[0154] Where W Pox is the position encoding matrix, W p is the mapping matrix, is the i-th slice, is the feature projection result of the i-th slice.

[0155] 4.3: Use each head h=1,2,…,H in the multi-head attention to convert it into a Query vector Q, a Key vector K and a Value vector V. The calculation formula is:

[0156]

[0157] In the formula, is the h-th Q-transform matrix, is the hth K transformation matrix, is the hth V transformation matrix.

[0158] 4.4: Perform Softmax normalization to obtain a weight matrix, where each value of the matrix is ​​greater than 0 and less than 1, and the sum is 1. Multiply the weight matrix by V and calculate the weighted sum to obtain the output result The calculation formula is as follows:

[0159]

[0160] Where, d k is the scaling factor.

[0161] 4.5: Build as Figure 3 The ETSformer model shown in the figure transforms the time series Q t-L:tIt is decomposed into seasonal and trend components, and the trend component is further divided into horizontal component and growth component; then it is transformed into latent space through the Embedding module.

[0162] 4.6: Use the encoder to extract the growth and seasonal parts from the time series of the concentration variable, and pass them to the next layer after nonlinear transformation. The calculation formula is as follows:

[0163]

[0164] Among them, t is the current time, L is the length of the historical sequence, n is the number of layers; LN is layer normalization, and FF is the position feedforward network; is the residual of the nth encoder; is the residual of the previous encoder layer, NH-ESA is the head exponential smoothing attention mechanism, FN is the frequency attention mechanism, For the historical growth part, For the historical seasonal part.

[0165] 4.7: Define the horizontal part of the current moment

[0166]

[0167] Among them, β is a smoothing parameter, is the growth part of the previous moment, is the seasonal component of the current moment.

[0168] 4.8: The prediction of the K-step growth part and the seasonal part are obtained through the decoder. The specific calculation formula is as follows:

[0169]

[0170] in, is the growth part at the current moment, and They represent the forecast of the K-step growth part and the seasonal part respectively; E t:t+K represents the prediction of the K-step horizontal section; I represents the number of stacks and LE is the growth damping.

[0171] 4.9: Combine level, growth and seasonal forecasts to obtain K-step forecasts of boiler NOx gas emission concentrations The calculation formula is as follows:

[0172]

[0173] 4.10: The input of boiler NOx concentration prediction model is historical NOx data and auxiliary variables (media characteristics, air supply volume, etc.), and the output is the NOx concentration prediction value. The input of PatchTST and ETSformer models is model parameters (number of neurons, learning rate, etc.), and the output is RMSE evaluation index.

[0174] 4.11: Initialize the initial population of the EVO algorithm through the adaptive t-distribution mutation strategy and the dynamic boundary strategy, and set the maximum number of iterations, population size, and upper and lower limits of the individual positions of the population.

[0175] 4.12: Update the individual fitness according to the objective function value, introduce the exponential moving average (EMA) to adjust the weight, and search for the global optimal solution according to the position update formula.

[0176] 4.13: If the current number of iterations is greater than the maximum number of iterations, decode the individual positions of the energy valleys and output the number of neurons, learning rate, regularization parameter, and number of hidden layers of the PatchTST and ETSformer models. Otherwise, return to step 4.11.

[0177] Step 5: Use the IEVO algorithm to perform weight fusion on the prediction results of the two models to obtain the optimal weight combination, and use GRELM to perform error correction on the weighted results to obtain the final prediction result.

[0178] According to the prediction results obtained in step 4, the input matrix is ​​established; the objective function of weight optimization is constructed, and the formula is as follows:

[0179]

[0180] Among them, R is the weighted result of the components, C i is the component sequence, W i is the weight of each component; the IEVO algorithm is used to adaptively optimize the weight of each component to obtain the final emission concentration prediction result.

[0181] Train the correction module GRELM:

[0182] First, the weight β from the hidden layer to the output layer is initialized; the data set passes through the hidden layer and enters the output layer. The output of GRELM is:

[0183]

[0184] Where β = [β 1 ,…,β L ] T It is the output weight between the hidden layer (L nodes) and the output layer (m nodes, m ≥ 1).

[0185] Calculate the loss under the current weight. The specific formula is as follows:

[0186] min||Hβ-T|| 2 ,β∈R L×m (34)

[0187] Use the gradient descent method to update the weights and reduce the prediction error. The specific formula is as follows:

[0188]

[0189] Among them, h j (x i ) is the output of the jth node in the hidden layer for the ith sample. In order to ensure that the error is minimized, the objective function is evaluated by minimizing the approximate square error so that the solution obtained is the optimal solution. The objective function is as follows:

[0190]

[0191] Among them, H is the output matrix of the hidden layer, and T is the target matrix of the training data. GRELM is used to perform error correction on the weighted results to obtain the final prediction results;

[0192] The historical NOx concentration and auxiliary variable data of boiler emissions are collected, and the Light-GBM feature selection is combined with KGDA data enhancement to expand the new data set; the optimized PatchTST and ETSformer models are used to make preliminary predictions on the expanded data, and the prediction results of the two models are weighted fused; then the actual NOx concentration data is subtracted from the weighted prediction results to obtain the error time series. The expression of this process is as follows:

[0193] Error=Svalue-Kvalue

[0194] Where Error represents the error value of boiler NOx concentration, Svalue represents the measured value, and Kvalue represents the initial predicted value.

[0195] The obtained error time series and NOx auxiliary variables are reorganized into the model input matrix, and then the GRELM model is used for error correction prediction to obtain the boiler NOx emission concentration error prediction result GError.

[0196] Finally, the initial prediction result and the error prediction result are added together to get the final prediction result, which is as follows:

[0197] ZError=Kvalue+GError

[0198] Where ZError represents the final boiler NOx concentration prediction value.

[0199] The present invention further provides a storage medium having a computer program stored thereon. When the computer program is executed by at least one processor, the steps of the boiler NOx emission concentration prediction method as described above are implemented.

[0200] The present invention also proposes an electronic device, comprising a memory and a processor, wherein: the memory is used to store a computer program that can be run on the processor; the processor is used to execute the steps of a boiler NOx emission concentration prediction method as described above when running the computer program.

[0201] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A method for predicting boiler NOx emission concentration, characterized in that: The following steps are involved: (1) Obtain historical NOx concentration data and auxiliary variable data of boiler emissions and normalize the data; (2) Use LightGBM to perform feature selection on multivariate data and select the top n most important features based on feature importance; Perform KGDA data enhancement on the dataset after feature selection and expand it into a new dataset; (3) Using the improved energy valley optimizer algorithm IEVO to optimize the hyperparameters of the two models, and using the optimal parameters to predict the output components of the models; the IEVO algorithm is based on the EVO algorithm and adopts the adaptive t-distribution mutation strategy and the dynamic boundary strategy to improve the original population initialization process, and introduces the exponential moving average to improve the adaptive weight; (4) Constructing the PatchTST model and the ETSformer model, inputting the data set obtained in step (2) into the two models for training, and optimizing the hyperparameters of the PatchTST and ETSformer models using the IEVO algorithm; (5) The IEVO algorithm is used to perform weight fusion on the prediction results of the PatchTST and ETSformer models to obtain the optimal weight combination, and GRELM is used to perform error correction on the weighted results to obtain the final prediction results.

2. A boiler NOx emission concentration prediction method according to claim 1, characterized in that: The auxiliary variable data in step (1) include medium characteristics, air supply volume, oxygen content, lower calorific value, unit load, air supply temperature and exhaust gas temperature.

3. A boiler NOx emission concentration prediction method according to claim 1, characterized in that: The implementation process of normalizing the data in step (1) is as follows: Among them, X' represents the normalized data value, X represents the original data value, and X max and X min Respectively represent the maximum and minimum values ​​of the column data.

4. A boiler NOx emission concentration prediction method according to claim 1, characterized in that: The implementation process of step (2) is as follows: (21) Use LightGBM to perform feature selection on multivariate data, sort them in descending order according to the absolute value of the gradient, select the first a% of the data as the large gradient subset A; randomly sample the remaining b*(1–a)% as the small gradient sample subset B, and add weight to B (22) The gradient of subset B is normalized and merged with A. According to the subset A∪B, the estimated variance gain is used To select the best splitting node for feature j, the specific formula is as follows: Among them, A l and A r They are the left and right branch sample sets after the large gradient sample subset A is split, and B l and B r are the left and right branch sample sets after the small gradient sample subset B is split, g i Represents sample x i The negative gradient of the loss function with respect to the model output, is the number of samples of the left branch with d as the splitting node according to feature j, is the number of samples of the right branch; (23) Calculate the feature importance score Score(i) to evaluate the correlation between each feature and the class label. Score(i) is expressed as: Score(i)={s∣s=w i x i } (3) Among them, w i represents the weight of each feature, x i Represents a feature set; (24) The maximum mean difference loss (MMD) adversarial training is introduced to improve the performance of the generative adversarial network (GAN). The maximum mean difference generative adversarial network (MMD-GAN) is constructed, and the traditional GDA data enhancement technology is further improved to KGDA. The parameters of the generator G(z) are fixed, and the noise distribution P z (z) randomly select m samples Z = {z1 z2,…,z n }, and extract i data X={x1x2……x i }; (25)MMD-GAN minimizes P r (x) and P g The MMD distance of (x') is used to generate samples, and the loss function of the discriminator is calculated according to formula (5). The Adam optimizer is used to update the network parameters of the discriminator. The loss functions of the generator and the discriminator are defined as: Where: k D is the kernel function, x i and x j are the i-th and j-th true samples in X, respectively, x' i and x' j are the i-th and j-th generated samples in X', respectively, and m and n are the number of samples in X and X' respectively; (26) Repeat step (24), calculate the loss function of the generator according to the result of step (24), and use the Adam optimizer to update the network parameters of the generator; determine whether the number of iterations has been reached, and if not, continue training; (27) The noise data that follows the prior distribution is taken as input, and the generator G(z) is used to create a dataset Q, which is then merged with the original dataset to form a new dataset M.

5. A boiler NOx emission concentration prediction method according to claim 1, characterized in that: The implementation process of step (3) is as follows: (31) Set the population size, dimension, number of iterations, and upper and lower limits of the search space of the EVO algorithm; (32) An adaptive t-distribution strategy is introduced to initialize the population, and a dynamic boundary strategy is used to adjust the search range when a potential optimal solution area is found to balance global and local searches. The improved formula is as follows: Where: is the position after disturbance; t(iter) is the t distribution mutation strategy with the iteration number t distribution as the degree of freedom parameter; (33) is randomly initialized, and the initial particle swarm can be expressed by a matrix: Among them, X i represents the i-th particle, g represents the initial particle swarm size, d represents the dimension of the variable to be optimized, represents the jth decision variable of the i-th particle; (34) Determine the enrichment boundary of the particles, evaluate the objective function of the particles, and determine the particle enrichment level and stability. The specific calculation formula is as follows: Where NEL i is the neutron enrichment level of the i-th particle, EB is the enrichment boundary of the particle; SL i is the stability level of the ith particle, BI and WI are the particles with the best and worst stability levels, respectively; (35) If NEL i >EB, then generate a random number in the range [0,1] as the stability bound SB; if SL i >SB, the position is updated according to the α-ray principle, two random integers are generated as decision variables, and they are replaced by particles or candidate particles with the best stability level; two random integers are generated in the range [1, d], representing the number of emitted α-rays Alpha IndexⅠ and the specific number of the defined ray Alpha IndexⅡ, respectively. The calculation formula is as follows: In the formula, is a newly generated particle in the universe, X i is the current position vector of the i-th particle in the search space, X BI is the position vector of the particle with the best stability; (36) According to the r-ray principle, the particle position is randomly adjusted, and the photons in the particles are used as decision variables. These variables will be replaced by neighboring particles or candidate solutions, and the position of the second candidate solution is updated as: In the formula, is the total distance between the ith particle and the kth adjacent particle, (x i ,y i ) and (x k ,y k ) represents the coordinates of the two particles in the search space, is the candidate solution newly generated in the search space, X Ng is the position vector of the neighboring particles around the i-th particle; (37) If SL i <SB, the positions of the two particles are updated according to the principle of beta decay, and the calculation formula is: Where, X CP is the position vector of the particle center, r1, r2, r3 and r4 are all random numbers, ranging between [0,1] (38) If NEL i <EB, determine a random motion in the search space, and the expression is: In the formula, is the i-th candidate solution in the search space, r is a random number ranging from [0,1]; (39) By introducing the exponential moving average (EMA) to adjust the weight, the global optimal solution is searched according to the position update formula; if the maximum number of iterations is not reached, return to step (3.3). Once it is reached, the optimal parameters are output and predicted. The calculation formula is as follows: w t =α·w t-1 +(1-a)·x t (17) Among them, w t is the weight at time t, w t-1 is the weight at time t-1; x t is the new data point at time t, and α is the exponential smoothing parameter that controls the influence of historical data.

6. A boiler NOx emission concentration prediction method according to claim 1, characterized in that: The implementation process of step (4) is as follows: (41) Construct the PatchTST model. Given a multivariate time series dataset, a sequence of length L (x1, x2....x n ) Input model, the data processing layer divides the sequence into N slices after receiving the data. The calculation formula of N is: Where L is the input sequence length, p is the patch length, and s is the step size; (42) The Transformer layer passes the patch through w p ∈R D×P Multiply and map to a space of dimension D to obtain feature projection and position encoding W Pox ∈R D×N Add together to get a sequence containing the time order of the patches Right now: Where W Pox is the position encoding matrix, W p is the mapping matrix, is the i-th slice, is the feature projection result of the i-th slice; (43) Each head h = 1, 2, ..., H in the multi-head attention is used to convert it into a Query vector Q, a Key vector K, and a Value vector V. The calculation formula is: In the formula, is the h-th Q-transform matrix, is the hth K transformation matrix, is the hth V transformation matrix; (44) Perform Softmax normalization to obtain a weight matrix, in which each value of the matrix is ​​greater than 0 and less than 1, and the sum is 1; Multiply the weight matrix by V and calculate the weighted sum to get the output result The calculation formula is as follows: Where, d k is the scaling factor; (45) Construct the ETSformer model and transform the time series Q t-L:t Decomposed into seasonal and trend components, and the trend component is further divided into horizontal component and growth component; then converted to latent space through the Embedding module; (46) The encoder is used to extract the growth and seasonal components from the time series of the concentration variable and pass them to the next layer after nonlinear transformation. The calculation formula is as follows: Among them, t is the current time, L is the length of the historical sequence, n is the number of layers; LN is layer normalization, and FF is the position feedforward network; is the residual of the nth encoder; is the residual of the previous encoder layer, NH-ESA is the head exponential smoothing attention mechanism, FN is the frequency attention mechanism, For the historical growth part, For the historical seasonal part; (47) Define the horizontal part at the current moment Among them, β is a smoothing parameter, is the growth part of the previous moment, is the seasonal part of the current moment; (48) The prediction of the K-step growth part and the seasonal part are obtained through the decoder. The specific calculation formula is as follows: in, is the growth part at the current moment, and They represent the forecast of the K-step growth part and the seasonal part respectively; E t:t+K represents the prediction of the K-step horizontal part; I represents the number of stacks, and LE is the growth damping; (49) Combining the level, growth and seasonal forecasts, we can obtain the K-step forecast of boiler NOx gas emission concentration: The calculation formula is as follows:

7. A boiler NOx emission concentration prediction method according to claim 1, characterized in that: In step (5), the IEVO algorithm is used to perform weight fusion on the prediction results of the PatchTST and ETSformer models. The implementation process is as follows: According to the prediction results obtained in step (4), an input matrix is ​​established; the objective function of weight optimization is constructed, and the formula is as follows: Among them, F is the weighted result of the components, K i is the component sequence, W i is the weight of each component; The IEVO algorithm is used to adaptively optimize the weights of each component to obtain the final NOx emission concentration prediction result.

8. A boiler NOx emission concentration prediction method according to claim 1, characterized in that: The implementation process of using GRELM to perform error correction on the weighted result in step (5) is as follows: Subtract the actual NOx concentration data from the weighted prediction result to obtain the error time series. The specific formula is as follows: Error=Svalue-Kvalue In the formula, Error represents the error value of boiler NOx concentration, Svalue represents the measured value, and Kvalue represents the initial predicted value; The obtained error time series and NOx auxiliary variables are reorganized into the model input matrix, and then the GRELM model is used for error correction prediction to obtain the boiler NOx emission concentration error prediction result GError; The final prediction result is obtained by adding the initial prediction result and the error prediction result. The formula is as follows: ZError=Kvalue+GError Where ZError represents the final boiler NOx concentration prediction value.

9. A storage medium, characterized in that: The storage medium stores a computer program, and when the computer program is executed by at least one processor, the steps of a method for predicting boiler NOx emission concentration as claimed in any one of claims 1 to 8 are implemented.

10. An electronic device, characterized in that: comprising a memory and a processor, wherein: A memory for storing computer programs that can be run on the processor; A processor is used to execute the steps of a method for predicting boiler NOx emission concentration as described in any one of claims 1 to 8 when running the computer program.

Images