Boiler NOx emission estimation method based on width learning system and integrated learning

Abstract

The invention discloses a boiler NOx emission estimation method based on a width learning system and integrated learning. The boiler NOx emission estimation method comprises the steps of 1, data processing; step 2, constructing an integrated learning prediction model; step 3, performing real-time estimation; according to the invention, the hybrid quantum dune cat swarm algorithm QSCSO is provided, a Levy flight strategy and a quantum strategy are added on the basis of the SCSO for improvement, the problems of low convergence precision, low optimization efficiency and the like caused by limitation of global search and easy falling into local optimum of the SCSO are solved, and a modeling method based on QSCSO-BLS is designed for individual learning device construction. According to the method, the training precision and generalization ability of the BLS model during modeling of the single-working-condition system are improved, meanwhile, the boiler NOx emission estimation method based on QSCSO-BLS and integrated learning is provided, and the problem that the model generalization ability is poor during modeling of the BLS full-working-condition system is solved.

Description

Technical Field

[0001] The present invention relates to the technical field of boiler emission prediction, and in particular to a boiler NOx emission prediction method based on a width learning system and integrated learning. Background Art

[0002] Boiler NOx (nitrogen oxide) emission estimation is an important part of the environmental protection field. As one of the atmospheric pollutants, NOx has caused serious impacts on the environment and human health, including the formation of environmental problems such as photochemical smog and acid rain, as well as harm to the human respiratory system. Therefore, reducing boiler NOx emissions has become a focus of attention. In order to effectively control boiler NOx emissions, companies need to accurately estimate their emissions in order to formulate scientific emission reduction measures and management strategies. Boiler NOx emission estimation not only helps companies understand their own emission status, but also provides a regulatory basis for environmental protection departments, and promotes the entire industry to develop in a more environmentally friendly and sustainable direction. Methods include statistical modeling based on historical data, using big data and machine learning algorithms for prediction; online monitoring and analysis of real-time operation data, combined with advanced algorithms for instant prediction; and combustion process simulation technology, which predicts NOx generation by simulating combustion conditions.

[0003] The existing boiler NOx emission estimation methods have the following defects: First, there is a BLS hyperparameter problem when using BLS modeling. The weights and biases between the BLS input data and feature nodes, and between feature nodes and enhanced nodes are randomly generated, and the quality of the network performance is random; and when BLS is used to model the full-condition system, the generalization ability of the model is poor, and it is difficult to cover the dynamic and static characteristics of the system; Second, SCSO has limited search range, and the optimization results are often not the global optimal, and it is easy to fall into local optimality in the later iteration; Third, at present, the NOx concentration measurement of the SCR denitrification system is mainly carried out by the continuous emission monitoring system (CEMS). Since CEMS is sampled and detected in the flue with more dust, there will be real-time problems when measuring NOx concentration. The time for flue transportation, gas sampling and detection is long, resulting in a lag in NOx concentration measurement; and the accuracy is poor. CEMS does not collect flue gas during the regular purge of the sampling probe. At this time, the NOx concentration measurement value fluctuates abnormally and deviates from the actual value. Summary of the invention

[0004] The purpose of the present invention is to provide a boiler NOx emission prediction method based on a width learning system and integrated learning to solve the problems raised in the above background technology.

[0005] To achieve the above object, the present invention provides the following technical solution: a boiler NOx emission estimation method based on a width learning system and integrated learning, comprising step one, data processing; step two, integrated learning estimation model construction; step three, real-time estimation;

[0006] In the above step 1, the full-operation operation data of the SCR denitrification system in the thermal power unit historical station is screened and preprocessed, the model input and output are determined, and the data set is divided;

[0007] In the above step 2, the training set is divided into working conditions, combiner training and model integration are performed, and model verification and testing are performed;

[0008] In the above step three, the preprocessed real-time data is input into the trained estimation model to obtain the real-time estimation result of NOx concentration.

[0009] As a further technical solution of the present invention, in the step one, the historical data of the full-operation condition operation of the SCR denitrification system is screened from the historical station of the on-site thermal power unit, the data is exported, and pre-processed such as filtering, outlier removal, and standardization are performed on it. According to the NOx concentration estimation requirements, the output and input of the estimation model, i.e., the estimation variables and auxiliary variables, are determined, and the data set is divided into a training set, a validation set, and a test set.

[0010] As a further technical solution of the present invention, in the step 2, the adaptive GMM two-step clustering method is used to divide the training set data into working conditions, and the working condition discriminator is obtained. According to the data set after the working condition division, the individual learners based on QSCSO-BLS are trained one by one, and the working condition membership and the output of the individual learner are used as the combiner input to train the combiner parameters, integrate the individual learners and the combiner to obtain the estimation model, use the validation set to verify the performance of the estimation model, and use the test set to test the performance of the estimation model.

[0011] As a further technical solution of the present invention, in the step 2, the individual learner is constructed using the QSCSO-BLS model, and each operating condition corresponds to an individual learner. When training the individual learner, the adaptive GMM two-step clustering method is first used to cluster the historical operation data. The clustering result is the category label, membership degree and number of operating conditions to which each historical data belongs, where the type label is a different operating condition. Then, the data under the same operating condition is input into the QSCSO-BLS model trainer for training to obtain the individual learner under this operating condition. The operating condition discriminator uses the clustering model obtained after clustering, and the input is the real-time operation data, and the output is the corresponding membership degree of each operating condition. The input of the combiner is the output of each QSCSO-BLS individual learner. And the degree of membership The output is the model estimation result:

[0012]

[0013] Where f is the combiner model, which usually adopts a parameter regression model, such as a neural network, an extreme learning machine, etc.

[0014] As a further technical solution of the present invention, in the step 2, in order to overcome the limitations of the global search of SCSO population individuals, it is easy to fall into local optimal problems, and improve the convergence accuracy and optimization efficiency of the algorithm. The Lévy flight strategy, quantum strategy and SCSO are combined to propose a hybrid quantum sand dune cat swarm algorithm (QSCSO). The specific improvement is to add the Lévy flight strategy to improve the position update formula of SCSO in the attack phase. In the optimization algorithm, the population individual position update formula of the Lévy flight strategy is as follows:

[0015]

[0016] In the formula They are the current position of the individual sand dune cat and the updated position of the individual sand dune cat, It means point multiplication. The ability of the sand cat population to update and optimize its individual positions is mainly affected by the parameter hearing sensitivity r, which is the hearing sensitivity range r. G is obtained randomly based on G As the iterative process is performed by S M When the linear speed drops to 0, the step length change rate remains unchanged, resulting in low randomness and easy to fall into the local optimum. Therefore, for the attack phase, the Lévy flight strategy is used to increase randomness and expand the search range. The improved attack phase formula is as follows:

[0017]

[0018] Where U(d) is the random walk step length of the Lévy flight strategy, and the constant c∈[0,1] is used to balance the random walk step length to ensure that the sand cat's attack on prey is not affected by the large walk step length.

[0019] As a further technical solution of the present invention, in the step 2, a quantum strategy is used to improve the SCSO so that the hunting behavior of the sand dune cat has quantum probabilistic meaning. Since the position of particles in the quantum space is arbitrary, when the hunting behavior of the sand dune cat has quantum probabilistic meaning, its hunting trajectory and moving speed have uncertainty, and it can appear at any position in the quantum space like a particle. In this way, the entire feasible solution space is used as the search range to overcome the limitations of the global search of SCSO. The Monte Carlo method is used to obtain the position update formula of the sand dune cat individual in the iterative process as follows:

[0020]

[0021] Where i is the serial number of the sand cat, is the updated position of the individual sand cat, u i is a random number in the interval [0,1], P i (t) is the potential well (optimal position) of the particle in quantum space, L i (t) is the characteristic length of the particle’s potential well (constraining the search range of the particle), P i (t) and L i The calculation formula of (t) is as follows:

[0022]

[0023] In the formula and are the optimal position of the entire population of sand dune cats and the current optimal position of the i-th sand dune cat individual; α and γ are random numbers in the interval [0,1]; mb is the aggregation center point of the optimal position of the entire sand dune cat population, and its calculation formula is as follows:

[0024]

[0025] Where N is the number of sand dune cat individuals in the population. In summary, the position update formula of sand dune cat individuals improved by quantum strategy is:

[0026]

[0027] At the same time, in order to avoid the problem that the quantum strategy will cause the population to decrease in the later stage of algorithm iteration, the attenuation factor λ is introduced into the formula for improvement. Finally, the individual position update formula of the sand dune cat improved by the quantum strategy is obtained as follows:

[0028]

[0029] In the formula, f best (t) is the global optimal fitness under the current number of iterations.

[0030] As a further technical solution of the present invention, in the step 2, a greedy algorithm is used to improve the SCSO to improve the algorithm's solution efficiency. After each iteration of SCSO, a greedy algorithm is used to divide the sand dune cat population into three parts according to fitness, and a mutation operation is performed. If the mutated sand dune cat individual is better than the one before mutation, the position is updated again, otherwise the original sand dune cat individual position is maintained unchanged. First, the average fitness of the sand dune cat population at the tth iteration is calculated:

[0031]

[0032] Where f i(t) is the fitness value of the individual sand dune cat; f avg (t) is the average fitness of the sand dune cat population, when f i (t)≥f avg (t), the current sand dune cat individual is in a divergent state, and the quantum strategy-improved sand dune cat individual position update formula is used for mutation; when f i (t) = f best (t), according to the population individual position update formula of Lévy flight strategy, the mutation is performed; when f i (t)<f avg At (t), the current sand dune cat individuals are in a clustered state. The Lévy flight strategy is used to mutate the sand dune cat individuals so that they move closer to the optimal sand dune cat individuals. The position update formula is as follows:

[0033]

[0034] As a further technical solution of the present invention, in step 2, the QSCSO flow chart is shown in the figure below, and the specific steps are as follows: (1) Initialize the sand dune cat population parameters (number of sand dune cats, S M , variable dimension d, maximum number of iterations iter Max , the upper and lower limits of the feasible solution ub and lb); (2) Calculate the fitness f of each sand dune cat i (t), and f i (t) Sort and find the location of the best sand dune cat individual and its optimal fitness f best (t); (3) Calculate the auditory sensitivity r and balance parameter R of each sand dune cat; (4) Each sand dune cat selects the corresponding hunting behavior to update its position according to the size of R. If |R|≤1, select the attack behavior and update the position according to the improved attack phase formula; if |R|>1, select the search behavior to update the position; (5) Update the fitness f of each sand dune cat individual i (t), and f i (t) Sort and find the location of the best sand dune cat individual and its optimal fitness f best (t), calculate the average fitness of the sand dune cat population f according to the average fitness formula of the sand dune cat population during iteration avg (t).

[0035] As a further technical solution of the present invention, in step 2, (6) traverse the sand dune cat population and perform mutation operations on the sand dune cat individuals. i (t)≥f avg (t) of the sand dune cat, the sand dune cat individual position update formula improved by quantum strategy is used to mutate the current individual; for f i (t) = fbest (t) of the sand dune cat, the population individual position update formula of the Lévy flight strategy in the optimization algorithm is used to mutate the current individual; for f i (t)<f avg (t) of the dune cat, the population individual position update formula of the Lévy flight strategy in the optimization algorithm is used to mutate the current individual; (7) the mutated sand dune cat individual is evaluated. If its fitness is better than that of the sand dune cat individual before mutation, the position and fitness of the sand dune cat individual after mutation are used to update the position and fitness of the sand dune cat individual before mutation, otherwise the position and fitness remain unchanged; (8) it is determined whether the current optimal sand dune cat individual meets the requirements, or the current number of iterations reaches the set maximum number of iterations. If so, the algorithm iteration ends and the position and fitness of the optimal sand dune cat individual are output. Otherwise, it returns to step 2 to continue iteration.

[0036] As a further technical solution of the present invention, in step three, the real-time operation data of the SCR denitrification system of the on-site thermal power unit is preprocessed and then input into the estimation model to obtain the real-time estimation result of NOx concentration.

[0037] Compared with the prior art, the beneficial effects of the present invention are as follows: the present invention proposes a hybrid quantum sand dune cat swarm algorithm QSCSO, and improves SCSO by adding Lévy flight strategy and quantum strategy, thereby overcoming the problems of low convergence accuracy and low optimization efficiency caused by the limitation of global search of SCSO and the easy falling into local optimum. A modeling method based on QSCSO-BLS is designed for individual learner construction, which improves the training accuracy and generalization ability of the BLS model when modeling a single-operating system. At the same time, a boiler NOx emission prediction method based on QSCSO-BLS and integrated learning is proposed, which solves the problem of poor model generalization ability when modeling the BLS full-operating system, improves the accuracy of NOx prediction, solves the real-time and accuracy problems when CEMS measures NOx concentration, and lays a foundation for precise ammonia injection control of SCR denitrification system. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 is a flow chart of the method of the present invention;

[0039] Figure 2 It is the QSCSO flow chart of the present invention;

[0040] Figure 3 It is a schematic diagram of the model algorithm of the present invention;

[0041] Figure 4 It is a flowchart of the method of the present invention;

[0042] Figure 5 This is a schematic diagram of the SCR denitration system of the present invention;

[0043] Figure 6 It is the algorithm iteration curve diagram in Example 1;

[0044] Figure 7 Schematic diagram of the RVFLNN and BLS network structure;

[0045] Figure 8 The clustering result distribution diagram of the active power and total wind volume of the generator changing with time steps under different working conditions in Example 2;

[0046] Fig. 9 The training set training effect and the test set test effect diagram in Example 2;

[0047] Fig.10 A comparison chart of the training effects of different modeling methods in Example 2 on the training set;

[0048] Fig.11 This is a comparison chart of the test results of different modeling methods in Example 2 on the test set;

[0049] Fig.12 This is a diagram showing the training effect of the individual learner in Example 2;

[0050] Fig.13 This is a training effect diagram of the NOx outlet concentration modeling training set in Example 2;

[0051] Fig.14 This is the test effect diagram of the NOx outlet concentration modeling test set in Example 2. DETAILED DESCRIPTION

[0052] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0053] Please see attached Figure 1 -Attached Fig.14 The present invention provides a technical solution: a boiler NOx emission estimation method based on a width learning system and integrated learning, comprising step one, data processing; step two, integrated learning estimation model construction; step three, real-time estimation;

[0054] In the above step 1, the full-operation operation data of the SCR denitrification system in the thermal power unit historical station is screened and preprocessed, the model input and output are determined, and the data set is divided;

[0055] In the above step 2, the training set is divided into working conditions, combiner training and model integration are performed, and model verification and testing are performed;

[0056] In the above step three, the preprocessed real-time data is input into the trained estimation model to obtain the real-time estimation result of NOx concentration.

[0057] The following will be specifically described in conjunction with the embodiments:

[0058] Embodiment 1:

[0059] QSCSO algorithm performance test:

[0060] In order to verify the effectiveness of QSCSO, QSCSO is compared with SCSO, particle swarm optimization (PSO), artificial hummingbird algorithm (AHA) and sparrow algorithm (SSA) in experiments. PSO and AHA are classic algorithms that have been successfully used in practical optimization problems. SSA is a powerful optimization algorithm proposed in recent years. Two types of benchmark functions, unimodal function and multimodal function, are selected for testing. Unimodal function has only one optimal solution, which is suitable for testing the convergence speed and global search ability of the algorithm; multimodal function has multiple local optimal solutions, which is suitable for testing the local optimal avoidance and exploration ability of the algorithm. The specific information of the two types of functions is shown in Table 1 and Table 2 respectively. The five optimization algorithms all set the population size to 50 and the maximum number of iterations to 500. The other parameters unique to each algorithm are set according to Table 3.

[0061] Table 1 Unimodal benchmark functions

[0062]

[0063] Table 2 Multimodal benchmark functions

[0064]

[0065] Table 3 Main parameter settings of each optimization algorithm

[0066] algorithm parameter PSO <![CDATA[Forgetting factor w = 0.6, cognitive factor c1 = 1.2, social factor c2 = 1.2]]> AHA Migration coefficient M = 2N (N is the population size, N = 50) SSA Safety value ST = 0.2, 20% discoverers, 80% joiners, 10% scouts SCSO <![CDATA[Maximum sensitivity S M = 2]]> QSCSO <![CDATA[Maximum sensitivity S M = 2]]>

[0067] The five optimization algorithms were independently run 30 times in the above eight benchmark test functions to reduce the impact of algorithm randomness on the experimental results. The average and standard deviation of the results were recorded as evaluation indicators of the five optimization algorithms. The comparison results and fitness curves of the five optimization algorithms are shown in Table 4 and Appendix. Figure 6 As shown;

[0068] Table 4 Benchmark function test results

[0069]

[0070]

[0071] Since the mean and standard deviation of the results of AHA, SSA, SCSO and QSCSO in the multi-peak benchmark functions F5 and F6 tests are all 0, the experimental results are the same, and the optimization effects of the four algorithms cannot be compared by the two evaluation indicators of mean and standard deviation. Therefore, the evaluation indicators are replaced by the average number of iterations and running time when the four algorithms reach the optimal value at 0.00E+00 accuracy. They are run independently 30 times in the F5 and F6 benchmark test functions. The experimental results are shown in Table 5.

[0072] Table 5 Average number of iterations and running time at 0.00E+00 accuracy

[0073]

[0074] Embodiment 2:

[0075] Boiler NOx emission estimation example and analysis:

[0076] 1. Research objects and data preprocessing

[0077] A 660MW supercritical thermal power unit was selected as the research object. The NOx outlet concentration of its SCR denitrification system was estimated. The historical data of the SCR denitrification system required for the NOx outlet concentration estimation model was established. The data were screened from the historical stations of the thermal power unit. According to the NOx generation mechanism and reaction mechanism, 11 variables were preliminarily selected as auxiliary variables of the estimation model, and the NOx outlet concentration of the SCR denitrification system was used as the estimation variable. The basic information of the auxiliary variables and the estimation variables is shown in Table 6. According to the sampling interval of 10s, the 2023 69120 data from 12:00 on November 4, 2020 to 12:00 on November 12, 2023 are used as experimental data. This section of data covers the load range of 35% to 100%. Considering that the NOx concentration measurement value of the CEMS device of the SCR denitrification system is unavailable during the purge period, this part of the data needs to be eliminated, and finally 63655 groups of data are obtained, and the 1st to 44000th groups of data are used as training sets, the 44001st to 54000th groups of data are used as validation sets, and the 54001st to 63655th groups of data are used as test sets;

[0078] Table 6 Basic information of auxiliary variables and estimated variables

[0079]

[0080] In order to ensure that there is "less redundant information and more relevant information" in the auxiliary variables and eliminate invalid information, the mutual information method is used to measure the correlation between the auxiliary variables and the estimated variables. The results are shown in Table 7:

[0081] Table 7 Mutual information between auxiliary variables and soft measurement variables

[0082]

[0083]

[0084] According to the results in Table 7, the total coal volume, total air volume, ammonia injection flow rate, inlet NOx concentration, outlet oxygen content, load, and inlet oxygen content are finally selected as the input features of the estimation model. The total coal volume and total air volume reflect the combustion conditions of the boiler and can directly reflect the generation of NOx; the ammonia injection flow rate and inlet NOx concentration have a direct impact on the outlet NOx concentration; the outlet oxygen content, generator active power and inlet oxygen content also have a strong correlation with the outlet NOx concentration;

[0085] Afterwards, the selected historical data are preprocessed by filtering, outlier removal, and standardization. At the same time, in order to eliminate the NOx concentration measurement lag and ensure that the output of the estimation model obtained after training is a real-time NOx outlet concentration estimation, the measurement lag time is estimated, and the measurement lag time of the NOx outlet concentration is obtained to be 60s. Then, the selected historical data are spatially reconstructed to offset the measurement lag in a 60s time displacement manner.

[0086] The processed data are clustered using the adaptive GMM two-step clustering method to divide the working conditions. The lower and upper limits of the number of clusters of the adaptive GMM two-step clustering method are set to 3 and 30 respectively, the probability density threshold ε1 is set to 0.01, and the cluster minimum sample ratio threshold ε2 is set to 0.005. After clustering, the working condition labels, working condition numbers, and working condition discriminators of the data are obtained. The number of working conditions after working condition division is 12. Figure 8 The clustering result distribution of the generator active power and total wind volume changing with time steps under different working conditions;

[0087] 2. QSCSO-BLS model establishment

[0088] Select the data of working condition 1, establish the QSCSO-BLS model under this working condition, and use QSCSO to optimize the weights and biases between BLS input data and feature nodes, and between feature nodes and enhanced nodes. The optimization range is set to [-1,1], and the fitness is the root mean square error calculated by five-fold cross validation of the training set. The training set training effect and test set test effect of the optimized BLS under the optimal parameters are shown in the attached figure. Fig. 9 As shown in the attached Fig. 9It can be seen that, whether from the training effect of the training set or the test effect of the test set, the output of the NOx outlet concentration estimation model built by the BLS after parameter optimization is basically consistent with the actual value, and the modeling accuracy and generalization ability are good;

[0089] The performance of the QSCSO-BLS model was further verified by setting up a comparative test and using the data from working condition 1 to establish the following: BLS model, AHA optimized BLS model (AHA-BLS), SSA optimized BLS model (SSA-BLS), SCSO optimized BLS model (SCSO-BLS). The training set was used to train the above five models. The training results of the training set are shown in the attached figure. Fig.10 The generalization ability of the above five models is tested using the test set. The test results are shown in the attached figure. Fig.11 The root mean square error of the training set and the test set is shown in Table 8:

[0090] Table 8 Root mean square error of different modeling methods in training and testing

[0091]

[0092] From the modeling effects of BLS, SCSO-BLS and QSCSO-BLS, it can be seen that in terms of the ability of different optimization algorithms to optimize BLS parameters, the QSCSO proposed in this paper can obtain better weights and biases, and has stronger search capabilities. In terms of the probability distribution of model errors, the error distribution of the BLS model optimized by QSCSO is the most concentrated, and the degree of fluctuation is the smallest.

[0093] In summary, optimizing BLS parameters can improve the network performance of BLS. Compared with other similar methods, the QSCSO-BLS model proposed in this paper has stronger training accuracy and generalization ability in establishing the NOx outlet concentration estimation model.

[0094] 3. Modeling of prediction model based on QSCSO-BLS and ensemble learning

[0095] According to the prediction method implementation process based on QSCSO-BLS and ensemble learning, a prediction model based on QSCSO-BLS and ensemble learning is constructed;

[0096] First, the individual learner under each working condition is obtained. The data after working condition division is used by adaptive GMM two-step clustering method. The auxiliary variables under each working condition are used as input and the NOx outlet concentration is used as output to train the individual learner model based on QSCSO-BLS under the working condition. The training results are shown in the attached figure. Fig.12As shown in the figure, the estimated value and actual value of the QSCSO-BLS model under each working condition are closely distributed around y=x, achieving good training accuracy. At the same time, there are also very few data points that deviate from y=x to a large extent, and the training accuracy is poor. The possible reasons are as follows: 1) After the working condition is divided, such data points are at the edge of the working condition, and the number of data points with the same properties is too small, resulting in insufficient training and excitation of the model when using QSCSO-BLS modeling; 2) There is no problem with the data point itself, but there is a disturbance in the data point during modeling, resulting in a large deviation between the estimated value and the actual value;

[0097] The combiner is trained. The function of the combiner is to assign corresponding weights to each individual learner according to the working condition membership signal of the working condition discriminator. After weighted summation, the integrated learning model is obtained. Its input is historical operation data, working condition membership signal and individual learner output, and its output is NOx outlet concentration. The training model adopts the extreme learning machine model. The combiner can make up for the shortcomings of the individual learner: since the membership of the working condition of the data point at the edge position is relatively small, and the combiner assigns weights to each individual learner according to the working condition membership corresponding to the data point during training, so for the data point at the edge position, the weight assigned by the combiner to the individual learner of the edge working condition is also small, thereby improving the modeling effect of the data point at the edge position;

[0098] The individual learner and the combiner are combined to obtain an ensemble learning prediction model, and the validation set is integrated into the training set. The integrated training set is used to retrain the ensemble learning prediction model, and then the test set is used to test the generalization ability of the model. At the same time, QSCSO-BLS is used to establish a global prediction model for NOx outlet concentration as a comparative experiment. The performance of the prediction model based on QSCSO-BLS and ensemble learning is compared horizontally. The effects of the two modeling methods on training set training and test set testing are shown in the attached figure. Fig.13 and attached Fig.14 shown.

[0099] Based on the above, the advantages of the present invention are as follows: Analysis of the results of the above comparative tests shows that the training accuracy and generalization ability of the unoptimized BLS model are the worst, while the training accuracy and generalization ability of the other optimized BLS models are significantly improved, indicating that optimizing the weights and biases between the input data and feature nodes, and between feature nodes and enhanced nodes of BLS can significantly improve the training accuracy and generalization ability of the BLS model; By comparing the modeling effects of AHA-BLS, SSA-BLS, SCSO-BLS and QSCSO-BLS, it can be seen that in terms of the ability of different optimization algorithms to optimize BLS parameters, the QSCSO proposed in this paper can obtain better weights and biases, and has stronger search capabilities, and in terms of the probability distribution of model errors, the error distribution of the BLS model after QSCSO optimization is the most concentrated. In summary, optimizing BLS parameters can improve the network performance of BLS. Compared with other similar methods, the QSCSO-BLS model proposed in this paper has stronger training accuracy and generalization ability in establishing the NOx outlet concentration estimation model. From the training set training and test set test results of the above-mentioned NOx outlet concentration modeling, it can be seen that the estimation model based on QSCSO-BLS and ensemble learning is significantly better than the global estimation model based on QSCSO-BLS in terms of training accuracy and model generalization ability. This shows that in NOx outlet concentration modeling, the combination of QSCSO-BLS and ensemble learning can improve the learning accuracy and fitting ability of complex feature objects without changing the complexity of a single model, and achieve full coverage of the full operating conditions of the feature objects.

[0100] It is obvious to those skilled in the art that the present invention is not limited to the details of the above exemplary embodiments, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

Claims

1. A boiler NOx emission estimation method based on a width learning system and integrated learning includes step 1, data processing; step 2, integrated learning estimation model construction; step 3, real-time estimation; characterized in that: In the above step 1, the full-operation operation data of the SCR denitrification system in the thermal power unit historical station is screened and preprocessed, the model input and output are determined, and the data set is divided; In the above step 2, the training set is divided into working conditions, combiner training and model integration are performed, and model verification and testing are performed; In the above step three, the preprocessed real-time data is input into the trained estimation model to obtain the real-time estimation result of NOx concentration.

2. The boiler NOx emission estimation method based on a wide learning system and integrated learning according to claim 1 is characterized in that: In the step 1, historical data of the full-operation condition operation of the SCR denitrification system is screened from the historical station of the on-site thermal power unit, the data is exported, and pre-processed by filtering, outlier removal, standardization, etc., and the output and input of the estimation model, i.e., the estimation variables and auxiliary variables, are determined according to the NOx concentration estimation requirements, and the data set is divided into a training set, a validation set, and a test set.

3. The boiler NOx emission estimation method based on a wide learning system and integrated learning according to claim 1 is characterized in that: In the step 2, the adaptive GMM two-step clustering method is used to divide the training set data into working conditions, and the working condition discriminator is obtained. According to the data set after the working condition division, the individual learners based on QSCSO-BLS are trained one by one, the working condition membership and the output of the individual learner are used as the combiner input, the combiner parameters are trained, the individual learners and the combiner are integrated to obtain the estimation model, the validation set is used to verify the performance of the estimation model, and the test set is used to test the performance of the estimation model.

4. The boiler NOx emission estimation method based on a width learning system and integrated learning according to claim 3 is characterized in that: In the step 2, the individual learner is constructed using the QSCSO-BLS model, and each operating condition corresponds to an individual learner. When training the individual learner, the adaptive GMM two-step clustering method is first used to cluster the historical operation data. The clustering result is the category label, membership degree and number of operating conditions to which each historical data belongs, where the type label is a different operating condition. Then, the data under the same operating condition is input into the QSCSO-BLS model trainer for training to obtain the individual learner under this operating condition. The operating condition discriminator uses the clustering model obtained after clustering, and the input is the real-time operation data, and the output is the corresponding membership degree of each operating condition. The input of the combiner is the output of each QSCSO-BLS individual learner. And the degree of membership The output is the model estimation result: Where f is the combiner model, which usually adopts a parameter regression model, such as a neural network, an extreme learning machine, etc.

5. The boiler NOx emission estimation method based on a width learning system and integrated learning according to claim 3 is characterized in that: In the step 2, in order to overcome the limitations of the global search of SCSO population individuals, it is easy to fall into local optimal problems, and improve the convergence accuracy and optimization efficiency of the algorithm. The Lévy flight strategy, quantum strategy and SCSO are combined to propose a hybrid quantum sand dune cat swarm algorithm (QSCSO). The specific improvement is to add the Lévy flight strategy to improve the position update formula of SCSO in the attack phase. In the optimization algorithm, the population individual position update formula of the Lévy flight strategy is as follows: In the formula They are the current position of the individual sand dune cat and the updated position of the individual sand dune cat, It means point multiplication. The ability of the sand cat population to update and optimize its individual positions is mainly affected by the parameter hearing sensitivity r, which is the hearing sensitivity range r. G is obtained randomly based on G As the iterative process is performed by S M When the linear speed drops to 0, the step size change rate remains unchanged, resulting in low randomness and easy to fall into the local optimum. Therefore, for the attack phase, the Lévy flight strategy is used to increase randomness and expand the search range. The improved attack phase formula is as follows: Where U(d) is the random walk step length of the Lévy flight strategy, and the constant c∈[0,1] is used to balance the random walk step length to ensure that the sand cat's attack on prey is not affected by the large walk step length.

6. The boiler NOx emission estimation method based on a width learning system and integrated learning according to claim 5 is characterized in that: In the step 2, the quantum strategy is used to improve the SCSO so that the hunting behavior of the sand dune cat has quantum probabilistic meaning. Since the position of particles in the quantum space is arbitrary, when the hunting behavior of the sand dune cat has quantum probabilistic meaning, its hunting trajectory and moving speed have uncertainty, and it can appear at any position in the quantum space like particles. In this way, the entire feasible solution space is used as the search range to overcome the limitations of the global search of SCSO. The Monte Carlo method is used to obtain the position update formula of the sand dune cat individual in the iterative process as follows: Where i is the serial number of the sand cat, is the updated position of the individual sand cat, u i is a random number in the interval [0,1], P i (t) is the potential well (optimal position) of the particle in quantum space, L i (t) is the characteristic length of the particle’s potential well (constraining the search range of the particle), P i (t) and L i The calculation formula of (t) is as follows: In the formula and are the optimal position of the entire population of sand dune cats and the current optimal position of the i-th sand dune cat individual; α and γ are random numbers in the interval [0,1]; mb is the aggregation center point of the optimal position of the entire sand dune cat population, and its calculation formula is as follows: Where N is the number of sand dune cat individuals in the population. In summary, the position update formula of sand dune cat individuals improved by quantum strategy is: At the same time, in order to avoid the problem that the quantum strategy will cause the population to decrease in the later stage of algorithm iteration, the attenuation factor λ is introduced into the formula for improvement. Finally, the individual position update formula of the sand dune cat improved by the quantum strategy is obtained as follows: In the formula, f best (t) is the global optimal fitness under the current number of iterations.

7. The boiler NOx emission estimation method based on a width learning system and integrated learning according to claim 6 is characterized in that: In the step 2, the greedy algorithm is used to improve the SCSO and improve the algorithm's solution efficiency. After each iteration, the SCSO uses the greedy algorithm to divide the sand dune cat population into three parts according to fitness, and performs a mutation operation. If the mutated sand dune cat individual is better than the one before mutation, the position is updated again, otherwise the original sand dune cat individual position is maintained unchanged. First, the average fitness of the sand dune cat population at the tth iteration is calculated: Where f i (t) is the fitness value of the individual sand dune cat; f avg (t) is the average fitness of the sand dune cat population, when f i (t)≥f avg (t), the current sand dune cat individual is in a divergent state, and the quantum strategy is used to improve the sand dune cat individual position update formula for mutation; when f i (t) = f best (t), according to the population individual position update formula of Lévy flight strategy, the mutation is performed; when f i (t)<f avg At (t), the current sand dune cat individuals are in a clustered state. The Lévy flight strategy is used to mutate the sand dune cat individuals so that they move closer to the optimal sand dune cat individuals. The position update formula is as follows:

8. The boiler NOx emission estimation method based on a width learning system and integrated learning according to claim 3 is characterized in that: In step 2, the QSCSO flow chart is shown in the figure below, and the specific steps are as follows: (1) Initialize the sand cat population parameters (number of sand cats, S M , variable dimension d, maximum number of iterations iter Max , the upper and lower limits of the feasible solution ub and lb); (2) Calculate the fitness f of each sand dune cat i (t), and f i (t) Sort and find the location of the best sand dune cat individual and its optimal fitness f best (t); (3) Calculate the auditory sensitivity r and balance parameter R of each sand dune cat; (4) Each sand dune cat selects the corresponding hunting behavior according to the size of R to update its position. If |R|≤1, select the attack behavior and update the position according to the improved attack phase formula; if |R|>1, select the search behavior to update the position; (5) Update the fitness f of each sand dune cat individual i (t), and f i (t) Sort and find the location of the best sand dune cat individual and its optimal fitness f best (t), calculate the average fitness of the sand dune cat population f according to the average fitness formula of the sand dune cat population during iteration avg (t).

9. The boiler NOx emission estimation method based on a wide learning system and integrated learning according to claim 1, characterized in that: In step 2, (6) traverse the sand dune cat population and perform mutation operations on the sand dune cat individuals. i (t)≥f avg (t) of the sand dune cat, the sand dune cat individual position update formula improved by quantum strategy is used to mutate the current individual; for f i (t) = f best (t) of the sand dune cat, the population individual position update formula of the Lévy flight strategy in the optimization algorithm is used to mutate the current individual; for f i (t)<f avg (t) of the dune cat, the population individual position update formula of the Lévy flight strategy in the optimization algorithm is used to mutate the current individual; (7) the mutated sand dune cat individual is evaluated. If its fitness is better than that of the sand dune cat individual before mutation, the position and fitness of the sand dune cat individual after mutation are used to update the position and fitness of the sand dune cat individual before mutation, otherwise the position and fitness remain unchanged; (8) it is determined whether the current optimal sand dune cat individual meets the requirements, or the current number of iterations reaches the set maximum number of iterations. If so, the algorithm iteration ends and the position and fitness of the optimal sand dune cat individual are output. Otherwise, it returns to step 2 to continue iteration.

10. The boiler NOx emission estimation method based on a wide learning system and integrated learning according to claim 1, characterized in that: In the step three, the real-time operation data of the SCR denitration system of the on-site thermal power unit is preprocessed and then input into the estimation model to obtain the real-time estimation result of NOx concentration.

Images