Thermal power plant SCR denitration system outlet NOx concentration prediction method and system

Abstract

The invention provides a thermal power plant SCR denitration system outlet NOx concentration prediction method and system, and relates to the technical field of artificial intelligence, and the method comprises the steps: carrying out the outlet NOx concentration prediction of a thermal power plant SCR denitration system based on a pre-established TTAO-CNN-LSTM model; the step of pre-establishing the TTAO-CNN-LSTM model comprises the steps that the historical NOx concentration and a plurality of historical SCR inlet factors are preprocessed; performing model input dimension reduction processing on the plurality of historical SCR inlet factors after preprocessing; a CNN-LSTM prediction model is constructed based on the multiple preprocessed historical SCR inlet factors subjected to dimension reduction processing and the preprocessed historical NOx concentration after model input; and setting an optimization range of a neural network hyper-parameter, carrying out optimization by adopting a triangular topology aggregation algorithm, and automatically updating an optimal value of the hyper-parameter to the CNN-LSTM prediction model to obtain a TTAO-CNN-LSTM model. The accuracy of predicting the NOx concentration at the outlet of the SCR denitration system of the thermal power plant is improved to a great extent, and automatic optimization of parameters is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of artificial intelligence, and particularly relates to a method and system for predicting the NOx concentration at the outlet of an SCR denitration system in a thermal power plant. Background Art

[0002] Currently, the main method for reducing NOx emissions is post-combustion control, which mainly treats the already generated NOx. In thermal power plants, the main post-combustion treatment method is selective catalytic reduction (SCR) technology. Due to its high denitration efficiency and mature process, SCR technology has become the preferred technology for large thermal power plants and highly polluting industries. The principle of SCR technology is that under the action of a catalyst, ammonia reacts with NOx in the flue gas, and after a reduction reaction, harmless nitrogen and water vapor are generated, thereby effectively reducing the NOx emissions. In the SCR denitration system, NOx and ammonia are mixed proportionally for the denitration reaction. If the ammonia injection amount is insufficient, the NOx emission concentration is too high, making it difficult to meet the environmental protection standards. If the ammonia injection amount is too much, it will cause ammonia slip and aggravate secondary pollution. With the rapid development and large-scale grid connection of new energy power generation, thermal power plants need to perform frequent deep peak shaving to cope with the fluctuations of new energy power generation, resulting in drastic changes in the unit operating conditions and often making it difficult to stably emit NOx.

[0003] With the rapid development of computer science and technology, the industrial field has begun to use machine learning to establish mathematical models to solve engineering problems. Currently, the main modeling methods based on the SCR denitration system are divided into two types: data modeling and mechanism modeling. Data modeling only needs to perform modeling input and modeling output on the operation data, and perform data identification to form corresponding mathematical relationships for solution. The modeling process is relatively simple and does not require understanding of the internal mechanism and structure. For complex SCR denitration systems, data modeling can use the operation data of thermal power plants for modeling. At the same time, thermal power plants are equipped with distributed control systems, which provide a large amount of operation data and provide strong support for data modeling. Therefore, more and more researchers have turned their attention to the field of data modeling. Common data modeling methods include shallow learning methods such as deep learning networks, artificial neural networks, and support vector machines. However, the models established using shallow learning methods have poor prediction effects, general prediction accuracy, and poor generalization ability, and some setting parameters in the modeling need to be manually adjusted, lacking scientificity.

[0004] In summary, it is necessary to establish a prediction model for the NOx concentration at the outlet of the SCR denitration system with high prediction accuracy and realize automatic parameter optimization. Summary of the Invention

[0005] One of the objectives of the present invention is to provide a method for predicting the NOx concentration at the outlet of the SCR denitration system in a thermal power plant. By pre-establishing a TTAO-CNN-LSTM model, based on the pre-established TTAO-CNN-LSTM model, the NOx concentration at the outlet of the SCR denitration system in the thermal power plant is predicted, which greatly improves the accuracy of predicting the NOx concentration at the outlet of the SCR denitration system in the thermal power plant and realizes automatic parameter optimization.

[0006] A method for predicting the NOx concentration at the outlet of the SCR denitration system in a thermal power plant provided by an embodiment of the present invention includes:

[0007] Based on the pre-established TTAO-CNN-LSTM model, the NOx concentration at the outlet of the SCR denitration system in the thermal power plant is predicted;

[0008] Among them, the steps for pre-establishing the TTAO-CNN-LSTM model are as follows:

[0009] Collect the historical NOx concentration and multiple historical SCR inlet factors of the SCR denitration system in the thermal power plant;

[0010] Preprocess the historical NOx concentration and multiple historical SCR inlet factors respectively;

[0011] Perform model input dimensionality reduction processing on the preprocessed multiple historical SCR inlet factors;

[0012] Based on the preprocessed multiple historical SCR inlet factors after model input dimensionality reduction processing and the preprocessed historical NOx concentration, construct a CNN-LSTM prediction model;

[0013] Set the optimization range of the neural network hyperparameters, use the triangular topology aggregation algorithm for optimization, and automatically update the optimal values of the hyperparameters into the CNN-LSTM prediction model to obtain the TTAO-CNN-LSTM model.

[0014] Optionally, the multiple historical SCR inlet factors at least include: ammonia injection amount, SCR inlet NOx concentration, SCR inlet O2 concentration, SCR inlet temperature, SCR inlet pressure, SCR inlet SO2 concentration, SCR inlet flue gas flow rate, and SCR inlet flue gas velocity.

[0015] Optionally, the preprocessing of the historical NOx concentration and multiple historical SCR inlet factors respectively includes:

[0016] Use the Gaussian weighted moving average filtering algorithm to filter and denoise the historical NOx concentration and multiple historical SCR inlet factors;

[0017] The K-nearest neighbor interpolation algorithm is used to interpolate and replace abnormal data points in historical NOx concentrations and multiple historical SCR inlet factors.

[0018] The Z-score standard is used to standardize historical NOx concentrations and multiple historical SCR inlet factors.

[0019] Optionally, the model input dimensionality reduction processing for the preprocessed multiple historical SCR inlet factors includes:

[0020] The principal component analysis method and the Pearson correlation coefficient algorithm are used to perform model input dimensionality reduction processing on the preprocessed multiple historical SCR inlet factors.

[0021] Optionally, constructing a CNN-LSTM prediction model based on the preprocessed multiple historical SCR inlet factors after model input dimensionality reduction processing and the preprocessed historical NOx concentrations includes:

[0022] The preprocessed multiple historical SCR inlet factors after model input dimensionality reduction processing and the preprocessed historical NOx concentrations are used as historical operation data;

[0023] The NOx concentration at the SCR inlet, the O2 content at the SCR inlet, the ammonia flow rate at the SCR inlet, the temperature at the SCR inlet, and the flue gas pressure at the SCR inlet are selected as the inputs of the input layer, and the NOx concentration at the SCR outlet is selected as the output of the output layer;

[0024] A CNN convolutional layer is set to filter the historical operation data and extract local features, sharing weights and biases;

[0025] A causal dilation convolution mechanism is set to extract the causal features of the output of the CNN convolutional layer;

[0026] A CNN pooling layer is set to perform dimensionality reduction sampling on the output of the causal dilation convolution mechanism;

[0027] A time-mode attention mechanism is set to process the output of the CNN pooling layer;

[0028] An LSTM layer is set to process the output of the time-mode attention mechanism;

[0029] The input layer, the CNN convolutional layer, the causal dilation convolution mechanism, the CNN pooling layer, the time-mode attention mechanism, the LSTM layer, and the output layer are used as the model to be trained;

[0030] Train the model to be trained based on historical operation data. During training, use the Adam algorithm to backpropagate the training error and gradually update the model parameters layer by layer. Use the Softmax activation function to classify the signal features and complete the classification task of the multi-feature input sequence. Finally, obtain the CNN-LSTM prediction model.

[0031] Optionally, the neural network hyperparameters at least include: learning rate, number of neurons in the hidden layer, and regularization coefficient.

[0032] Optionally, setting the optimization range of the neural network hyperparameters includes:

[0033] Set the optimization ranges of the learning rate, number of neurons in the hidden layer, and regularization coefficient to be [0.1, 1e-6], [5, 250], and [0.1, 1e-8] respectively.

[0034] Optionally, the pre-establishment step of the TTAO-CNN-LSTM model further includes:

[0035] Use the test data to verify the prediction effect of the TTAO-CNN-LSTM prediction model.

[0036] Optionally, the using the test data to verify the prediction effect of the TTAO-CNN-LSTM prediction model includes:

[0037] Use RMSE, MAPE, RPD, and R 2 Verify the prediction effect of the TTAO-CNN-LSTM prediction model according to the test data.

[0038] A NOx concentration prediction system at the outlet of the SCR denitration system in a thermal power plant provided by an embodiment of the present invention includes:

[0039] A concentration prediction module for predicting the NOx concentration at the outlet of the SCR denitration system in a thermal power plant based on the pre-established TTAO-CNN-LSTM model;

[0040] Among them, the pre-establishment steps of the TTAO-CNN-LSTM model are as follows:

[0041] Collect the historical NOx concentration and historical multiple SCR inlet factors of the SCR denitration system in a thermal power plant;

[0042] Preprocess the historical NOx concentration and historical multiple SCR inlet factors respectively;

[0043] Perform model input dimensionality reduction processing on the preprocessed historical multiple SCR inlet factors;

[0044] Based on the preprocessed historical multiple SCR inlet factors after model input dimensionality reduction processing and the preprocessed historical NOx concentration, a CNN-LSTM prediction model is constructed;

[0045] Set the optimization range of the neural network hyperparameters, use the triangular topology aggregation algorithm for optimization, and automatically update the optimal values of the hyperparameters into the CNN-LSTM prediction model to obtain the TTAO-CNN-LSTM model.

[0046] The present invention has achieved the following beneficial effects:

[0047] 1. The TTAO-CNN-LSTM prediction model has higher prediction accuracy and stronger generalization ability compared with other prediction models.

[0048] 2. The causal dilation convolution mechanism is introduced to extract the causal features in the sample data sequence, and the time pattern attention mechanism is introduced to enhance the attention of the neural network to critical moments at different time steps.

[0049] 3. During the training process of the CNN-LSTM model, the TTAO algorithm is used to automatically optimize the learning rate, the number of neurons in the hidden layer, and the L2 regularization coefficient in the neural network, which is scientific and avoids the disadvantages of manual parameter tuning.

[0050] Before model training, the Gaussian weighted moving average filtering algorithm is used for filtering and denoising of sample data, the K-nearest neighbor interpolation algorithm is used for interpolation and replacement of abnormal data points in sample data, and the Z-score standard is used for standardization of sample data. The principal component analysis method is used to extract features from sample data to determine the number of principal components; the Pearson correlation coefficient algorithm is used to determine the correlation between factors for dimensionality reduction processing. It effectively improves the quality of sample data, reduces the dimension of sample data, and reduces the impact of sample data quality on prediction accuracy during the training process.

[0051] Other features and advantages of the present invention will be described in the subsequent specification, and part of them will be obvious from the specification or understood by implementing the present invention. The objectives and other advantages of the present invention can be achieved and obtained through the structures specifically pointed out in the written specification and the drawings.

[0052] Next, through the drawings and embodiments, the technical solutions of the present invention will be further described in detail. Description of the Drawings

[0053] The drawings are used to provide a further understanding of the present invention, and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention, and do not constitute a limitation to the present invention. In the drawings:

[0054] Figure 1This is the principal component analysis Pareto chart in the embodiment of the present invention;

[0055] Figure 2 This is the Pearson correlation coefficient chart in the embodiment of the present invention;

[0056] Figure 3 This is the structural diagram of the CNN-LSTM prediction model in the embodiment of the present invention;

[0057] Figure 4 This is the flow chart of parameter optimization for the CNN-LSTM prediction model based on the TTAO algorithm in the embodiment of the present invention;

[0058] Figure 5 This is the optimization iteration chart of the TTAO algorithm in the embodiment of the present invention;

[0059] Figure 6 This is the comparison chart of the prediction results of the training sets of four prediction models in the embodiment of the present invention;

[0060] Figure 7 This is the comparison chart of the prediction results of the test sets of four prediction models in the embodiment of the present invention;

[0061] Figure 8 This is the relative error chart of the prediction results of the TTAO-CNN-LSTM prediction model in the embodiment of the present invention;

[0062] Figure 9 This is the comparison chart of the evaluation criteria of four prediction models in the embodiment of the present invention. Detailed implementation manners

[0063] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are only for the purpose of illustrating and explaining the present invention, and are not intended to limit the present invention.

[0064] The embodiment of the present invention provides a method for predicting the NOx concentration at the outlet of the SCR denitration system in a thermal power plant, including:

[0065] S1. Based on the pre-established TTAO-CNN-LSTM model, predict the NOx concentration at the outlet of the SCR denitration system in the thermal power plant; the pre-established TTAO-CNN-LSTM model will predict the NOx concentration at the outlet according to the current multiple SCR inlet factors (ammonia injection amount, SCR inlet NOx concentration, SCR inlet O2 concentration, SCR inlet temperature, SCR inlet pressure, SCR inlet SO2 concentration, SCR inlet flue gas flow rate, and SCR inlet flue gas velocity) of the SCR denitration system in the thermal power plant;

[0066] Among them, the steps for pre-establishing the TTAO-CNN-LSTM model are as follows:

[0067] Collect the historical NOx concentration and multiple historical SCR inlet factors of the thermal power plant SCR denitration system;

[0068] For the historical NO X Concentration and multiple historical SCR inlet factors are respectively preprocessed;

[0069] Perform model input dimensionality reduction processing on the preprocessed multiple historical SCR inlet factors;

[0070] Based on the preprocessed multiple historical SCR inlet factors and the preprocessed historical NOx concentration after model input dimensionality reduction processing, construct a CNN-LSTM prediction model;

[0071] Set the optimization range of the neural network hyperparameters, use the triangular topology aggregation algorithm for optimization, and automatically update the optimal values of the hyperparameters into the CNN-LSTM prediction model to obtain the TTAO-CNN-LSTM model.

[0072] This application pre-establishes a TTAO-CNN-LSTM model. Based on the pre-established TTAO-CNN-LSTM model, predict the outlet NOx concentration of the thermal power plant SCR denitration system, greatly improving the prediction accuracy of the outlet NOx concentration of the thermal power plant SCR denitration system and realizing automatic parameter optimization.

[0073] In one embodiment, the multiple historical SCR inlet factors at least include: ammonia injection amount, SCR inlet NOx concentration, SCR inlet O2 concentration, SCR inlet temperature, SCR inlet pressure, SCR inlet SO2 concentration, SCR inlet flue gas flow rate, and SCR inlet flue gas velocity.

[0074] In one embodiment, the preprocessing of the historical NOx concentration and the multiple historical SCR inlet factors respectively includes:

[0075] Use the Gaussian weighted moving average filtering algorithm to filter and denoise the historical NOx concentration and the multiple historical SCR inlet factors;

[0076] Use the K-nearest neighbor interpolation algorithm to interpolate and replace the abnormal data points in the historical NOx concentration and the multiple historical SCR inlet factors;

[0077] Use the Z-score standard to standardize the historical NOx concentration and the multiple historical SCR inlet factors.

[0078] The data preprocessing method mainly includes: filtering and denoising, outlier interpolation and replacement, and standardization.

[0079] The main steps of using the Gaussian weighted moving average filtering algorithm for filtering and denoising sample data (historical NOx concentration and multiple historical SCR inlet factors) are as follows:

[0080] The mathematical expression of the Gaussian weighted moving average filtering algorithm is as follows:

[0081]

[0082] Among them, w(i) is the weight of the i-th data point in the sample data, c is the current time point, σ is the standard deviation of the Gaussian function, exp(…) is the natural exponential function, and i is the i-th data point.

[0083] Define a filtering window, including the current sample point and several surrounding sample points.

[0084] Calculate the weight of each sample point in the window. The weight is calculated according to the Gaussian distribution, and the points farther away from the current sample point have smaller weights.

[0085] Perform weighted averaging on the values of each sample point in the window to obtain the filtering result of the current sample point.

[0086] Move the filtering window forward by one position and repeat the above steps until all sample points are processed.

[0087] The main steps of using the K-nearest neighbor interpolation algorithm for interpolating and replacing abnormal data points in sample data are as follows:

[0088] Determine the value of K: Usually, K is a small integer, and the common value range is between 1 and 5.

[0089] Calculate the distance: For each sample with missing values or anomalies, calculate its distance from all other samples in the training set. The commonly used distance metric method is the Euclidean distance.

[0090] Select neighbors: According to the calculated distances, select the K samples with the smallest distances as neighbors.

[0091] Estimated value: Use the corresponding values of the neighbors to estimate or fill in the missing values and anomalies. For the case of missing values, usually calculate the average or weighted average of the corresponding features of these neighbors as the filling value.

[0092] The main step of standardizing sample data using the Z-score standard is to calculate the difference between each data point and the mean of the data set and divide it by the standard deviation of the data set, thereby converting the original data into a new data set with zero mean and unit variance:

[0093]

[0094] Among them, Z is the standardized value, X is the original data point in the sample data, μ is the mean of the original data set in the sample data, and σ is the standard deviation of the original data set in the sample data.

[0095] In one embodiment, the model input dimensionality reduction processing of the preprocessed historical multiple SCR inlet factors includes:

[0096] Perform model input dimensionality reduction processing on the preprocessed historical multiple SCR inlet factors by using the principal component analysis method and the Pearson correlation coefficient algorithm.

[0097] Use the principal component analysis method to extract features from the sample data to determine the number of principal components; use the Pearson correlation coefficient algorithm to determine the correlation between factors for dimensionality reduction processing;

[0098] The main steps of using the principal component analysis method to extract features from the sample data to determine the number of principal components are as follows:

[0099] Data standardization: Standardize the original data so that features in different dimensions have the same scale. The mathematical expression is as follows:

[0100]

[0101] Among them, is the mean of the j-th factor in the sample matrix X of the sample data, is the standard deviation of the j-th factor in the sample matrix X, z ij is the value of the j-th factor at the i-th sampling after standardization processing, n is the total number of factors in the sample matrix, x ij is the value of the j-th factor at the i-th sampling in the sample matrix X.

[0102] Calculate the covariance matrix: Calculate the covariance matrix based on the standardized data, and this matrix describes the correlation between each dimension in the data. The mathematical expression is as follows:

[0103]

[0104] Among them, is the mean of the j-th factor in the standardized matrix Z; r ij is the relationship between the i-th factor and the j-th factor, z ij is the value of the j-th factor at the i-th sampling after standardization processing, n is the number of factors.

[0105] Perform eigenvalue decomposition on the covariance matrix to obtain eigenvalues and corresponding eigenvectors. The eigenvalues represent the variance contribution of each principal component in the data, and the eigenvectors describe the direction of each principal component. The mathematical expression is as follows:

[0106]

[0107] Among them, ω i is the contribution degree of the main component, ρ is the cumulative contribution degree of the first m main components, and λ i is the eigenvalue of the i-th factor, and λ j is the eigenvalue of the j-th factor, and n is the number of factors.

[0108] Eigenvalue sorting and principal component selection: Sort according to the magnitude of the eigenvalues, and select the eigenvectors corresponding to the first k eigenvalues as the basis vectors of the new coordinate system. The mathematical expression is as follows:

[0109]

[0110] Among them, y i is the projection of the original data in the principal component space, x j is the j-th factor, e ij is the eigenvector corresponding to the eigenvalue λ i in e i , k is the number of principal components.

[0111] Data mapping: Project the original data into the new low-dimensional space composed of the selected eigenvectors to obtain the data after dimensionality reduction.

[0112] Sort the contribution rates from large to small, and determine the number of principal components N based on the standard of cumulative contribution rate of 95%.

[0113] Use the Pearson correlation coefficient algorithm to determine the correlation between factors, and the main steps of dimensionality reduction are as follows:

[0114] Based on the research of principal component analysis, determine to select N inlet factors from 8 SCR inlet factors as the input of the data model.

[0115] Determine the correlation between the outlet NOx concentration and 8 SCR inlet factors, and sort the correlations from large to small. The mathematical expression is as follows:

[0116]

[0117] Among them, X and Y are the observed values of two variables, ρ(X, Y) is the Pearson correlation coefficient between the two variables, cov(X, Y) is the covariance of X and Y, and ρXρY is the standard deviation of X and Y.

[0118] Select N SCR inlet factors with large correlation with the outlet NOx concentration as the input of the prediction model.

[0119] In one embodiment, constructing a CNN-LSTM prediction model based on the preprocessed historical multiple SCR inlet factors and the preprocessed historical NOx concentration after model input dimensionality reduction processing, includes:

[0120] Taking the preprocessed historical multiple SCR inlet factors and the preprocessed historical NOx concentration after model input dimensionality reduction processing as historical operation data;

[0121] Selecting the SCR inlet NOx concentration, SCR inlet O2 content, SCR inlet ammonia flow rate, SCR inlet temperature, and SCR inlet flue gas pressure as the inputs of the input layer, and selecting the SCR outlet NOx concentration as the output of the output layer;

[0122] Setting a CNN convolutional layer to filter the historical operation data and extract local features, sharing weights and biases;

[0123] Setting a causal dilated convolution mechanism to extract the causal features of the output of the CNN convolutional layer;

[0124] Setting a CNN pooling layer to perform dimensionality reduction sampling on the output of the causal dilated convolution mechanism;

[0125] Setting a time-mode attention mechanism to process the output of the CNN pooling layer;

[0126] Setting an LSTM layer to process the output of the time-mode attention mechanism;

[0127] Taking the input layer, CNN convolutional layer, causal dilated convolution mechanism, CNN pooling layer, time-mode attention mechanism, LSTM layer, and output layer as the model to be trained;

[0128] Training the model to be trained based on the historical operation data. During training, using the Adam algorithm to backpropagate the training error and gradually update the model parameters layer by layer, using the Softmax activation function to classify the signal features, completing the classification task of the multi-feature input sequence, and finally obtaining the CNN-LSTM prediction model.

[0129] First, input the preprocessed sample data into the input layer. Second, the sample data is first processed by the CNN. The convolutional layer filters the sample data and extracts local features, sharing weights and biases, improving the learning efficiency of the model. The pooling layer realizes effective dimensionality reduction sampling of the sample data, continuously reducing the size of the sample data space, and ensuring the invariance of the features. Third, the CNN is fused with the LSTM to capture the time series features of the sample data. Finally, the prediction result is output through the fully connected layer, that is, the prediction result of the SCR outlet NOx concentration.

[0130] The causal dilated convolution mechanism is a variant of the conventional convolution, which is composed of the causal convolution and the dilated convolution.

[0131] Causal convolution adjusts the convolutional kernel so that the convolution operation only involves the input data at the current time and previous times, avoiding the influence of future information on the prediction at the current time of the model and ensuring that the model follows the causal relationship. The definition of causal convolution is:

[0132]

[0133] where y(t) is the output signal, x(t) is the input signal, w(k) is the convolutional kernel weight, (t - k) is the value of the input at time t - k, and k is the length of the convolutional kernel.

[0134] Atrous convolution can cover a larger range of input data by introducing gaps or dilations between each weight of the convolutional kernel, thereby effectively expanding the receptive field of the convolutional kernel. The definition of atrous convolution is:

[0135]

[0136] where d is the dilation factor, K is the index of the convolutional kernel.

[0137] Causal atrous convolution combines the advantages of causal convolution and atrous convolution, which can not only ensure that the convolution operation does not leak future information but also expand the receptive field. The mathematical definition of causal atrous convolution is:

[0138]

[0139] where w(k) is the convolutional kernel weight; (t - k) is the value of the input at time t - k; k is the length of the convolutional kernel; d is the dilation factor.

[0140] The principle of the temporal pattern attention mechanism is to assign a weight to each time step, which reflects the importance of that time step in the overall sequence. In this way, the model can focus on the time steps with more information or more contributions to the final task. This optimization process is as follows:

[0141] Input time series. Given a time series input X = [x1, x2,..., x t , where x T represents the input feature at time t.

[0142] The attention weight α t is calculated based on the similarity between the query vector and all key vectors, and the commonly used calculation method is dot product attention.

[0143]

[0144] where α tis the attention weight, score() is the scoring function, and Q t ·K t' is the dot product of the query and the key, T is the time, and i is the time step.

[0145] The scoring function score is usually defined as the dot product of the query vector and the key vector, and is scaled by a scaling factor for scaling.

[0146]

[0147] where score is the scoring function; Q t ·K t' is the dot product of the query and the key; d k is the dimension of the key.

[0148] The output is obtained through the weighted value vector. For each time step t, the output y t is the weighted sum of all time step value vectors.

[0149]

[0150] where α t is the attention weight, v t is the value vector corresponding to the time step i, T is the time, and i is the time step.

[0151] The output Y = [y1, y2,..., y t of the final temporal pattern attention mechanism is the representation of all time steps after weighting.

[0152] Output result. Subsequent task processing is performed based on the weighted context vector.

[0153] In one embodiment, the neural network hyperparameters at least include: learning rate, the number of neurons in the hidden layer, and the regularization coefficient;

[0154] The search range for setting the neural network hyperparameters includes:

[0155] The search ranges for setting the learning rate, the number of neurons in the hidden layer, and the regularization coefficient are [0.1, 1e-6], [5, 250], and [0.1, 1e-8], respectively.

[0156] The process of optimizing the learning rate, the number of neurons in the hidden layer, and the L2 regularization coefficient of the CNN-LSTM neural network using the TTAO algorithm is as follows:

[0157] Initialization: In the initialization stage, N / 3 individuals are randomly generated in the feasible region. The mathematical expressions for generating each individual are as follows:

[0158]

[0159] Among them, is the first search individual in the \(i\)-th triangular topological unit; \(i\) is a positive integer between 1 and \(N / 3\); \(r_0\) is a random number between \([0, 1]\); are the lower and upper limits of the variable, \([lb_1,..., lb\) D is the lower limit vector, \([ub_1,..., ub\) D is the upper limit vector.

[0160] Formation of triangular topological unit: Starting from the first vertex, locate a new direction vector in the spherical coordinate system with a length of Convert it to the ordinary coordinate system through trigonometric functions to form the second vertex. The generated direction vector with a length of is rotated counterclockwise by \(\pi / 3\), and then transformed through the coordinate system to obtain the third vertex. The mathematical expressions for generating these vertices are as follows:

[0161]

[0162]

[0163] Among them, is the first vertex, is the second vertex, is the third vertex, is the direction vector, is the direction vector, \(l\) is the size of the triangular topological unit, \(t\) is the current iteration number, and \(T\) is the maximum iteration number.

[0164] Each group of triangular topological units is aggregated internally to form the fourth vertex in a linearly weighted manner. The mathematical expression for generating the fourth vertex is as follows:

[0165]

[0166] Among them, \(r_1\), \(r_2\), and \(r_3\) are random numbers between \([0, 1]\).

[0167] Generic aggregation: Information interaction occurs between the optimal individual in each triangular topological unit and the optimal individual in an arbitrarily selected unit set, and a new individual is generated among the better two-vertex connections. The mathematical expression for generating the new individual is as follows:

[0168]

[0169] Among them, is the latest individual at the \(i\)-th iteration, \(r_4\) is a random number between \([0, 1]\), is the optimal individual at the \(i\)-th iteration, To randomly generate the optimal individual;

[0170] The mathematical expressions for the optimal and sub-optimal individuals updated at the (t + 1)-th iteration are as follows:

[0171]

[0172] Where, is the optimal individual at the i-th iteration, is the latest individual at the i-th iteration, is the sub-optimal individual at the i-th iteration, and f(·) is the function of the given problem.

[0173] Local aggregation: The position of the optimal individual is locally perturbed in terms of direction and step size based on the motion vector difference formed by the optimal and sub-optimal individuals, so as to re-search each group within a certain local area and achieve the exploitation of each topological triangular unit. The mathematical expression is as follows:

[0174]

[0175] Where, is the optimal individual at the i-th iteration, is the latest individual at the i-th iteration, is the sub-optimal individual at the i-th iteration, f(·) is the function of the given problem, i is the number of iterations, and t is the time.

[0176] Among them, α decreases, which adjusts the size of the aggregation range. The mathematical expression of α is as follows:

[0177]

[0178] Where, T is the time.

[0179] If the new individual is better than the original individual, update the position, otherwise do not update. The mathematical expression is as follows:

[0180]

[0181] Where, is the optimal individual at the i-th iteration, is the latest individual at the i-th iteration, f(·) is the function of the given problem, i is the number of iterations, and t is the time.

[0182] In one embodiment, the pre-establishment steps of the TTAO-CNN-LSTM model further include:

[0183] Using test data to verify the prediction effect of the TTAO-CNN-LSTM prediction model;

[0184] Verifying the prediction effect of the TTAO-CNN-LSTM prediction model using test data includes:

[0185] Using RMSE, MAPE, RPD, and R 2 Verifying the prediction effect of the TTAO-CNN-LSTM prediction model according to the test data.

[0186] In one embodiment, the provided data is sourced from the historical operation data collected by the CEMS system of a certain thermal power plant, collected every 5 seconds for approximately 3 hours, a total of 2000 groups. This set of data can represent the sample data during the normal operation period of the unit, and the data is divided into a training set and a test set according to 4:1.

[0187] Step 1: Obtain the historical operation data in the CEMS system of a certain thermal power plant in Shanxi, arranged in order from 1 to 9, mainly including: 1. Ammonia injection volume, 2. SCR inlet NOx concentration, 3. SCR inlet O2 concentration, 4. SCR inlet temperature, 5. SCR inlet pressure, 6. SCR inlet SO2 concentration, 7. SCR inlet flue gas flow rate, 8. SCR inlet flue gas velocity, and 9. SCR outlet NOx concentration.

[0188] Step 2: Perform preprocessing and dimensionality reduction on the collected data. The processing methods are as follows:

[0189] Data preprocessing: Use the Gaussian weighted moving average filtering algorithm to filter and denoise the sample data, use the K-nearest neighbor interpolation algorithm to interpolate and replace the abnormal data points in the sample data, and use the Z-score standard to standardize the sample data.

[0190] Principal component analysis: Use the PCA method for dimensionality reduction. The data has been filtered, denoised, standardized, and outlier replaced. The Pareto chart of the calculated principal component analysis is as Figure 1 shown. From Figure 1 it can be seen that the contribution rate of the first principal component is 36.5581%, the cumulative contribution rate of the first 4 principal components to the original data set is 88.6899%, and the cumulative contribution rate of the first 5 principal components is 96.5664%. Taking the cumulative contribution rate of 95% as the screening criterion, select the first 5 principal components as the input variables of the model. These principal component variables reflect 95% of the information of the original input variable data.

[0191] Correlation analysis: Use the Pearson correlation coefficient for screening. The principle is that if the correlation coefficient is close to 1, it means there is a high positive correlation between the two variables; if the correlation coefficient is close to 0, it means the two variables are independent of each other and have no correlation; if the correlation coefficient is negative, it means there is a negative correlation between the two variables. Select the collected historical data to generate a correlation coefficient matrix and visualize it, as Figure 2As shown in the figure. The generated Pearson correlation coefficient heat map can be used to screen out the parameters that have a large correlation with each other and a high correlation with the outlet NOx concentration as input parameters. Among them, the parameter with the highest correlation coefficient with the outlet NOx concentration is the ammonia water flow rate at the SCR inlet (0.8343), followed by the O2 content at the SCR inlet (-0.6138), the NOx concentration at the SCR inlet (0.5989), the temperature at the SCR inlet (-0.4851), and the flue gas pressure at the SCR inlet (-0.4055). A negative parameter represents a negative correlation. Therefore, the above five influencing factors are selected for subsequent modeling.

[0192] Step 3: Select the NOx concentration at the SCR inlet, the O2 content at the SCR inlet, the ammonia water flow rate at the SCR inlet, the temperature at the SCR inlet, and the flue gas pressure at the SCR inlet as the input of the model, and select the NOx concentration at the SCR outlet as the output of the model. In the CNN, the convolutional layer filters the historical data and extracts local features, sharing weights and biases to improve the learning efficiency of the model. The pooling layer realizes effective dimensionality reduction sampling of the historical data, continuously reducing the size of the historical data space to ensure the invariance of features. Among them, the causal dilation convolution mechanism is introduced to extract the causal features in the sample data sequence. Again, the CNN is fused with the LSTM to capture the time series features of the historical data. Among them, the time pattern attention mechanism is introduced to enhance the attention of the neural network to critical moments at different time steps. Train the neural network and automatically learn the sequence features, use the Adam algorithm to backpropagate the training error, update the model parameters layer by layer step by step, use the Softmax activation function to classify the signal features, complete the classification task of the multi-feature input sequence, and construct the CNN-LSTM prediction model, as Figure 3 shown.

[0193] Step 4: Set the optimization ranges of the learning rate, the number of neurons in the hidden layer, and the L2 regularization coefficient, which are [1e-6, 0.1], [5, 250], and [1e-8, 0.1] respectively. Set the initial number scale of the TTAO algorithm to 10 and the number of iterations to 30. The flow chart of the CNN-LSTM parameter optimization based on the TTAO algorithm is as Figure 4 shown. After the optimization by the TTAO algorithm, the optimal values of the three parameters are determined. After 30 optimizations, the iteration is paused. At this time, the fitness value at the second iteration is the smallest, and the iteration curve is as Figure 5 shown. Therefore, the learning rate, regularization coefficient, and the number of neurons in the hidden layer are selected as 0.0286, 0.0007, and 100 respectively. The parameter settings of the TTAO-CNN-LSTM neural network are shown in Table 1.

[0194] Table 1 Parameter settings of the TTAO-CNN-LSTM neural network

[0195]

[0196] Step 5: Set that the input layer mainly performs data preprocessing on the initial data and divides the training set and the test set. The hidden layer is a 5-layer CNN-LSTM neural network. The output layer performs inverse normalization on the data to obtain the final result predicted by the model. Compare the prediction effects of the TTAO-CNN-LSTM prediction model with the CNN prediction model, the LSTM prediction model, and the CNN-LSTM prediction model, as Figure 6 and Figure 7 shown. Compare the prediction result with the actual SCR outlet NOx concentration and calculate the relative error, as Figure 8 shown. Calculate the RMSE, MAPE, RPD, and R 2 of the four prediction models, verify the accuracy of the TTAO-CNN-LSTM prediction model and compare it with the CNN prediction model, the LSTM prediction model, and the CNN-LSTM prediction model, as Figure 9 shown,

[0197] As Figure 9 can be seen, the RMSE and MAPE of the designed TTAO-CNN-LSTM prediction model are the smallest, which are 0.4204 and 0.0082 respectively. The RPD reaches 2.2153, and R 2 is closest to 1. Compared with the CNN-LSTM model, the RMSE of this model is reduced by 0.1052 mg / m 3 , and the MAPE is reduced by 0.0021. The RPD is increased by 0.3941, and R 2 is increased by 0.1046. This proves that the prediction result of this model has a smaller deviation, higher accuracy, and better prediction effect on the SCR outlet NOx concentration.

[0198] In one embodiment, the method for predicting the NOx concentration at the outlet of the SCR denitration system in a thermal power plant further includes:

[0199] S2. Obtain multimodal record data on the prediction of the NOx concentration at the outlet of the SCR denitration system of the thermal power plant based on the TTAO-CNN-LSTM model within the first historical time period;

[0200] In S2, the multimodal record data includes: the current multiple SCR inlet factors of the SCR denitration system of the thermal power plant based on which the TTAO-CNN-LSTM model makes each prediction, the corresponding results of each prediction of the TTAO-CNN-LSTM model, the operating parameters of the TTAO-CNN-LSTM model, and the impacts on the SCR denitration system of the thermal power plant, etc.; the first historical time period can be within the past 30 days;

[0201] S3. Perform feature map representation on the multimodal record data to obtain data feature maps;

[0202] In S3, when performing feature map representation, a large number of data features of the multimodal record data are extracted, and these data features are paired according to the feature pairing rules to obtain multiple paired feature sets; the pairing rules include: Rule 1, pair the data features of the same feature type and store them in the same paired feature set; Rule 2, pair the data features that are suitable for simultaneous display for the expert group to make optimization decisions on the TTAO-CNN-LSTM model and store them in the same paired feature set; Rule 3, pair the data features with a time-dependent relationship. For example, the data at the same moment of different sensors (such as temperature, pressure, flow rate) can be divided by a time series window to ensure that the paired features are within the same time range; Rule 4, according to expert experience or physical principles, pair the features that are closely related in the actual physical process. For example, the flue gas flow rate and the reaction temperature in the SCR system may be a strong correlation and have a connection in the physical process; Rule 5, pair the data features that present dynamic change characteristics. For example, considering the impact of fuel quality changes on the SCR reaction, the temperature, flow rate, and fuel characteristics during the reaction process can be paired; then all the data features in the same paired feature set are set in the same blank area of the feature map template. After all settings are completed, the current feature map template is used as the data feature map;

[0203] S4. Interact with the expert group based on the data feature map;

[0204] In S4, interact with the expert group based on the data feature map; the expert group includes a large number of personnel dedicated to model training, the use of the SCR denitration system in thermal power plants, etc.; during the interaction with the expert group, the expert group can decide on the correction plan for optimizing the TTAO-CNN-LSTM model;

[0205] S5. When the expert group inputs the decision-making correction plan, optimize the TTAO-CNN-LSTM model based on the correction plan;

[0206] In S5, when the expert group decides on the correction plan, optimize the TTAO-CNN-LSTM model based on the correction plan, and the correction plan indicates how to optimize the TTAO-CNN-LSTM model;

[0207] S6. Based on the optimized TTAO-CNN-LSTM model, continue to predict the outlet NOx concentration of the SCR denitration system in the thermal power plant;

[0208] In S6, after the TTAO-CNN-LSTM model is optimized, continue to predict the outlet NOx concentration of the SCR denitration system in the thermal power plant based on it.

[0209] In the embodiments of the present invention, firstly, through the feature map representation of multi-modal recorded data, multi-source data inside and outside the system can be efficiently integrated, and at the same time, the time dependence and dynamic change characteristics in the physical process are fully considered, thereby significantly improving the accuracy of the data feature map representation. Secondly, the expert group interacts with the system based on the data feature map, enabling the system to dynamically optimize the model by combining the experience and knowledge of the experts, further enhancing the adaptability and accuracy of the prediction model. Through the two-way feedback mechanism that combines expert knowledge and machine learning optimization, the TTAO-CNN-LSTM model can exert powerful prediction capabilities in the complex and changing denitration system of thermal power plants and continuously adapt to changes under different operating conditions. This process provides strong support for the real-time adjustment of the SCR denitration system and pollutant emission control, significantly improving the environmental governance effect and operating efficiency.

[0210] In one embodiment, S4, interacting with the expert group based on the data feature map, includes:

[0211] S41. Display the data feature map to the expert group;

[0212] In S41, after the data feature map is displayed to the expert group, each person in the expert group can view the displayed data feature map;

[0213] S42. Obtain a first map area jointly viewed by the expert group for more than a first duration threshold in the data feature map displayed to the expert group and a second map area independently viewed by at least one person in the expert group for more than a second duration threshold within a second historical time period;

[0214] In S42, the second historical time period can be the time period when the expert group views the displayed data feature map within the most recent 10 minutes in history; Joint viewing means that more than a threshold proportion of the people in the expert group view simultaneously, where the number threshold can be 2 / 3; Independent viewing means that a person views independently by himself; The first duration threshold can be 3 minutes; The second duration threshold can be 1 minute; The fact that the expert group jointly views a first map area for more than the first duration threshold and at least one person in the expert group independently views a second map area for more than the second duration threshold in the data feature map displayed to the expert group indicates that the expert group may be mainly making decisions on how to optimize the TTAO-CNN-LSTM model based on the first map area and in cooperation with the second map area. If this possibility is confirmed, decision-making assistance and guidance can be started for the expert group. Therefore, it is necessary to confirm whether this possibility is confirmed; The second duration threshold is less than the first duration threshold;

[0215] S43. Perform feature description on the first map area, the second map area, and the correlation relationship between the first map area and the second map area to obtain a feature description vector;

[0216] In S43, when performing feature description, the features of the first map region, the second map region, and the association relationship between the first map region and the second map region are extracted, including at least: the feature pairing rule based on which data features are paired in the paired feature set corresponding to the first map region, the data feature type in the first map region, the feature pairing rule based on which data features are paired in the paired feature set corresponding to the second map region, the data feature type in the second map region, the type of the association relationship, etc.; then, the extracted features are represented in the form of a vector to obtain a feature description vector.

[0217] S44. Based on the feature description vector, query the trigger indication knowledge base to determine the trigger indication knowledge; where the trigger indication knowledge includes: the trigger degree and the trigger event.

[0218] In S44, there is trigger indication knowledge corresponding to different feature description vectors in the trigger indication knowledge base, which includes the trigger degree and the trigger event. The trigger degree represents the degree of likelihood that the expert group may be making a decision on how to optimize the TTAO-CNN-LSTM model mainly based on the first map region and in cooperation with the second map region. The trigger event represents the event that the expert group may be making a decision on how to optimize the TTAO-CNN-LSTM model mainly based on the first map region and in cooperation with the second map region. For example, if the feature represented as the feature description vector is that the data feature type of the first map region is the result graph of the TTAO-CNN-LSTM model analysis (showing the change trend of the model prediction error), and the data feature type of the second map region is the graph reflecting the change trend of sensor data (mainly related to some parameters in the system), it indicates that the expert group is very likely to be making an optimization decision. The corresponding trigger degree is 8, and the corresponding trigger event is to analyze the association relationship between sensor data and model prediction error.

[0219] S45. When the trigger degree exceeds the trigger degree threshold, generate a guiding region sequence corresponding to the logical sequence of the trigger event; the i-th guiding region in the guiding region sequence conforms to the region constraint conditions of the first i logics in the logical sequence.

[0220] In S45, the trigger threshold can be 5. When the trigger level exceeds the trigger threshold, it indicates that the expert group may be making a decision on how to optimize the TTAO-CNN-LSTM model mainly based on the first map area and in cooperation with the second map area; i takes an integer from 1 to N, where N is the total number of guiding areas in the guiding area sequence. Taking i = 3 as an example, the third guiding area in the guiding area sequence meets the area constraint conditions of the first 3 logics in the logical sequence. The logical sequence of the trigger event includes the sequence of logics to be followed for further executing the trigger event. For example, if the trigger event is to analyze the correlation between sensor data and model prediction error, then its logical sequence further includes collecting sensor data, analyzing the correlation between sensor data and model prediction error, and verifying the analysis result. The area constraint condition of the logic is that the guiding area needs to contain relevant data features for executing the logic. For example, if the logic is to further collect sensor data, the corresponding area constraint condition is that the area contains richer sensor data. Then the third guiding area meets the area constraint conditions of the first 3 logics, which enables the expert group to make up for the previous analysis logic when viewing the third guiding area, improving convenience and guiding accuracy.

[0221] S46. In the data feature map displayed to the expert, starting from the first midpoint of the first map area, sequentially connect the second midpoints of each guiding area in the guiding area sequence in the order of sequence positions.

[0222] In S46, each area has a midpoint, and they are sequentially connected to form a virtual connection line, which is not displayed to the expert group.

[0223] S47. When, within a future time period, the expert group jointly views a third map area that exceeds the third duration threshold in the data feature map displayed to the expert group and is passed through by a target segment on the virtual connection line that exceeds the length threshold, determine whether the target segment contains a second midpoint.

[0224] In S47, the future time period can be within the next 30 minutes; the third duration threshold can be 2 minutes; the third duration threshold is less than the first duration threshold and greater than the second duration threshold. The third map area is newly jointly viewed by the expert group. The length threshold can be, for example, 6 cm. The target segment is the local segment of the virtual connection line that passes through the third map area. Passing through the third map area means that the target segment falls within the third map area.

[0225] S48. When it is yes, determine the order value of the guiding area containing the second midpoint on the target segment in the guiding area sequence.

[0226] In S48, when the second midpoint is included in the target segment, it indicates that the expert group can be assisted and guided at this time. Determine the sequence value of the guiding area of the second midpoint included in the target segment in the guiding area sequence. The sequence value is the number corresponding to the position of the guiding area in the guiding area sequence from the front to the back;

[0227] S49. When the difference between the sequence value exceeding the historical sequence value is less than the difference threshold, guide the expert group to jointly view the guiding area of the second midpoint included in the target segment.

[0228] In S49, when it has been confirmed historically that the expert group may be mainly making decisions on how to optimize the TTAO-CNN-LSTM model based on the first graph area and in cooperation with the second graph area, the historical sequence value is taken as the sequence value of the guiding area of the second midpoint included in the target segment determined when it was last confirmed that the expert group may be mainly making decisions on how to optimize the TTAO-CNN-LSTM model based on the first graph area and in cooperation with the second graph area. If it has not been confirmed, the historical sequence value is taken as 0; the difference threshold can be 3; when the difference between the sequence value exceeding the historical sequence value is less than the difference threshold, it indicates that guiding the expert group to jointly view the guiding area of the second midpoint included in the target segment will not cause excessive logical disconnection.

[0229] By precisely analyzing the interaction behavior of the expert group during the process of viewing data feature graphs, the embodiments of the present invention can intelligently guide the decision-making of the expert group, thereby optimizing the TTAO-CNN-LSTM model. Specifically, first, by tracking the behavior of the expert group when jointly and independently viewing graph areas, possible decision-making situations are identified, etc.; then, through feature description and querying the trigger indication knowledge base, the efficient triggering of the expert decision-making process is realized, improving the accuracy and convenience of decision-making; through the generation of the guiding area sequence and the application of virtual connection lines, the system can dynamically guide the expert group to further analyze according to their behavior, ensuring the logical coherence of the decision-making process; by analyzing the duration threshold of expert behavior and regional association, the expert is precisely guided to make decisions in the most relevant areas, thereby improving the accuracy and decision-making efficiency of the decision-making assistance system, and can dynamically adjust the guiding strategy to avoid logical disconnection, etc., significantly improving the intelligent level of decision-making guidance.

[0230] The embodiments of the present invention provide a system for predicting the NOx concentration at the outlet of the SCR denitration system of a thermal power plant, including:

[0231] A concentration prediction module for predicting the NOx concentration at the outlet of the SCR denitration system of a thermal power plant based on a pre-established TTAO-CNN-LSTM model;

[0232] Among them, the steps for pre-establishing the TTAO-CNN-LSTM model are as follows:

[0233] Collect the historical NOx concentration of the SCR denitration system in a thermal power plant and multiple historical SCR inlet factors;

[0234] Preprocess the historical NOx concentration and multiple historical SCR inlet factors respectively;

[0235] Perform dimensionality reduction processing on the preprocessed multiple historical SCR inlet factors for model input;

[0236] Construct a CNN-LSTM prediction model based on the preprocessed multiple historical SCR inlet factors after dimensionality reduction processing for model input and the preprocessed historical NOx concentration;

[0237] Set the optimization range of the neural network hyperparameters, use the triangular topology aggregation algorithm for optimization, and automatically update the optimal values of the hyperparameters into the CNN-LSTM prediction model to obtain the TTAO-CNN-LSTM model.

[0238] Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalent technologies, the present invention is also intended to include these changes and modifications.

Claims

1. A method for predicting NOx concentration at the outlet of an SCR denitration system in a thermal power plant, characterized in that: include: Based on the pre-established TTAO-CNN-LSTM model, the outlet NOx concentration of the SCR denitrification system of the thermal power plant is predicted; Among them, the pre-establishment steps of the TTAO-CNN-LSTM model are as follows: Collect historical NOx concentrations and historical multiple SCR inlet factors of the SCR denitrification system of thermal power plants; Pre-process the historical NOx concentration and multiple historical SCR inlet factors respectively; Perform model input dimension reduction on the pre-processed historical multiple SCR entry factors; Based on the pre-processed historical multiple SCR inlet factors and pre-processed historical NOx concentration after dimensionality reduction of model input, a CNN-LSTM prediction model was constructed; The optimization range of the neural network hyperparameters is set, the triangular topology aggregation algorithm is used for optimization, and the optimal values ​​of the hyperparameters are automatically updated to the CNN-LSTM prediction model to obtain the TTAO-CNN-LSTM model.

2. The method for predicting NOx concentration at the outlet of the SCR denitration system of a thermal power plant according to claim 1, characterized in that: The historical multiple SCR inlet factors include at least: ammonia injection amount, SCR inlet NOx concentration, SCR inlet O2 concentration, SCR inlet temperature, SCR inlet pressure, SCR inlet SO2 concentration, SCR inlet flue gas flow rate and SCR inlet flue gas flow velocity.

3. The method for predicting NOx concentration at the outlet of the SCR denitration system of a thermal power plant according to claim 1, characterized in that: The pre-processing of the historical NOx concentration and the historical multiple SCR inlet factors respectively includes: The Gaussian weighted moving average filtering algorithm is used to filter and denoise the historical NOx concentration and multiple historical SCR inlet factors; The K nearest neighbor interpolation algorithm is used to interpolate and replace abnormal data points of historical NOx concentration and multiple historical SCR inlet factors; The Z-score standard is used to standardize the historical NOx concentration and historical multiple SCR inlet factors.

4. The method for predicting NOx concentration at the outlet of the SCR denitration system of a thermal power plant according to claim 1, characterized in that: The model input dimension reduction processing is performed on the pre-processed historical multiple SCR entry factors, including: The principal component analysis method and Pearson correlation coefficient algorithm are used to reduce the model input dimension of multiple historical SCR entry factors after preprocessing.

5. The method for predicting NOx concentration at the outlet of the SCR denitration system of a thermal power plant according to claim 1, characterized in that: The CNN-LSTM prediction model is constructed based on the pre-processed historical multiple SCR inlet factors and the pre-processed historical NOx concentration after the model input dimension reduction processing, including: The model inputs the pre-processed historical multiple SCR inlet factors after dimensionality reduction and the pre-processed historical NOx concentration as historical operation data; Select SCR inlet NOx concentration, SCR inlet O2 content, SCR inlet ammonia flow, SCR inlet temperature and SCR inlet flue gas pressure as the input of the input layer, and select SCR outlet NOx concentration as the output of the output layer; Set up CNN convolutional layers to filter historical running data and extract local features, sharing weights and biases; Set up a causal dilated convolution mechanism to extract the causal features of the output of the CNN convolutional layer; Set up a CNN pooling layer to downsample the output of the causal dilated convolution mechanism; Set the temporal mode attention mechanism to process the output of the CNN pooling layer; Set up the LSTM layer to process the output of the temporal pattern attention mechanism; The input layer, CNN convolution layer, causal dilated convolution mechanism, CNN pooling layer, temporal pattern attention mechanism, LSTM layer and output layer are used as the models to be trained; The model to be trained is trained based on historical operating data. During training, the Adam algorithm is used to back-propagate the training error, and the model parameters are gradually updated layer by layer. The Softmax activation function is used to classify the signal features, complete the classification task of multi-feature input sequences, and finally obtain the CNN-LSTM prediction model.

6. The method for predicting NOx concentration at the outlet of the SCR denitration system of a thermal power plant according to claim 1, characterized in that: The neural network hyperparameters include at least: learning rate, number of hidden layer neurons and regularization coefficient.

7. The method for predicting NOx concentration at the outlet of the SCR denitration system of a thermal power plant according to claim 6, characterized in that: The optimization range of setting the neural network hyperparameters includes: The optimization ranges of the learning rate, the number of hidden layer neurons, and the regularization coefficient are set to [0.1, 1e-6], [5, 250], and [0.1, 1e-8] respectively.

8. The method for predicting NOx concentration at the outlet of the SCR denitration system of a thermal power plant according to claim 1, characterized in that: The pre-building steps of the TTAO-CNN-LSTM model also include: The test data is used to verify the prediction effect of the TTAO-CNN-LSTM prediction model.

9. The method for predicting NOx concentration at the outlet of the SCR denitration system of a thermal power plant according to claim 8, characterized in that: The test data is used to verify the prediction effect of the TTAO-CNN-LSTM prediction model, including: Using RMSE, MAPE, RPD and R 2 The prediction effect of the TTAO-CNN-LSTM prediction model is verified based on the test data.

10. A NOx concentration prediction system at the outlet of the SCR denitration system of a thermal power plant, characterized in that: include: The concentration prediction module is used to predict the outlet NOx concentration of the SCR denitrification system of a thermal power plant based on the pre-established TTAO-CNN-LSTM model; Among them, the pre-establishment steps of the TTAO-CNN-LSTM model are as follows: Collect historical NOx concentrations and historical multiple SCR inlet factors of the SCR denitrification system of thermal power plants; Pre-process the historical NOx concentration and multiple historical SCR inlet factors respectively; Perform model input dimension reduction on the pre-processed historical multiple SCR entry factors; Based on the pre-processed historical multiple SCR inlet factors and pre-processed historical NOx concentration after dimensionality reduction of model input, a CNN-LSTM prediction model was constructed; The optimization range of the neural network hyperparameters is set, the triangular topology aggregation algorithm is used for optimization, and the optimal values ​​of the hyperparameters are automatically updated to the CNN-LSTM prediction model to obtain the TTAO-CNN-LSTM model.

Images