Flue gas desulfurization slurry pH value prediction method based on time convolution gated recurrent neural network

Abstract

The invention discloses a flue gas desulfurization slurry pH value prediction method based on a time convolution gated recurrent neural network, and the method comprises the steps: analyzing a flue gas desulfurization process, and combining a large amount of process monitoring data, a maximum mutual information analysis method is adopted to carry out correlation analysis on process variables such as flue gas pressure, inlet flue gas temperature, flue gas oxygen content, flue gas fly ash content, inlet flue gas sulfur dioxide concentration, inlet flue gas flow and the like, and key variables influencing the pH value of slurry are obtained. And then, combining the time convolution neural network with the gating circulation unit to form a time convolution gating circulation network. And finally, completing the prediction of the pH value of the slurry in the desulfurization process by using the selected key variables and the convolution gating circulation network. The pH value is monitored and adjusted in real time, shutdown can be prevented, and system reliability and economic benefits are improved. Therefore, accurate prediction of the pH value of the slurry has important significance on stability, efficiency and environmental protection performance of the flue gas desulfurization process.

Description

Technical Field

[0001] The invention belongs to the field of data processing and is a method for predicting the pH value of flue gas desulfurization slurry based on a time convolution gated recurrent neural network. Background Art

[0002] Sulfur dioxide in coal-fired flue gas is one of the main pollutants, and its hazards cannot be ignored. Sulfur dioxide can generate sulfuric acid and sulfates through a series of chemical reactions in the atmosphere, leading to the formation of acid rain. Acid rain can seriously damage the acid-base balance of soil and water bodies, damage agricultural and aquatic ecosystems, and corrode buildings and infrastructure. In addition, sulfur dioxide also has significant harm to human health. Long-term exposure to low concentrations of sulfur dioxide may cause chronic respiratory diseases such as asthma and chronic bronchitis; short-term high-concentration exposure may cause acute dyspnea and decreased lung function. Therefore, controlling sulfur dioxide emissions has become one of the focuses of environmental protection policy and technology research. Desulfurization technology, such as wet flue gas desulfurization technology, can efficiently remove sulfur dioxide from flue gas, thereby reducing its harm to the environment and health, and has important practical significance and research value.

[0003] Traditional desulfurization process modeling methods have significant limitations in capturing the complex dynamic characteristics of the process. These methods usually rely on a series of simplified assumptions and linear approximations, mainly using static equilibrium models and empirical models. Although these models can reflect the basic characteristics of the desulfurization process to a certain extent, they are unable to deal with multiphase reactions, nonlinear mass transfer and heat transfer processes, and time-varying operating conditions, and it is difficult to provide accurate descriptions and predictions. First, the desulfurization process involves complex multiphase reactions. For example, in a wet flue gas desulfurization (FGD) system, limestone slurry reacts chemically with sulfur dioxide (SO2) in the flue gas to produce calcium sulfite and calcium sulfate. This process not only includes mass transfer between the gas phase and the liquid phase, but also involves the reaction kinetics between the liquid phase and the solid phase. However, traditional models often simplify these complex multiphase reactions into single-phase reactions, ignoring the changes in mass transfer resistance and reaction rate between the phases. Although this simplified treatment can reduce the complexity of the calculation, it cannot accurately reflect the interactions and dynamic changes between the phases in the actual process, resulting in a significant decrease in model accuracy. Secondly, the mass transfer and heat transfer processes in the desulfurization process have significant nonlinear characteristics. Taking wet flue gas desulfurization as an example, SO2 in the flue gas enters the slurry through the gas-liquid interface, and then reacts with the alkaline substances in the slurry to generate calcium sulfate. In this process, the mass transfer rate at the gas-liquid interface is affected by many factors, such as flue gas flow rate, slurry flow state, interface area, etc. The changes in these factors will cause the mass transfer process to present strong nonlinear characteristics. However, traditional models often use linear approximation methods, ignoring the influence of these nonlinear factors, and thus cannot accurately describe the actual situation of mass transfer and heat transfer processes. This simplification not only affects the prediction accuracy of the model, but also limits the adaptability of the model under different operating conditions. In addition, the operating conditions and environmental parameters in the desulfurization process often change. For example, load changes in coal-fired power plants can cause fluctuations in flue gas flow and composition, and changes in ambient temperature and pressure can also affect the efficiency of the reaction process and product generation. Traditional static equilibrium models and empirical models lack the ability to respond to these dynamic changes, making it difficult to adjust and respond to changes in operating conditions in real time, resulting in poor prediction results of the model in practical applications. Especially under high load changes or complex operating conditions, the limitations of traditional models are more prominent, and they cannot provide effective control strategies and optimization solutions.

[0004] In summary, the traditional desulfurization process modeling method has significant limitations in capturing complex dynamic characteristics and cannot meet the needs of efficient desulfurization process optimization and precise control. The flue gas desulfurization slurry pH prediction method based on time convolution gated recurrent neural network realizes the effective combination of time convolution neural network and gated recurrent unit. Among them, TCN can effectively extract long-term dependencies and time series features in sequence data, process sequence data through one-dimensional convolution operation, and capture information at different time steps. GRU, as an improved recurrent neural network variant, selectively remembers or forgets historical information through the gate mechanism, thereby reducing the gradient vanishing problem when processing long sequence data and improving the efficiency of sequence modeling. In addition, the residual structure introduces jump connections, which enables the gradient to propagate more smoothly, helps to train deeper neural networks, avoids the problem of gradient vanishing or explosion, and thus improves the training effect and performance of the model. The combination of TCN's parallel processing capability and GRU's memory capability makes the TCN-GRU model perform well in time series data processing, can efficiently capture complex time series patterns, and provide more accurate slurry pH prediction results. Summary of the invention

[0005] The present invention proposes a method for predicting the pH value of flue gas desulfurization slurry based on a time convolutional gated recurrent neural network. The method uses actual data of the flue gas desulfurization process in a coal-fired power plant as input, and selects process variables that have a greater impact on the pH value of the slurry through a mutual information analysis method. A time convolutional neural network and a gated recurrent unit are used to form a time convolutional gated recurrent neural network, wherein the time convolutional network can effectively extract long-term dependent time series features through a one-dimensional convolution operation, and the gated recurrent unit can efficiently selectively memorize and forget information through a gate mechanism, thereby enhancing the sequence modeling capability. The introduction of the residual structure alleviates the gradient vanishing problem and ensures the stable training of the deep network. Overall, the TCN-GRU model performs well in processing long time series data, can capture complex time series patterns, and provide more accurate predictions of the slurry pH value.

[0006] For the collected sulfur dioxide concentration sequence of flue gas, the variational mode decomposition method is used to decompose the time series into multiple internal mode functions. The original complex signal contains multiple frequency components. Directly modeling these complex signals may require considering the interaction of multiple frequencies. The decomposed sub-time series corresponds to a specific frequency range or scale, which can effectively simplify the processing of frequency or scale in the modeling process. Then, the feedforward cascade neural network, long short-term memory network and autoregressive integral moving average model are used to form a basic learner, and the multilayer perceptron is selected as the meta-learner. The two types of learners are combined according to the hierarchical relationship to complete the construction of the ensemble model. The ensemble model is used to learn the dynamic characteristics of each internal mode function separately, and the prediction sequence under the function is obtained. By reorganizing the prediction sequences of all integrated models, the final prediction sequence of the SO2 concentration of the flue gas is obtained.

[0007] A method for predicting the pH value of flue gas desulfurization slurry based on a time convolution gated recurrent neural network, characterized in that it comprises the following steps:

[0008] Step 1: During the power generation process of a coal-fired power plant, the process monitoring system conducts real-time monitoring and data collection on the entire process of flue gas desulfurization.

[0009] Step 2: Based on the actual situation on the field and theoretical knowledge, select candidate variables that can be used to characterize the variables to be predicted. The mutual information index is used to evaluate the degree of correlation between any two candidate variables. First, select variables with high mutual information scores with the predicted variables, which can be regarded as variables that are highly correlated with the variables to be predicted. For the selected variables, look for variables with similar mutual information scores, which indicates that there is a close relationship between them. In this case, some of them are redundant variables and should be eliminated to reduce the computational complexity.

[0010] Step 3: Build a time convolutional gated recurrent neural network and use the measured values ​​of the selected variables and the measured values ​​of the slurry pH as training data to complete the training of the slurry pH prediction model for the desulfurization process.

[0011] Step 4: Evaluate the established model using performance metrics such as root mean square error, coefficient of determination, and mean absolute percentage error.

[0012] The flue gas desulfurization slurry pH prediction method based on the time convolution gated recurrent neural network of the embodiment of the present invention takes the desulfurization process data of a coal-fired thermal power plant as the research object, performs correlation analysis on process variables such as flue gas pressure, inlet flue gas temperature, flue gas oxygen content, flue gas fly ash content, inlet flue gas sulfur dioxide concentration, and inlet flue gas flow, and selects key variables affecting the slurry pH value. Then, the time convolution neural network and the gated recurrent unit are combined to form a time convolution gated recurrent network. Finally, the selected key variables and the convolution gated recurrent network are used to complete the prediction of the slurry pH value of the desulfurization process. The slurry pH value directly affects the desulfurization reaction rate and efficiency. Deviating from the optimal range will lead to a decrease in the desulfurization effect, and even cause equipment corrosion and unstable operation. By accurately predicting the slurry pH value, on the one hand, the use of desulfurizers can be optimized and the operating cost can be reduced; on the other hand, the operating parameters of the desulfurization system can be optimized and the desulfurization efficiency can be improved, which is of great significance for atmospheric environmental protection.

[0013] In addition, the flue gas desulfurization slurry pH value prediction method based on time convolution gated recurrent neural network according to this embodiment also has the following additional technical features:

[0014] In step 2, the correlation between process variables and production capacity indicators needs to be analyzed based on the actual desulfurization process data using the mutual information coefficient analysis method to obtain the best predictor of slurry pH value. This method measures the degree of mutual dependence between the two variables by calculating the mutual information between the two variables. M is used to represent the random variable, and the calculation expression of M corresponding to the entropy is:

[0015]

[0016] Where p(m) is the probability density function corresponding to M. Assuming N is a known variable, the conditional entropy defined by M and N can be expressed as

[0017]

[0018] In the formula, p(m,n) and p(m|n) represent the joint probability and posterior probability defined by M and N respectively. Then, the calculation expression of mutual information I(M,N) is

[0019]

[0020] The main steps of using mutual information analysis method to analyze the correlation between variables are as follows:

[0021] (1) Before performing feature selection for the desulfurization process, data preprocessing is required. This mainly includes: 1) data collection and organization to ensure data integrity and quality; 2) processing missing values, filling or deleting missing values ​​in the data; 3) standardizing or normalizing data: scaling the data to the same scale.

[0022] (2) Based on the desulfurization process data, the mutual information between each characteristic variable and the target variable, as well as the mutual information values ​​between the characteristic variables, are calculated.

[0023] (3) Analyze the calculated mutual information value and complete the selection of feature variables.

[0024] In step 3, the sulfur dioxide concentration of the flue gas is predicted based on the temporal convolutional neural network. The temporal convolutional gated recurrent neural network combines the long-term dependency capture capability of the temporal convolutional network and the efficient sequence processing capability of the gated recurrent unit, and adopts a residual structure to enhance the stability and performance of the model. The temporal convolutional network achieves parallel processing by extending convolution to improve computational efficiency, while the gated recurrent unit effectively processes sequence data through a gating mechanism. The residual structure solves the gradient problem in deep networks, making information easier to transfer and enhancing the expressiveness and generalization ability of the model in complex time series tasks.

[0025] For univariate time series The output h(k) of the temporal convolutional neural network at time k can be expressed as

[0026]

[0027] Where w is the filter, F and d represent the filter size and dilation factor respectively.

[0028] The Gated Recurrent Unit (GRU) is an improved recurrent neural network structure that aims to solve the gradient vanishing and gradient exploding problems of traditional recurrent neural networks when processing long sequence data. GRU dynamically adjusts the information flow of hidden states by introducing update gates and reset gates, thereby more effectively capturing long-term dependencies in the sequence. Compared with long short-term memory networks, GRU has a simpler structure with only two gating units, which reduces computational complexity and speeds up training. GRU has excellent performance in sequence prediction, speech recognition, and natural language processing, and has become an important tool in time series data modeling.

[0029] d k =σ(W d [x k ,h k-1 ]+b d ) (14)

[0030] s k =σ(W s [x k ,h k-1 ]+b s ) (15)

[0031]

[0032] Among them, x is the external input of the network, and h represent the intermediate hidden layer and hidden layer state respectively. d ,W s and represents the weight matrix of each gate control unit in turn. The outputs of the reset gate and update gate are denoted as s and d, b respectively. * (* indicates d, s and ) represents the bias vector and ⊙ represents element-by-element vector multiplication.

[0033] The data processing process of the temporal convolutional gated recurrent neural network is mainly divided into the following steps:

[0034] (1) Data acquisition: A large number of different types of sensors are deployed in the flue gas desulfurization process, which can be used to record the measured values ​​of various process variables in real time. These measured values ​​are then transmitted to the built-in management information system. Accessing this system can obtain historical data of the corresponding variables at any time.

[0035] (2) Model input variable selection: Based on the actual situation on site and theoretical knowledge, candidate variables that can be used to characterize the variables to be predicted are selected.

[0036] (3) Feature selection, using the mutual information metric to evaluate the correlation between any two variables. First, select variables with high mutual information scores with the predicted variable, which can be regarded as variables that are highly correlated with the predicted variable. For the selected variables, look for variables with similar mutual information scores, which indicates that there is a close relationship between them. In this case, some of them are redundant variables and should be eliminated to reduce computational complexity.

[0037] (4) Data preprocessing: preprocessing the measured values ​​of the selected variables, including outlier detection, missing data filling and data standardization. Data preprocessing can effectively improve data quality and consistency, and enhance the accuracy and stability of the established model.

[0038] (5) Model building: Using the preprocessed data obtained in the previous step, complete the training of the temporal convolutional gated recurrent neural network.

[0039] In step 4, the correlation coefficient (R 2), mean absolute percentage error (MAPE) and root mean square error (RMSE) are used to test the generalization performance of the proposed prediction method. The calculation expressions of the three performance indicators are:

[0040]

[0041] in y(i) and are the true value and predicted value at time i, respectively, and n is the total number of samples. The correlation coefficient ranges from 0 to 1, R 2 The higher the value, the stronger the model's ability to fit the data. MAPE and RMSE are both statistical indicators that measure the difference between the model's predicted value and the actual value. The smaller the value of the two indicators, the higher the prediction accuracy of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 It is a schematic diagram of mutual information feature selection.

[0043] Figure 2 This is a diagram of the gated recurrent unit structure.

[0044] Figure 3 It is a schematic diagram of temporal convolution operation.

[0045] Figure 4 Diagram of the temporal convolutional gated recurrent neural network structure. DETAILED DESCRIPTION

[0046] The embodiments of the present invention are described in detail below. This embodiment is implemented on the premise of the technical solution of the present invention, and a detailed implementation method and a specific operation process are given, but the protection scope of the present invention is not limited to the following embodiments.

[0047] The technical solution adopted by the present invention is a method for predicting the pH value of flue gas desulfurization slurry based on a time convolution gated recurrent neural network, and the method comprises the following steps:

[0048] Step 1: In a coal-fired power plant, the distributed control system performs real-time monitoring and data collection during the flue gas desulfurization process.

[0049] Step 2: Use the mutual information feature analysis method to analyze the correlation between various process variables and obtain the key variables that affect the pH value of the slurry.

[0050] Step 3: Combine the temporal convolutional neural network and the gated recurrent unit to form a temporal convolutional gated recurrent network. This combined model has the advantages of both models and has a strong ability to capture process dynamics.

[0051] Step 4: Use the selected key variables and convolutional gated recurrent network to complete the prediction of the slurry pH value in the desulfurization process, and use different performance indicators to comprehensively evaluate the prediction results.

[0052] According to the actual slurry pH value, data is collected to extract the key variables that affect the slurry pH value. The correlation between variables is analyzed using the mutual information feature analysis method, which measures the degree of mutual dependence between the two variables by calculating the mutual information between the two variables. M is used to represent the random variable, and the calculation expression of M corresponding to the entropy is

[0053]

[0054] Where p(m) is the probability density function corresponding to M. Assuming N is a known variable, the conditional entropy defined by M and N can be expressed as

[0055]

[0056] In the formula, p(m,n) and p(m|n) represent the joint probability and posterior probability defined by M and N respectively. Then, the calculation expression of mutual information I(M,N) is

[0057]

[0058] The main steps of using mutual information analysis method to analyze the correlation between variables are as follows:

[0059] Step 1: Before performing feature selection for the desulfurization process, data preprocessing is required. This mainly includes: (1) data collection and organization to ensure data integrity and quality; (2) processing missing values, filling or deleting missing values ​​in the data; (3) standardizing or normalizing data: scaling the data to the same scale.

[0060] Step 2: Based on the desulfurization process data, calculate the mutual information between each characteristic variable and the target variable, as well as the mutual information values ​​between the characteristic variables.

[0061] Step 3: Analyze the calculated mutual information value and complete the selection of feature variables.

[0062] The temporal convolutional neural network is combined with the gated recurrent unit to construct a combined deep learning model with the ability to characterize complex dynamic characteristics to complete the prediction of the slurry pH value in the desulfurization process.

[0063] Temporal Convolutional Networks (TCN) are a class of deep learning models specifically designed to process time series data. The network uses one-dimensional convolution operations to replace traditional recurrent neural networks, showing significant advantages in capturing long-term dependencies and parallel computing capabilities. One of the core features of TCN is causal convolution, which ensures that the output depends only on the current and previous inputs, thus conforming to the natural order of the time series. In addition, TCN introduces dilated convolutions, which introduce intervals between convolution kernels to enable the receptive field to expand rapidly, thereby capturing dependencies over a longer time range without significantly increasing the amount of computation. Another key advantage is the application of residual connections, which allows the model to directly pass information in each layer, effectively alleviating the common gradient vanishing and gradient exploding problems in deep network training, and ensuring stable propagation of gradients. Compared with recurrent neural networks and long short-term memory networks, the parallel computing capabilities of TCN significantly improve the training speed and efficiency of the model. TCN also has the ability to process variable-length input sequences without the need for fixed-length preprocessing of the input, which makes it more flexible and efficient in many practical application scenarios. By learning features directly from raw time series data, TCN achieves end-to-end learning without the need for complex feature engineering.

[0064] For univariate time series The output h(k) of the temporal convolutional neural network at time k can be expressed as

[0065]

[0066] Where w is the filter, F and d represent the filter size and dilation factor respectively.

[0067] The Gated Recurrent Unit (GRU) is an improved recurrent neural network structure that aims to solve the gradient vanishing and gradient exploding problems of traditional recurrent neural networks when processing long sequence data. GRU dynamically adjusts the information flow of hidden states by introducing update gates and reset gates, thereby more effectively capturing long-term dependencies in the sequence. Compared with long short-term memory networks, GRU has a simpler structure with only two gating units, which reduces computational complexity and speeds up training. GRU has excellent performance in sequence prediction, speech recognition, and natural language processing, and has become an important tool in time series data modeling.

[0068] d k =σ(W d [x k ,h k-1 ]+b d ) (5)

[0069] s k =σ(W s [x k ,h k-1 ]+b s ) (6)

[0070]

[0071] Among them, x is the external input of the network, and h represent the intermediate hidden layer and hidden layer state respectively. d ,W s and represents the weight matrix of each gate control unit in turn. The outputs of the reset gate and update gate are denoted as s and d, b respectively. * (* indicates d, s and ) represents the bias vector and ⊙ represents element-by-element vector multiplication.

[0072] The temporal convolutional gated recurrent neural network consists of three parts: input layer, temporal convolution layer and gated recurrent unit layer. The input layer completes the selection and preprocessing of input variables. Specifically, candidate features are evaluated and screened based on mutual information measurement, and then preprocessing operations such as missing value filling and outlier removal are performed on the data corresponding to the selected features. The temporal convolution layer consists of a dilated convolution layer, a weight normalization layer, a dropout layer and a rectified linear unit. The dilated convolution layer is responsible for extracting the dynamic features of the input sequence; the weight normalization layer accelerates the training process by re-parameterizing the weights of the neural network, improves the convergence speed, and stabilizes the training process; the role of the dropout layer is to prevent overfitting by randomly ignoring a part of the neurons during the training process, thereby improving the generalization ability of the model. The role of the rectified linear unit is to introduce nonlinear characteristics, so that the model can learn complex patterns and accelerate the training process through simple and efficient calculations. The output of the temporal convolution layer is the input of the gated recurrent unit layer, which learns the long-term and short-term relationships in the extracted feature sequence, further improving the model's ability to capture complex nonlinear dynamics in the time series. In the temporal convolutional gated recurrent neural network, the temporal convolution layer has a powerful ability to extract temporal features. Combined with the superiority of the gated recurrent unit layer in long-term information memory, the combined model can achieve accurate identification of systems with severe nonlinear dynamics and large time delays.

[0073] The data processing process of the temporal convolutional gated recurrent neural network is mainly divided into the following steps:

[0074] Step 1: Data acquisition, the flue gas desulfurization process is deployed with a large number of different types of sensors, which can be used to record the measured values ​​of various process variables in real time, which are then transmitted to the built-in management information system. Accessing the system can obtain historical data of the corresponding variables at any time.

[0075] Step 2: Model input variable selection: Based on the actual situation on site and theoretical knowledge, select candidate variables that can be used to characterize the variables to be predicted.

[0076] Step 3: Feature selection, using the mutual information metric to evaluate the degree of correlation between any two variables. First, select variables with high mutual information scores with the predicted variable, which can be considered as variables that are highly correlated with the predicted variable. For the selected variables, look for variables with similar mutual information scores, which indicates that there is a close relationship between them. In this case, some of them are redundant variables, so they should be eliminated to reduce computational complexity.

[0077] Step 4: Data preprocessing, preprocessing the measured values ​​of the selected variables, including outlier detection, missing data filling and data standardization. Data preprocessing can effectively improve data quality and consistency, and enhance the accuracy and stability of the established model.

[0078] Step 5: Model building, using the preprocessed data obtained in the previous step, complete the training of the temporal convolutional gated recurrent neural network.

[0079] Step 6: Model validation: Use root mean square error (RMSE), R square (R 2 ) and mean absolute percentage error (MAPE) to evaluate the established models.

[0080] The combined model includes a residual structure. The residual structure is a design method used to build deeper neural networks in deep learning. The core idea is to introduce a skip connection to add the input directly to the output to form a residual block. The formula of the residual block is as follows:

[0081] y=f(x,{W i})+x (9)

[0082] Among them, x represents the input of the residual block, f(x,{W i}) represents a series of transformations performed on the module input, which is usually a complex function formed by stacking layers such as convolutional layers, activation functions, and batch normalization. i} is the learnable weight parameter in the transformation, and y represents the output of the residual block.

[0083] The importance of residual structure for deep learning prediction lies in that it solves the key challenges in deep neural network training. First, the residual structure directly transmits inputs through jump connections, effectively alleviating the gradient vanishing problem, allowing the network to transfer gradients more stably, thereby supporting deeper network architectures. Second, the residual structure simplifies the optimization process of deep networks, making the model easier to train and able to converge in a shorter time. This improvement not only improves the prediction accuracy and generalization ability of the model, but also significantly enhances its performance in complex tasks.

[0084] Example

[0085] Step 1: The distributed control system deployed in the flue gas desulfurization system completes real-time monitoring and data collection of the pH value of the limestone slurry in the reaction tank.

[0086] Step 2: Use the mutual information feature analysis method to analyze the correlation between various process variables and obtain the key variables that affect the pH value of the slurry.

[0087] Step 3: Build a time convolutional gated recurrent neural network and use the measured values ​​of the selected variables and the measured values ​​of the slurry pH as training data to complete the training of the slurry pH prediction model for the desulfurization process.

[0088] Step 4: Evaluate the established model using performance metrics such as root mean square error, coefficient of determination, and mean absolute percentage error.

Claims

1. A method for predicting the pH value of flue gas desulfurization slurry based on time convolution gated recurrent neural network, characterized in that: The method comprises the following steps: Step 1: In a coal-fired power plant, the distributed control system performs real-time monitoring and data collection during the flue gas desulfurization process; Step 2: Use mutual information feature analysis method to analyze the correlation between various process variables and obtain the key variables affecting the pH value of the slurry; Step 3: Combine the temporal convolutional neural network and the gated recurrent unit to form a temporal convolutional gated recurrent network; Step 4: Use the selected key variables and convolutional gated recurrent network to complete the prediction of the slurry pH value in the desulfurization process, and use different performance indicators to comprehensively evaluate the prediction results.

2. The method for predicting pH value of flue gas desulfurization slurry based on time convolution gated recurrent neural network according to claim 1 is characterized in that: According to the actual slurry pH value, data is collected to extract the key variables that affect the slurry pH value; the correlation between the variables is analyzed using the mutual information feature analysis method, which measures the degree of mutual dependence between the two variables by calculating the mutual information between the two variables; Use M to represent a random variable, and the calculation expression of entropy corresponding to M is Where p(m) is the probability density function corresponding to M, and N is a known variable. Then the conditional entropy defined by M and N is expressed as In the formula, p(m,n) and p(m|n) represent the joint probability and posterior probability defined by M and N respectively; then, the calculation expression of mutual information I(M,N) is The steps of analyzing the correlation between variables using the mutual information analysis method are as follows: Step 1: Before performing feature selection for the desulfurization process, data preprocessing is required; this mainly includes: (1) data collection and organization to ensure data integrity and quality; (2) processing missing values, filling or deleting missing values ​​in the data; (3) Standardize or normalize data: scale the data to the same scale; Step 2: Based on the desulfurization process data, calculate the mutual information between each characteristic variable and the target variable, as well as the mutual information value between the characteristic variables; Step 3: Analyze the calculated mutual information value and complete the selection of feature variables.

3. The method for predicting the pH value of flue gas desulfurization slurry based on time convolutional gated recurrent neural network according to claim 1 is characterized in that: Combine the time convolutional neural network with the gated recurrent unit to build a combined deep learning model with the ability to characterize complex dynamic characteristics and complete the prediction of the slurry pH value in the desulfurization process; In the time convolution neural network TCN, for univariate time series The output h(k) of the temporal convolutional neural network at time k can be expressed as Where w is the filter, F and d represent the filter size and dilation factor respectively; In the gated recurrent unit GRU, d k =σ(W d [x k ,h k-1 ]+b d ) (5) s k =σ(W s [x k ,h k-1 ]+b s ) (6) Among them, x is the external input of the network, and h represent the intermediate hidden layer and hidden layer state respectively; W d ,W s and represents the weight matrix of each gate control unit in turn; the outputs of the reset gate and update gate are represented as s and d, b respectively * (* indicates d, s and ) represents the bias vector, ⊙ represents element-by-element vector multiplication; The data processing process of the time convolution gated recurrent neural network is divided into the following steps: Step 1: Data acquisition. Different types of sensors are deployed in the flue gas desulfurization process to record the measured values ​​of various process variables in real time. The measured values ​​are then transmitted to the built-in management information system; the management information system is accessed to obtain the historical data of the corresponding variables; Step 2: Model input variable selection: based on the actual situation on site and theoretical knowledge, select candidate variables to characterize the variables to be predicted; Step 3: Feature selection, using the mutual information index to evaluate the correlation between any two variables; first select variables with high mutual information scores with the predicted variable, these variables are considered to be highly correlated with the predicted variable; for the selected variables, look for variables with similar mutual information scores, which indicates that there is a close relationship between them; Step 4: Data preprocessing: preprocess the measured values ​​of the selected variables, including outlier detection, missing data filling and data standardization; data preprocessing can effectively improve data quality and consistency, and enhance the accuracy and stability of the established model; Step 5: Model building, using the preprocessed data obtained in the previous step to complete the training of the temporal convolutional gated recurrent neural network; Step 6: Model validation: using root mean square error RMSE, R squared R 2 And mean absolute percentage error MAPE performance indicators are used to evaluate the established models.

4. The method for predicting the pH value of flue gas desulfurization slurry based on time convolutional gated recurrent neural network according to claim 1 is characterized in that: The combined model contains a residual block, and the formula of the residual block is as follows: y=f(x,{W i })+x (9) Among them, x represents the input of the residual block, f(x,{W i }) represents a series of transformations on the module input, which is usually a complex function stacked by convolutional layers, activation functions, and batch normalization layers; {W i } is the learnable weight parameter in the transformation, and y represents the output of the residual block.

Images