A SO2 prediction method based on GWO-biGRU dynamic prediction model

Abstract

The invention discloses a SO2 prediction method based on a GWO-biGRU dynamic prediction model, which belongs to the technical field of glass factory desulfurization. The technical key points are: comprising the following steps: step one: constructing a bidirectional gated cyclic unit, and adopting the Gray Wolf optimization algorithm and the LightGBM algorithm to establish the GWO-biGRU dynamic prediction model; step two: establishing an outlet SO2 concentration prediction model, selecting auxiliary variables, and after the input characteristic variables of the model are determined, training the GWO-biGRU dynamic prediction model; step three: comparing the biGRU prediction model and the biGRU prediction model optimized by the Gray Wolf optimization algorithm, analyzing the prediction results, and having the advantages of predicting the outlet SO2 concentration in the desulfurization system, guiding the desulfurization operation, and reducing the SO2 concentration.

Description

Technical Field

[0001] The invention relates to the technical field of glass factory desulfurization, and in particular to a SO2 prediction method based on a GWO-biGRU dynamic prediction model. Background Art

[0002] At present, there are many desulfurization technologies and the technologies are mature. Generally speaking, there are three main types of coal desulfurization technologies, namely, desulfurization before combustion, during combustion and after combustion. All three types of desulfurization technologies contain methods worth learning from, among which the flue gas desulfurization technology in post-combustion desulfurization is the most representative.

[0003] Pre-combustion desulfurization, that is, desulfurization operations are performed before combustion, mainly including washing, microwave and biochemical treatment processes. Among them, washing technology is the most widely used, but this technology is not effective in removing organic sulfur, and the overall desulfurization rate is not high. The desulfurization rate of chemical treatment is relatively high, but it requires relatively harsh environmental conditions and requires catalyst support. The cost of use is high and the practicality is poor. Compared with other methods, biological treatment mainly uses microorganisms to optimize coal quality, has lower requirements for environmental conditions, and is not costly, but the desulfurization speed is slow. It is still in the research stage and has not been applied on a large scale.

[0004] Desulfurization during combustion means that desulfurization is performed during the combustion process in the furnace. The principle of this technology is to add limestone particles during the combustion process to react with SO2 in the flue gas to achieve the purpose of desulfurization. However, the control accuracy of limestone particles is very high, and it is easy to cause the furnace to flame out during the addition process, resulting in economic losses to the power plant. This method is not widely used.

[0005] Post-combustion desulfurization is the desulfurization of flue gas generated by the boiler after combustion, which mainly includes dry flue gas desulfurization, semi-dry flue gas desulfurization and wet flue gas desulfurization. Among them, wet flue gas desulfurization technology is the most widely used. After coal is burned, the flue gas discharged through the boiler flue mainly contains acidic substances, which need to be neutralized by alkaline substances such as limestone, ammonia or seawater to achieve the purpose of desulfurization.

[0006] Among wet flue gas desulfurization technologies, the limestone-gypsum wet flue gas desulfurization method is the most commonly used, and more than 90% of the units put into operation in my country use this method. In this method, limestone is used as a detergent to react with SO2 in the flue gas to generate CaSO3 and CaSO4, which are oxidized by air to eventually generate CaSO4 slurry, which is then dehydrated into gypsum. Compared with other methods, this method has the advantage of mature technology and is also highly environmentally friendly. Summary of the invention

[0007] In view of the deficiencies in the prior art, an embodiment of the present invention aims to provide a SO2 prediction method based on a GWO-biGRU dynamic prediction model to solve the problems in the above-mentioned background technology.

[0008] To achieve the above object, the present invention provides the following technical solutions:

[0009] A SO2 prediction method based on the GWO-biGRU dynamic prediction model comprises the following steps:

[0010] Step 1: Construct a bidirectional gated recurrent unit and use the Gray Wolf Optimization Algorithm and LightGBM algorithm to establish a GWO-biGRU dynamic prediction model;

[0011] Step 2: Establish an outlet SO2 concentration prediction model, select auxiliary variables, and after the input characteristic variables of the model are determined, train the GWO-biGRU dynamic prediction model;

[0012] Step 3: Compare the biGRU prediction model and the biGRU prediction model optimized by the Gray Wolf Optimization Algorithm, and analyze the prediction results.

[0013] As a further solution of the present invention, the grey wolf optimization algorithm in step 1 includes:

[0014] S101: Surrounding the prey. During the hunting process, the gray wolf group will first surround the prey. The optimization mathematical model is:

[0015]

[0016]

[0017] Where: t is the number of iterations at the current moment, is the coordination vector, where (in To control the parameters, it decreases linearly from 2 to 0. is a random vector in [0,1]), is the position vector of the prey at the tth iteration, is the position vector of the gray wolf at the tth iteration;

[0018] S102: Hunting. After the prey is surrounded, the wolf pack will hunt under the leadership of the three alpha wolves, β and δ. During the optimization process, the position information of the three alpha wolves is saved in each iteration, and the rest of the wolf pack updates its own position according to the position of the alpha wolves. The optimization mathematical model is:

[0019]

[0020]

[0021]

[0022] Where: is the position vector of the three alpha wolves in the current iteration round, is the individual position vector in t iterations, is a random vector, is the distance between the individual and the alpha wolf, is the updated individual position vector;

[0023] S103: Attack prey, through the iterative process The linear decrease of is realized, and the group obtains the optimal solution of the objective function after the iteration.

[0024] As a further solution of the present invention, the bidirectional gated recurrent unit in step 1 includes:

[0025] GRU introduces a new gating mechanism based on LSTM, including reset gate and update gate. GRU uses update gate to control the flow of information. The specific formula of GRU is as follows:

[0026] z t =σ(W z ·[h t-1 ,x t ])

[0027] r t =σ(W r ·[h t-1 ,x t ])

[0028]

[0029]

[0030] Among them, z t and r t Represent the outputs of the update gate and reset gate respectively; σ represents the sigmoid activation function; W z , W r , W represents the relevant weight matrix; [h t-1 ,x t ] means including h t-1 and x t A two-element vector; x t Indicates the current input, h t represents the hidden state passed to the next moment, represents the candidate hidden state;

[0031] The transmission state of GRU is unidirectional transmission from forward to backward. We choose to use a bidirectional gated recurrent unit neural network with recurrent units in both forward and backward directions. The formula of the biGRU network is as follows:

[0032]

[0033]

[0034]

[0035] in, represents the forward hidden layer state output, represents the backward hidden layer state output, w t , v t Respectively and The corresponding weight of b i Indicates the corresponding offset.

[0036] As a further solution of the present invention, the LightGBM in step 1 includes:

[0037] LightGBM uses a unilateral gradient sampling algorithm, which will provide marginal contribution when estimating information gain. LightGBM introduces a mutually exclusive feature bundling method, which merges mutually exclusive features to ensure that the meaning of the features remains unchanged while reducing the number of features. LightGBM uses a histogram method to divide the feature values ​​into multiple small intervals. LightGBM uses a leaf growth algorithm to locate the leaves with the highest gain and split them, while limiting the leaves with low gains.

[0038] As a further solution of the present invention, the establishment of the outlet SO2 concentration prediction model in step 2 includes:

[0039] The 3σ criterion is used to determine the outliers, and the average of the two values ​​before and after the outliers is used to fill them. The mean filling method is used for processing, and wavelet denoising is used to distinguish the original signal from the noise signal, so as to achieve the best recovery of the original signal.

[0040] As a further solution of the present invention, the selection of auxiliary variables in step 2 includes:

[0041] The samples were subjected to three feature selections, and the average value was taken as the final feature weight. Finally, the slurry pH value, total furnace air, flue gas flow rate at the desulfurization tower inlet, SO2 concentration at the inlet, kiln load, and slurry supply flow rate with larger characteristics were selected. The measurement points of each input characteristic of the combustion system were affected by factors such as the pipeline length. The leading moment was taken as the feature and the lag time with the SO2 emission concentration was used to determine its time step, and then the lag time of other characteristics was iterated to achieve the temporal consistency of the selected features.

[0042] As a further solution of the present invention, the step 2 of selecting and training the GWO-biGRU dynamic prediction model includes:

[0043] The data set is divided into a training set and a test set, and normalized. The training set samples are input into the optimized biGRU network to train the GWO-biGRU prediction model.

[0044] As a further solution of the present invention, the prediction result analysis includes:

[0045] 7500 sets of data sampled at 10s intervals provided by its DCS system were selected. The unit load corresponding to the time series of the data used showed a wide range of changes, including load fluctuation segments and stable segments.

[0046] In summary, compared with the prior art, the embodiments of the present invention have the following beneficial effects:

[0047] When it is widely used in power plant boiler DCS systems, the parameter information of each link reflecting the production operation status is transmitted to the monitoring computer through sensors for unified monitoring and storage, which lays a solid foundation for the implementation of this method and also lays a solid foundation for the practical application of power plant DCS.

[0048] Predicting the outlet SO2 concentration in the desulfurization system and guiding the desulfurization operation to reduce the SO2 concentration are of great significance to environmental protection and human health.

[0049] In order to more clearly illustrate the structural features and effects of the present invention, the present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 This is a GWO-biGRU flow chart of an embodiment of the invention.

[0051] Figure 2 These are the prediction results of different prediction models in the embodiments of the invention. DETAILED DESCRIPTION

[0052] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0053] The specific implementation of the present invention is described in detail below in conjunction with specific embodiments.

[0054] In one embodiment, a SO2 prediction method based on the GWO-biGRU dynamic prediction model is described in detail. Figure 1-2 , including the following steps:

[0055] Step 1: Construct a bidirectional gated recurrent unit and use the Gray Wolf Optimization Algorithm and LightGBM algorithm to establish a GWO-biGRU dynamic prediction model;

[0056] Step 2: Establish an outlet SO2 concentration prediction model, select auxiliary variables, and after the input characteristic variables of the model are determined, train the GWO-biGRU dynamic prediction model;

[0057] Step 3: Compare the biGRU prediction model and the biGRU prediction model optimized by the Gray Wolf Optimization Algorithm, and analyze the prediction results.

[0058] For further information, see Figure 1-2 , the grey wolf optimization algorithm in step 1 includes:

[0059] S101: Surrounding the prey. During the hunting process, the gray wolf group will first surround the prey. The optimization mathematical model is:

[0060]

[0061]

[0062] Where: t is the number of iterations at the current moment, is the coordination vector, where (in To control the parameters, it decreases linearly from 2 to 0. is a random vector in [0,1]), is the position vector of the prey at the tth iteration, is the position vector of the gray wolf at the tth iteration;

[0063] S102: Hunting. After the prey is surrounded, the wolf pack will hunt under the leadership of the three alpha wolves, β and δ. During the optimization process, the position information of the three alpha wolves is saved in each iteration, and the rest of the wolf pack updates its own position according to the position of the alpha wolves. The optimization mathematical model is:

[0064]

[0065]

[0066]

[0067] Where: is the position vector of the three alpha wolves in the current iteration round, is the individual position vector in t iterations, is a random vector, is the distance between the individual and the alpha wolf, is the updated individual position vector;

[0068] S103: Attack prey, through the iterative process The linear decrease of is realized, and the group obtains the optimal solution of the objective function after the iteration.

[0069] For further information, see Figure 1-2 In step 1, the bidirectional gated recurrent unit comprises:

[0070] GRU introduces a new gating mechanism based on LSTM, including reset gate and update gate. GRU uses update gate to control the flow of information. The specific formula of GRU is as follows:

[0071] z t =σ(W z ·[h t-1 ,x t ])

[0072] r t =σ(W r ·[h t-1 ,x t ])

[0073]

[0074]

[0075] Among them, z t and r t Represent the outputs of the update gate and reset gate respectively; σ represents the sigmoid activation function; W z , W r , W represents the relevant weight matrix; [h t-1 ,x t ] means including h t-1 and x t A two-element vector; x t Indicates the current input, h t represents the hidden state passed to the next moment, represents the candidate hidden state;

[0076] The transmission state of GRU is unidirectional transmission from forward to backward. We choose to use a bidirectional gated recurrent unit neural network with recurrent units in both forward and backward directions. The formula of the biGRU network is as follows:

[0077]

[0078]

[0079]

[0080] in, represents the forward hidden layer state output, represents the backward hidden layer state output, w t , v t Respectively and The corresponding weight of b i Indicates the corresponding offset.

[0081] For further information, see Figure 1-2 , in step 1, LightGBM includes:

[0082] LightGBM uses a unilateral gradient sampling algorithm, which will provide marginal contribution when estimating information gain. LightGBM introduces a mutually exclusive feature bundling method, which merges mutually exclusive features to ensure that the meaning of the features remains unchanged while reducing the number of features. LightGBM uses a histogram method to divide the feature values ​​into multiple small intervals. LightGBM uses a leaf growth algorithm to locate the leaves with the highest gain and split them, while limiting the leaves with low gains.

[0083] For further information, see Figure 1-2 The step 2 of establishing the outlet SO2 concentration prediction model includes:

[0084] The 3σ criterion is used to determine the outliers, and the average of the two values ​​before and after the outliers is used to fill them. The mean filling method is used for processing, and wavelet denoising is used to distinguish the original signal from the noise signal, so as to achieve the best recovery of the original signal.

[0085] For further information, see Figure 1-2 , the auxiliary variables selected in step 2 include:

[0086] The samples were subjected to three feature selections, and the average value was taken as the final feature weight. Finally, the slurry pH value, total furnace air, flue gas flow rate at the desulfurization tower inlet, SO2 concentration at the inlet, kiln load, and slurry supply flow rate with larger characteristics were selected. The measurement points of each input characteristic of the combustion system were affected by factors such as the pipeline length. The leading moment was taken as the feature and the lag time with the SO2 emission concentration was used to determine its time step, and then the lag time of other characteristics was iterated to achieve the temporal consistency of the selected features.

[0087] For further information, see Figure 1-2 The step 2 of selecting and training the GWO-biGRU dynamic prediction model includes:

[0088] The data set is divided into a training set and a test set, and normalized. The training set samples are input into the optimized biGRU network to train the GWO-biGRU prediction model.

[0089] For further information, see Figure 1-2 , the prediction result analysis includes:

[0090] 7500 sets of data sampled at 10s intervals provided by its DCS system were selected. The unit load corresponding to the time series of the data used showed a wide range of changes, including load fluctuation segments and stable segments.

[0091] In this embodiment, 1.1 Gray Wolf Optimization Algorithm, Gray Wolf Optimization Algorithm (GWO) is a swarm intelligence optimization algorithm with fast convergence speed, few parameters and easy implementation. Its optimization steps are as follows: first, a random gray wolf group is generated in the designated search space, and then the gray wolves are divided into different social levels according to the individual fitness. The top three fitness are set as α, β, δ, and the rest of the wolf groups are set as ω. The optimization process is guided by the best three solutions, followed by the process of surrounding prey, hunting and attacking prey, and finally finding the optimal solution of the objective function. Each process is described in detail as follows:

[0092] (1) Surrounding the prey

[0093] During the hunting process, the gray wolf group will first surround the prey, and its optimization mathematical model is:

[0094]

[0095]

[0096] Where: t is the number of iterations at the current moment, is the coordination vector, where (in To control the parameters, it decreases linearly from 2 to 0. is a random vector in [0,1]), is the position vector of the prey at the tth iteration, is the position vector of the gray wolf at the tth iteration.

[0097] (2) Hunting

[0098] When the prey is surrounded, the wolf pack will hunt under the leadership of the three alpha wolves, β, and δ. In the optimization process, the position information of the three alpha wolves is saved in each iteration, and the rest of the wolf pack updates its own position according to the position of the alpha wolves. The optimization mathematical model is:

[0099]

[0100]

[0101]

[0102] Where: is the position vector of the three alpha wolves in the current iteration round, is the individual position vector in t iterations, is a random vector, is the distance between the individual and the alpha wolf, is the updated individual position vector.

[0103] (3) Attacking prey

[0104] This stage is the process of finding the optimal solution of the objective function, mainly through the iterative process. The linear decrease of is realized, and the group obtains the optimal solution of the objective function after the iteration.

[0105] 1.2 Bidirectional Gated Recurrent Unit

[0106] The Gated Recurrent Unit (GRU) is a variant of the recurrent neural network (RNN) proposed by Cho et al.

[19] in 2014 to solve the gradient explosion problem of traditional RNNs. It is used to process sequence data. GRU introduces a new gating mechanism based on LSTM, which only includes two key gating mechanisms, namely the reset gate and the update gate, making the structure simpler. This simplification helps to reduce the occurrence of the gradient vanishing problem and reduces the number of parameters in the model, thereby improving the training and inference speed. GRU uses the update gate to control the flow of information, which is similar to the input gate and forget gate of LSTM. The specific formula of GRU is as follows:

[0107] z t =σ(W z ·[h t-1 ,xt ])

[0108] r t =σ(W r ·[h t-1 ,x t ])

[0109]

[0110]

[0111] Among them, z t and r t Represent the outputs of the update gate and reset gate respectively; σ represents the sigmoid activation function; W z , W r , W represents the relevant weight matrix; [h t-1 ,x t ] means including h t-1 and x t A two-element vector; x t Indicates the current input, h t Represents the hidden state passed to the next moment; represents a candidate hidden state.

[0112] The transmission state of GRU is unidirectional transmission from the front to the back, which is usually used to solve the problem that the output at the current moment is only related to a single state variable in the past

[20] . Therefore, this paper chooses to use a bidirectional gated recurrent unit neural network with recurrent units in both forward and backward directions, allowing it to consider the past and future of sequence data at the same time. This can mine complex dependencies in the data and is conducive to the prediction of nonlinear data. Its network structure is as follows: Figure 1 As shown. The formula of the biGRU network is as follows:

[0113]

[0114]

[0115]

[0116] in, represents the forward hidden layer state output, represents the backward hidden layer state output; w t , v t Respectively and The corresponding weight of b i Indicates the corresponding offset.

[0117] 1.3 LightGBM

[0118] LightGBM is a machine learning algorithm based on gradient boosting decision trees. As an open source distributed gradient boosting framework, it can efficiently perform data modeling and prediction.

[0119] The basic idea of ​​LightGBM is to gradually merge multiple weak regression trees to form a strong regression tree, thereby reducing the prediction error rate. Compared with XGBoost, it has the following advantages: (1) LightGBM uses a gradient-based one-side sampling algorithm (GOSS). GOSS discards data instances with small gradients because data instances with small gradients will provide marginal contributions when estimating information gain. Through the GOSS algorithm, LightGBM reduces the number of data instances, improves the utilization of high-gradient data instances, and greatly reduces the sampling time. (2) LightGBM introduces the exclusive feature bundling (EFB) method, which aims to reduce the number of features by bundling mutually exclusive features together. In the generalized feature space, there are many mutually exclusive properties, and they cannot all be non-zero values. In this way, these mutually exclusive features may be reasonably combined together to become a single feature to achieve feature dimensionality reduction, thereby improving computational efficiency. In other words, the EFB method reduces the number of features while ensuring that the meaning of the features remains unchanged by merging mutually exclusive features. (3) LightGBM uses a histogram method to divide the eigenvalues ​​into multiple small intervals for data segmentation and modeling on the decision tree. This method greatly reduces the calculation time and usually obtains better prediction results. (4) LightGBM uses a leaf-wise growth algorithm. This algorithm can locate the leaves with the highest gain and split them, while limiting the leaves with low gains to avoid wasting resources. This leaf-oriented growth method can improve training efficiency. In addition, LightGBM can also reduce the risk of overfitting while ensuring efficiency by limiting the maximum depth of the leaves.

[0120] 2 Establishing a prediction model for outlet SO2 concentration

[0121] The DCS system of a glass factory provided data with a time span of 7 days and a time interval of 10s. This paper selected 5300 sets of data with large unit load changes, of which the first 5000 sets of data were used as the training set of the model, and the last 300 sets of data were used as the test set of the model.

[0122] Since the data collected by DCS may be missing, affected by outliers and noise, the 3σ criterion is used to determine the outliers first, and the average of the two values ​​before and after the outliers is used to fill them. For missing values, the mean filling method is also used for processing. The noise is reduced by wavelet noise reduction, which is a method that uses a given metric to optimally approximate the original signal, thereby distinguishing the original signal from the noise signal and achieving the best recovery of the original signal.

[0123] 2.1 Selection of auxiliary variables

[0124] More than 5,000 continuous samples were selected from the original data, and the samples were feature selected three times, and the average value was taken as the final feature weight. Finally, the slurry pH value, total furnace wind, flue gas flow at the inlet of the desulfurization tower, inlet SO2 concentration, kiln load, and slurry supply flow with larger characteristics were selected. The measurement points of each input characteristic of the combustion system are affected by factors such as pipeline length, and there is a large lag between them and the measured value of the outlet SO2 concentration. Considering that a combustion cycle is at most 300 seconds, the maximum delay time of 300 seconds is selected as the feature on the basis of the maximum delay time, and the rest are unchanged, each time is a time step, and then LightGBM is used to calculate the importance of each feature until 300 seconds. The most important moment is selected, and its leading moment is used as the lag time between the feature and the SO2 emission concentration, and its time step is determined. Then, the lag time of other features is iterated to achieve the temporal consistency of the selected features. Table 1 shows the optimal delay time between the selected features and the outlet SO2.

[0125] Table 1 Variation range of each variable and time delay compensation value

[0126] variable Transformation Range Time compensation / s Slurry pH 5.6-5.8 20 Total furnace air / t·h-1 1350-1900 100 <![CDATA[Flue gas flow rate at the inlet of the desulfurization tower / km 3 ·h-1]]> 1235-1879 60 <![CDATA[Inlet SO2 concentration / mg·m-3]]> 5454-7896 50 Kiln load / MW 403.7-575.3 210 <![CDATA[Supply pulp flow rate / m 3 ·h-1]]> 61.43-79.46 130

[0127] 2.2 Prediction model construction based on GWO-biGRU

[0128] The data set is divided into a training set and a test set, and then normalized. The Gray Wolf Optimization Algorithm is used to optimize the hyperparameters of the biGRU network, such as the learning rate, number of hidden layers, and number of training rounds. The learning rate of the model is 0.0024, the number of hidden layers is 8, the minimum training batch is 95, and the maximum number of training rounds is 104. The training set samples are input into the optimized biGRU network to train the GWO-biGRU prediction model.

[0129] 3. Analysis of prediction results

[0130] The data used in the above method are 5300 sets of data sampled at 10s intervals provided by its DCS system. The unit load corresponding to the time series of the data used has a large range of changes, including load fluctuation segments and stable segments. The first 5000 sets of data are used as training sets, and the last 300 sets are used as test sets.

[0131] Comparison of the biGRU prediction model and the biGRU prediction model optimized by the Gray Wolf Optimization Algorithm Figure 2 As shown in the figure, the experimental results show that the prediction accuracy of the biGRU prediction model optimized by the Gray Wolf Optimization Algorithm is greatly improved compared with the single model. Its prediction results are advanced and have the effect of guiding intelligent optimization control.

[0132] The prediction model proposed in this paper after parameter optimization using the GWO algorithm is closer to the actual value of SO2 concentration than the traditional biGRU prediction model without optimization, and the root mean square error and maximum error are better than the ordinary biGRU prediction model. The root mean square error (RMSE), mean absolute percentage error (MAPE) and mean absolute error (MAE) are shown in Table 2.

[0133] Table 2 Comparison of errors of different models

[0134] algorithm MAPE MAE RMSE biGRU prediction model 1.642 2.324 2.245 GWO-biGRU prediction model 1.524 1.856 1.657

[0135] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A SO2 prediction method based on the GWO-biGRU dynamic prediction model, characterized in that: The following steps are involved: Step 1: Construct a bidirectional gated cycle unit and use the Gray Wolf Optimization Algorithm and LightGBM algorithm to establish a GWO-biGRU dynamic prediction model to predict the outlet SO2 concentration in the desulfurization system; Step 2: Establish the outlet SO2 concentration prediction model, and select auxiliary variables in the following way: perform three feature selections on the sample, and take the average value as the final feature weight, and finally select slurry pH value, furnace total air, desulfurization tower inlet flue gas flow, inlet SO2 concentration, kiln load, and slurry supply flow as auxiliary variables. Considering that the measurement points of various input characteristics of the combustion system are affected by the pipeline length factor, there is a large lag between it and the outlet SO2 concentration measurement value. Considering that a combustion cycle is at most 300 seconds, the maximum delay time of 300 seconds is selected on the basis of the maximum delay time. As features, the rest are unchanged, each time is a time step, and LightGBM is used to calculate the importance of each feature within 300 seconds. The most important moment is selected, and its leading moment is used as the feature and the lag time of SO2 emission concentration, its time step is determined, and the lag time of other features is iterated to achieve the temporal consistency of the selected features; after determining the input feature variables of the model, the data set is divided into a training set and a test set, and normalized, and then the training set samples are input into the optimized biGRU network to train the GWO-biGRU prediction model; Step 3: Compare the biGRU prediction model and the biGRU prediction model optimized by the Gray Wolf Optimization Algorithm, and analyze the prediction results to verify the prediction effect of the GWO-biGRU prediction model.

2. The SO2 prediction method based on the GWO-biGRU dynamic prediction model according to claim 1 is characterized in that: The gray wolf optimization algorithm in step 1 includes: S101: Surrounding the prey. During the hunting process, the gray wolf group will first surround the prey. The optimization mathematical model is: Where: t is the number of iterations at the current moment, is the coordination vector, where (in To control the parameters, it decreases linearly from 2 to 0. is a random vector in [0,1]), is the position vector of the prey at the tth iteration, is the position vector of the gray wolf at the tth iteration; S102: Hunting. After the prey is surrounded, the wolf pack will hunt under the leadership of the three alpha wolves, β and δ. During the optimization process, the position information of the three alpha wolves is saved in each iteration, and the rest of the wolf pack updates its own position according to the position of the alpha wolves. The optimization mathematical model is: Where: is the position vector of the three alpha wolves in the current iteration round, is the individual position vector in t iterations, is a random vector, is the distance between the individual and the alpha wolf, is the updated individual position vector; S103: Attack prey, through the iterative process The linear decrease of is realized, and the group obtains the optimal solution of the objective function after the iteration.

3. The SO2 prediction method based on the GWO-biGRU dynamic prediction model according to claim 2 is characterized in that: The bidirectional gated cycle unit in step 1 comprises: GRU introduces a new gating mechanism based on LSTM, including reset gate and update gate. GRU uses update gate to control the flow of information. The specific formula of GRU is as follows: z t =σ(W z ·[h t-1 ,x t ]) r t =σ(W r ·[h t-1 ,x t ]) Among them, z t and r t Represent the outputs of the update gate and reset gate respectively; σ represents the sigmoid activation function; W z , W r , W represents the relevant weight matrix; [h t-1 ,x t ] means including h t-1 and x t A two-element vector; x t Indicates the current input, h t represents the hidden state passed to the next moment, represents the candidate hidden state; The transmission state of GRU is unidirectional transmission from forward to backward. We choose to use a bidirectional gated recurrent unit neural network with recurrent units in both forward and backward directions. The formula of the biGRU network is as follows: in, represents the forward hidden layer state output, Represents the backward hidden layer state output, wt, vt represent and The corresponding weight of b i Indicates the corresponding offset.

4. The SO2 prediction method based on the GWO-biGRU dynamic prediction model according to claim 3 is characterized in that: In step 1, LightGBM includes: LightGBM uses a unilateral gradient sampling algorithm, which will provide marginal contribution when estimating information gain. LightGBM introduces a mutually exclusive feature bundling method, which merges mutually exclusive features to ensure that the meaning of the features remains unchanged while reducing the number of features. LightGBM uses a histogram method to divide the feature values ​​into multiple small intervals. LightGBM uses a leaf growth algorithm to locate the leaves with the highest gain and split them, while limiting the leaves with low gains.

5. The SO2 prediction method based on the GWO-biGRU dynamic prediction model according to claim 1 is characterized in that: The step 2 of establishing the outlet SO2 concentration prediction model includes: The 3σ criterion is used to determine the outliers, and the average of the two values ​​before and after the outliers is used to fill them. The mean filling method is used for processing, and wavelet denoising is used to distinguish the original signal from the noise signal, so as to achieve the best recovery of the original signal.

6. The SO2 prediction method based on the GWO-biGRU dynamic prediction model according to claim 5 is characterized in that: The selection of auxiliary variables in step 2 includes: The samples were subjected to three feature selections, and the average value was taken as the final feature weight. Finally, the slurry pH value, total furnace air, flue gas flow rate at the desulfurization tower inlet, SO2 concentration at the inlet, kiln load, and slurry supply flow rate were selected. The measurement points of the input characteristics of the combustion system were affected by factors such as the pipeline length. The leading moment was taken as the feature. The lag time with the SO2 emission concentration was used to determine its time step, and then the lag time of other features was iterated to achieve the temporal consistency of the selected features.

7. The SO2 prediction method based on the GWO-biGRU dynamic prediction model according to claim 1 is characterized in that: The step 2 of selecting and training the GWO-biGRU dynamic prediction model includes: The data set is divided into a training set and a test set, and normalized. The training set samples are input into the optimized biGRU network to train the GWO-biGRU prediction model.

8. The SO2 prediction method based on the GWO-biGRU dynamic prediction model according to claim 1 is characterized in that: The prediction result analysis includes: 7500 sets of data sampled at 10s intervals provided by its DCS system were selected. The unit load corresponding to the time series of the data used showed a wide range of changes, including load fluctuation segments and stable segments.

Images