Tranform-based modeling method for mechanism hybrid prediction model of SCR (Selective Catalytic Reduction) denitration system

Abstract

The invention discloses a modeling method of an SCR (Selective Catalytic Reduction) denitration system mechanism hybrid prediction model based on Transform. Feature extraction is carried out on the collected data, a reaction kinetic equation is constructed, an Euler method is adopted for solving, and unknown model parameters are reserved; a section of working condition stable data is selected from the collected data and imported into a mechanism model, optimal model parameters are found, the mechanism model of the SCR denitration reaction is obtained, and the A side outlet flue concentration and the B side outlet flue concentration are calculated through the model; a Transform model is trained to compensate the SCR system mechanism model, and the compensated SCR system mechanism model serves as real-time soft measurement values output by the A side and the B side; input variable selection is carried out on all-condition data, A-side and B-side outlet flue concentrations obtained by the mechanism model in real time and outputs at the first two moments are included into an input variable set, and time delay processing is carried out on input variables; and the mechanism models of the A-side flue and the B-side flue are connected in series with a Transform model, the Transform model needs to be imported into the output of the mechanism models in real time, and finally the SCR denitration system hybrid prediction model is obtained.

Description

Technical Field

[0001] The present invention relates to the field of thermal power generation industrial control, and particularly to a method for modeling a mechanism hybrid prediction model of an SCR denitration system based on a Transformer. Background Art

[0002] With the acceleration of the industrialization process and the continuous growth of energy demand, thermal power generation, as the main way of power supply in China, while promoting economic development, also brings serious environmental pollution problems. Among them, nitrogen oxides (NO x ) as one of the main pollutants generated during thermal power generation pose a serious threat to the atmospheric environment and human health. NO x not only directly stimulates the human respiratory system, causing diseases such as asthma and bronchitis, but also reacts with other pollutants in the atmosphere to generate secondary particulate matter and ozone, leading to the formation of photochemical smog and acid rain, and further damaging the ecosystem, crops and buildings. In addition, NO x as an important greenhouse gas, also exacerbates global climate change. To address the NO x pollution problem, China has introduced a series of strict environmental protection policies in recent years. The NO x emission concentration of the thermal power industry is restricted to below 50 mg / m 3 . The implementation of these policies has forced thermal power plants to adopt efficient and reliable denitration technologies to meet the increasingly strict environmental protection requirements.

[0003] Among many denitration technologies, selective catalytic reduction (SCR) technology has become the preferred solution for thermal power plants to reduce NO x emissions due to its high denitration efficiency, stability and adaptability. The SCR technology uses ammonia (NH3) or urea as a reducing agent under the action of a catalyst to convert NO x in the flue gas into harmless nitrogen (N2) and water (H2O), and its denitration efficiency can reach more than 90%. The SCR system is usually installed between the economizer and the air preheater of the boiler, and uses the heat in the flue gas and the action of the catalyst to achieve efficient reaction in the temperature range of 300 - 400 °C. In addition, the SCR technology can also work in coordination with other environmental protection equipment (such as electrostatic precipitators and wet flue gas desulfurization systems) to form a complete flue gas purification chain, further reducing pollutant emissions. However, the application of the SCR technology also faces some challenges, such as the high cost of the catalyst, the possible secondary pollution caused by ammonia slip, and the precise control requirements for temperature, flow rate and reducing agent dosage during the operation of the system. Therefore, how to optimize the modeling and control strategy of the SCR system to improve the denitration efficiency and reduce the operation cost has become a hot research issue. Summary of the Invention

[0004] The object of the present invention is to overcome the above-mentioned disadvantages and deficiencies of the prior art, and provide a modeling method for a mechanism hybrid prediction model of an SCR denitration system based on Transformer. Through the chemical reaction mechanism of the SCR denitration system, the present invention constructs a reaction kinetic equation to obtain a system mechanism model, and inputs the real-time outputs of the A and B side flue gas outlet NO x concentrations and other directly measured variables such as load and air volume into the Transformer prediction module, and finally establishes a hybrid prediction model.

[0005] The present invention is realized through the following technical solutions:

[0006] A modeling method for a mechanism hybrid prediction model of an SCR denitration system based on Transformer, the modeling method comprising the following steps:

[0007] S1. Extract effective characteristic data from the data collected by the on-site unit, and at the same time ensure that the full operating conditions of the unit are covered;

[0008] S2. Construct a reaction kinetic equation through the SCR denitration reaction mechanism, and use the Euler method for solution, retaining unknown model parameters;

[0009] S3. Select a section of data with relatively stable operating conditions from the collected data, import it into the mechanism model, and find the optimal model parameters through the grey wolf optimization algorithm. Thus, the mechanism model of the SCR denitration reaction is obtained, and the concentrations of the A and B side outlet flue gases can be calculated by means of this model;

[0010] S4. Train the Transformer model to compensate the SCR system mechanism model, improve the output accuracy of the mechanism model under variable operating conditions, and use it as the real-time soft measurement value of the A and B side outputs;

[0011] S5. Select input variables from the full operating condition data, incorporate the concentrations of the A and B side outlet flue gases obtained in real time by the mechanism model and the outputs of the previous two moments (for example, the previous two seconds) into the input variable set, and perform time delay processing on the input variables;

[0012] S6. Train the tail flue Transformer prediction model with the input variable set obtained in S5 to obtain the prediction model for this part;

[0013] S7. Connect the mechanism models of the A and B side flues in series with the Transformer model. The Transformer model needs to import the output of the mechanism model in real time, as well as other variables that can be directly measured by sensors, and finally obtain the hybrid prediction model of the SCR denitration system.

[0014] In step S3, after the reaction kinetics equation is constructed according to the Eley-Rideal mechanism, the equation is solved by the Euler method, and the number of iterative calculations is determined according to the physical model of the reaction vessel and the residence time of the flue gas, so that iterative calculations are performed in the process of calculating the numerical solution; the mechanism model differential equation solving method retains the unknown parameters of the mechanism model in the process, uses the historical data actually collected from the power plant units, adopts the Gray Wolf optimization method for identification, and substitutes the identified parameters to obtain the mechanism model.

[0015] In step S4, the Transformer model is used to perform accuracy compensation on the mechanism model.

[0016] In step S5, the selection of the input variable set includes the total set value under the clean flue gas mode, the SCR inlet flue concentration on the A side, the SCR ammonia flow rate on the A side, the SCR inlet flue concentration on the B side, the SCR ammonia flow rate on the B side, the total air volume, the operating load, the coal feed rate of the coal feeder, the denitrification inlet flue gas temperature on the A side and the denitrification inlet flue gas temperature on the B side, and the SCR outlet flue concentration on the A side and the SCR outlet flue concentration on the B side output by the mechanism model are included in the variable set.

[0017] For multi-input systems such as SCR denitrification, there is redundancy between some information, and the partial mutual information method is used to decouple the input variables.

[0018] The input data monitored and recorded by the sensor have different degrees of time delay for the output, i.e., the concentration at the tail flue outlet. The mutual information method is used to calculate the MI value between each input variable and the output variable, find the optimal time delay for each variable, and perform data alignment.

[0019] Compared with the prior art, the present invention has the following advantages and effects:

[0020] (1) Establishing an accurate controlled object model is the prerequisite for designing a high-performance controller, which can achieve precise control of the amount of ammonia injection, ensure that the concentration at the tail flue outlet of the SCR denitrification system accurately tracks the set value, and strictly ensure compliance with the flue gas emission standards.

[0021] (2) The model is based on the internal reaction mechanism, and the model parameters have clear physical meanings, which are easy to understand and interpret. It can still maintain good prediction accuracy under changing operating conditions or new scenarios, and is suitable for design optimization and extreme condition analysis.

[0022] (3) The Transformer algorithm is used to predict the tail flue concentration, and the output of the mechanism model and historical data such as load are used as input to make the prediction result more accurate and make up for the lack of output accuracy of the mechanism model for variable operating conditions.

[0023] (4) The hybrid prediction model combines mechanism modeling, effectively avoiding the model mismatch problem commonly existing in data-driven modeling. Description of the Drawings

[0024] Figure 1 It is the flowchart of the optimization of the grey wolf algorithm of the present invention.

[0025] Figure 2 It is the mechanism compensation structure diagram of the SCR denitration system of the present invention.

[0026] Figure 3 It is the flowchart of the selection of input variables of partial mutual information of the present invention.

[0027] Figure 4 It is the structure diagram of the hybrid prediction model of the SCR denitration system of the present invention.

[0028] Figure 5 It is the model effectiveness verification diagram of the present invention.

[0029] Figure 6 It is the experimental comparison result diagram of the present invention.

[0030] Figure 7 It is the experimental comparison result diagram of the present invention. Detailed Embodiments

[0031] The present invention will be further described in detail below with reference to specific embodiments.

[0032] As Figures 1-7 shown. The present invention discloses a modeling method for a mechanism hybrid prediction model of an SCR denitration system based on Transformer. The modeling method of the prediction model can be realized through the following steps:

[0033] (1) Extract effective characteristic data from the data collected by the on-site unit, and at the same time, it is necessary to ensure that the full operating conditions of the unit are covered.

[0034] (2) Construct a reaction kinetic equation through the SCR denitration reaction mechanism, and use the Euler method for solution, retaining the unknown model parameters.

[0035] (3) Select a section of data with relatively stable operating conditions from the collected data, import it into the mechanism model, and find the optimal model parameters through the grey wolf optimization algorithm. Thus, the mechanism model of the SCR denitration reaction is obtained, and the NO x concentration at the outlet flue of sides A and B can be calculated by means of this model.

[0036] (4) Train the Transformer model to compensate the SCR system mechanism model and improve the output accuracy of the mechanism model under variable operating conditions.

[0037] (5) Select input variables from the full operating condition data, and include the NO concentrations at the outlet flues on the A and B sides and the outputs at the previous two moments in the input variable set, and perform time delay processing on the input variables. x Concentrations and the outputs at the previous two moments are incorporated into the input variable set, and time delay processing is performed on the input variables.

[0038] (6) Train the Transformer prediction model with the input variable set obtained in the previous step to obtain the prediction model for this part.

[0039] (7) Connect the mechanism models of the flues on the A and B sides in series with the Transformer model. The Transformer model needs to import the output of the mechanism model in real time, as well as other variables that can be directly measured by sensors, and finally obtain the hybrid prediction model of the SCR denitration system.

[0040] In step (2), mechanism modeling can quantitatively describe the relationship between flue gas flow rate, temperature, catalyst activity and NO conversion rate through mass conservation, energy conservation and reaction kinetics equations, and construct differential equations. Solve the equations by the Euler method, adopt preset step size parameters in the numerical simulation process, and carry out calculation and analysis based on the SCR reaction device with specific geometric parameters. According to the standard flue gas flow rate characteristics at the outlet of the economizer under rated operating conditions, combined with the theoretical residence time requirements derived from fluid mechanics characteristics, complete the solution process through multiple iterative calculations. x In step (2), since there are still many parameters to be solved in the model, there are large coupling relationships among the parameters, and as the unit operating conditions change, the model parameters will also change accordingly, making it very difficult to directly obtain their numerical values. Therefore, in this study, the historical data actually collected from the power plant units is used, and the grey wolf optimization method is adopted for identification. The SCR denitration system is divided into the outlet on the A side and the outlet on the B side. The ultimate goal is to make the NO concentrations at the outlets on both sides of the mechanism model as close as possible to the actual values, and find the corresponding ten parameters at this time. The specific algorithm implementation process is as follows:

[0041] In step (2), since there are still many parameters to be solved in the model, there are large coupling relationships among the parameters, and as the unit operating conditions change, the model parameters will also change accordingly, making it very difficult to directly obtain their numerical values. Therefore, in this study, the historical data actually collected from the power plant units is used, and the grey wolf optimization method is adopted for identification. The SCR denitration system is divided into the outlet on the A side and the outlet on the B side. The ultimate goal is to make the NO concentrations at the outlets on both sides of the mechanism model as close as possible to the actual values, and find the corresponding ten parameters at this time. The specific algorithm implementation process is as follows: x Concentrations are as close as possible to the actual values, and find the corresponding ten parameters at this time. The specific algorithm implementation process is as follows:

[0042] 1. Algorithm initialization. Set the population size to 50, the problem dimension to 10, and the number of iterations in the optimization process to 600.

[0043] 2. Initialize the population. Randomly and uniformly generate the initial positions of the wolf pack.

[0044] 3. Fitness evaluation. Set the fitness function as:

[0045]

[0046] Where y A (k) and y B (k) are the actual NO concentrations at the outlets on the A and B sides of the SCR denitration system respectively.x The average concentration, and The NO on the A and B sides of the mechanism model are x The average value of the concentration.

[0047] 4. Use the quick sort algorithm to determine the α, β, and δ wolves, update the parameters, move to the new position, calculate the current population fitness, and update the status of the α, β, and δ wolves.

[0048] 5. Trigger the termination condition and extract the optimal solution.

[0049] For detailed algorithm flow, see Figure 1 .

[0050] In step (4), the mechanism model of the SCR denitrification system constructs a reaction kinetic equation based on the Eley-Rideal mechanism, and obtains the dynamic relationship between the elements through the law of conservation of mass and energy. Taking the A side channel as an example, according to equations 2.10-2.17, the reaction kinetic equation indirectly uses the A side SCR inlet flue NO x concentration, flue gas temperature at side A outlet, ammonia flow rate at side A and total air volume, and then through the unit conversion processing of equations 2.38-2.40, the NO at the flue gas outlet at side A can be directly calculated by the mechanism model. x Concentration, similarly, the NO concentration of the flue gas outlet on the B side can be obtained. x concentration.

[0051] The output of the mechanism model is the NO of the flue outlet on the A and B sides. x The concentration, as well as the coal feeder coal quantity, operating load and other variables not included in the calculation of the mechanism model will also be used as important inputs of the compensation model. These variables mainly make up for the lack of accuracy of the tail flue concentration output under variable conditions of the mechanism model. x The concentration always keeps a steady change, not a sudden change, so the output of the first two seconds is considered as the input variable of the Transformer compensation model. Figure 2 .

[0052] In step (5), the correlation between variables is measured by partial mutual information. For multi-input systems such as SCR denitrification, some information is redundant. For example, variable X i The mutual information between the output Y contains part of the variable X j Information, I(Y, X i ) will be greater than the actual value, and we need to use conditional expectation to convert X j Eliminate relevant information. Use PMI to measure the correlation between variables. Eliminate X j After Y, X iDenoted as M and N respectively, they are expressed as:

[0053] M = Y - f Y (X j )

[0054]

[0055] f Y (X j ) is the information of X contained in Y j . is the information of X contained in X i . j .

[0056]

[0057] p(x) is the kernel density estimation function. The partial mutual information PMI between Y and X is: i PMI(X

[0058] , Y) = I(N, M) i .

[0059] Suppose P is the set of candidate independent variables, Q is the set of dependent variables, and R is the set of optimal input variables. First, calculate f Q (R) of Q on the set R, and obtain the residual

[0060] M = Q - f Q (R)

[0061] For each variable P in P j calculate its kernel regression estimation on the set R Meanwhile, calculate the residual of P j .

[0062]

[0063] Calculate the mutual information I(M, N) between M and N, and find the candidate variable P with the maximum mutual information R , and finally calculate the Akaike information criterion (AIC) value.

[0064]

[0065] Where n is the number of samples, M i is the regression residual of Y calculated according to the selected variables, and p is the number of selected variables. If the Akaike information criterion decreases, P R needs to be placed into the set R, and this process is looped. When the Akaike information criterion no longer decreases, the optimal set R is output, and the loop ends. For the detailed algorithm flow, refer to Figure 3 .

[0066] In step (6), through the analysis of the SCR denitration system structure and in combination with the on-site equipment data, the current NO x concentration emission standard and the total NO x set value under the unit's clean flue gas mode both correspond to the concentration in the tail flue. The controller controls the NO x concentration at the outlets of the flues on both sides of A and B respectively, rather than directly affecting the NO x concentration in the tail flue. Moreover, the relationship between the NO x concentration in the tail flue and the NO x concentration at the outlets of the flues on both sides of A and B is not a simple linear relationship. It will also be affected by factors such as the air volume and the adsorption degree of the catalyst at different temperatures, resulting in uneven NO x concentration distribution.

[0067] The Transformer hybrid prediction model is precisely to solve the above problems and achieve real-time prediction of the NO x concentration in the tail flue. Here, a section of data with drastic changes in working conditions is selected as the training set and the test set. The specific implementation steps of the algorithm are as follows:

[0068] 1. Define the input and output. Let the number of samples be N, the feature dimension be d, and the input matrix be:

[0069]

[0070] The output vector is:

[0071]

[0072] 2. Standardize the processing. Perform Z-score standardization on each feature column:

[0073]

[0074] 3. Use the sliding window method to construct time series samples. Let the window length be L

[0075]

[0076] The corresponding output is y t , representing the NO x concentration at the moment of t + Δt.

[0077] 4. Divide the training set (X train , y train ) and the test set (X test , y test )

[0078] 5. Create the model architecture.

[0079] ① Input embedding layer. Map the d-dimensional input features in the dataset to dmodel The H-dimensional space (0) = SW e + b e

[0080] where are model parameters.

[0081] ② Positional encoding. Injecting temporal position information:

[0082]

[0083] where j ∈ {1,..., [d model / 2]}.

[0084] ③ Transformer encoder. It contains N identical layers, and each layer contains h multi-head self-attention:

[0085] MultiHead(Q, K, V) = Concat(head1,..., head h )W O

[0086]

[0087] W O , is a linear transformation matrix, and the feed-forward network:

[0088] FFN(x) = ReLU(xW1 + b1)W2 + b2

[0089] W1 and W2 are weight matrices, and b1 and b2 are bias terms.

[0090] ④ Output layer. Taking the hidden state at the last moment for regression prediction:

[0091]

[0092] where are all learnable parameters, and h L is the hidden state at the last moment.

[0093] 6. Design a training strategy and use the mean squared error loss as the loss function.

[0094]

[0095] 7. Monitor the loss of the validation set and terminate the training when it has not decreased for E patience = 10 consecutive epochs.

[0096] In step (7) above, in the designed Transformer hybrid prediction model, the NO at the A and B side outlet flues output by the mechanism modelx As a real-time soft measurement, the concentration is further used as a key input variable and input into the prediction model in real time for predicting the concentration in the tail flue. At the same time, other influencing factors such as the coal feeding amount of the coal feeder and the air volume for the NO x concentration in the tail flue will also be used as important inputs to the prediction model. Among them, this part of variables mainly makes up for the deficiency of the output accuracy of the concentration in the tail flue caused by the simple mapping form of the pure mechanism model. The final output of the hybrid model will be given by the Transformer prediction model. Since the NO x concentration in the tail flue of the unit always remains stable during actual operation and does not change suddenly, it is considered to use the output of the previous two seconds as the input variable of the hybrid prediction model. The structure of the hybrid prediction model of the SCR denitration system is referred to Figure 4 . The comparison between the output of the SCR hybrid prediction model and the actual value is referred to Figure 5 .

[0097] To verify the advantages of the hybrid prediction model proposed in this study, the prediction accuracies of the Transformer hybrid prediction model, the pure mechanism model, the random forest hybrid model, and the support vector machine hybrid model are compared below respectively.

[0098] Comparison between the hybrid prediction model and the mechanism model:

[0099] In the modeling method currently adopted by a certain power grid company, for the concentration in the tail flue of the power plant unit, simple numerical calculations are used. According to the equipment information, the average value of the NO x concentration at the outlets of the A and B sides is directly taken as the output value of the model. Here, this method is used as the predicted value of the mechanism model for the NO x concentration in the tail flue. The comparison of the output accuracy between the hybrid prediction model and the mechanism model is referred to Figure 6 .

[0100] It can be seen from Figure 6 that the hybrid prediction model fits closely with the actual value in the entire time domain. Especially under steady-state conditions, its prediction error is controlled within ±3mg / Mm 3 within, which is significantly better than the deviation of 10–15mg / Nm of the mechanism model 3 . This shows that the fusion of data-driven and mechanism knowledge effectively improves the modeling accuracy. In the stage of variable working conditions, the actual concentration fluctuates violently due to sudden load changes. Compared with the pure mechanism model, the hybrid model significantly shortens the response delay, and the peak capture accuracy is improved by 40%. This is mainly because the hybrid prediction model comprehensively considers more relevant input variables. The experimental results show that under all working conditions, the prediction accuracy of the Transformer hybrid prediction model for the NO x concentration in the tail flue is significantly better than that of the pure mechanism model.

[0101] Comparison between the Transformer hybrid prediction model, the random forest hybrid prediction model, and the support vector machine hybrid prediction model:

[0102] The main difference between the hybrid prediction model and the data model is that two key variables, the NO concentrations at the flue gas outlets on the A and B sides obtained through the mechanism model, are added to the input variable set. These variables have a very strong correlation with the NO concentration in the tail flue and greatly affect the accuracy of the output prediction. For the comparison of the accuracy of the three hybrid prediction models, refer to x the following. x The concentration has a very strong correlation with the NO concentration in the tail flue and greatly affects the accuracy of the output prediction. For the comparison of the accuracy of the three hybrid prediction models, refer to Figure 7 .

[0103] From Figure 7 it can be seen that all three models can closely fit the actual output for the training set. Among them, the RMSE of the Transformer hybrid prediction model is 0.29426, and the MAPE is 0.50402%. The RMSE of the random forest hybrid prediction model is 0.12298, and the MAPE is 0.10245%. The RMSE of the support vector machine hybrid prediction model is 0.14874, and the MAPE is 0.12218%. Under the steady-state conditions of the test set, the prediction curve of the Transformer hybrid model almost coincides with the true value. Its self-attention mechanism captures the long-term temporal correlations of the ammonia injection rate, system load, and A and B side outputs, etc., and controls the prediction error within ±2mg / Nm 3 . Due to the limitations of their own algorithms, the prediction accuracy of the random forest hybrid model and the support vector machine hybrid model is limited. During the variable operating conditions stage, the Transformer model, relying on the global dependence modeling of historical temporal features, quickly responds to dynamic changes. The peak prediction error is only 4%, significantly better than that of the random forest (error 12%) and SVM (error 20%). Moreover, during the severe disturbance section, the prediction curves of the two deviate seriously from the actual trend. Among them, the RMSE of the Transformer hybrid prediction model is 0.34637, and the MAPE is 0.84222%. The RMSE of the random forest hybrid prediction model is 1.5723, and the MAPE is 2.4672%. The RMSE of the support vector machine hybrid prediction model is 4.9115, and the MAPE is 9.0792%. This result verifies the significant advantages of the Transformer model in applying to hybrid mechanism modeling compared with other data-driven modeling methods.

[0104] As described above, the present invention can be preferably realized.

[0105] The implementation manners of the present invention are not limited by the above embodiments. Any other changes, modifications, substitutions, combinations, and simplifications made without departing from the spirit and principle of the present invention shall be equivalent replacement manners and are all included in the protection scope of the present invention.

Claims

1. A modeling method for a mechanism hybrid prediction model of an SCR denitration system based on Transformer, characterized in that, The modeling method includes the following steps: S1. Extract effective characteristic data from the data collected by on-site units, and at the same time, it is necessary to ensure that the full operating conditions of the units are covered; S2. Construct a reaction kinetic equation through the SCR denitration reaction mechanism, and use the Euler method for solution, retaining unknown model parameters; S3. Select a section of data with relatively stable operating conditions from the collected data, import it into the mechanism model, and find the optimal model parameters through the grey wolf optimization algorithm, that is, obtain the mechanism model of the SCR denitration reaction. The concentrations of the flue gas at the outlets of sides A and B can be calculated from this model; S4. Train the Transformer model to compensate the SCR system mechanism model, improve the output accuracy of the mechanism model under variable operating conditions, and use it as the real-time soft measurement value of the outputs of sides A and B; S5. Select input variables from the full operating condition data, include the concentrations of the flue gas at the outlets of sides A and B obtained by the mechanism model in real time and the outputs of the previous two moments into the input variable set, and perform time delay processing on the input variables; S6. Use the input variable set obtained in S5 to train the Transformer prediction model for the tail flue, and obtain the prediction model for this part; S7. Connect the mechanism models of sides A and B of the flue gas in series with the Transformer model. The Transformer model needs to import the output of the mechanism model in real time and the variables that can be directly measured by sensors. Finally, obtain the hybrid prediction model of the SCR denitration system.

2. The modeling method of the mechanism hybrid prediction model for the SCR denitration system based on Transformer according to claim 1, wherein In step S3, after constructing the reaction kinetic equation according to the Eley-Rideal mechanism, solve the equation by the Euler method, and determine the number of iterative calculations according to the physical model of the reaction vessel and the residence time of the flue gas, so as to perform iterative calculations in the process of calculating the numerical solution.

3. The modeling method of the SCR denitration system mechanism hybrid prediction model based on Transformer according to claim 2, wherein, For the solution method of the differential equation of the mechanism model, the unknown parameters of the mechanism model in this process are retained. Using the historical data actually collected by the power plant units, the grey wolf optimization method is used for identification, and the identified parameters are substituted to obtain the mechanism model.

4. The modeling method of the mechanism hybrid prediction model of the SCR denitration system based on Transformer according to claim 1, wherein, In step S4, the Transformer model is used to compensate the accuracy of the mechanism model.

5. The modeling method of the mechanism hybrid prediction model for the SCR denitration system based on Transformer according to claim 1, characterized in that, In step S5, the selection of the input variable set includes the total set value in the net flue gas mode, the concentration of the flue gas at the inlet of side A of SCR, the ammonia flow rate of side A of SCR, the concentration of the flue gas at the inlet of side B of SCR, the ammonia flow rate of side B of SCR, the total air volume, the operating load, the coal feeding amount of the coal feeder, the flue gas temperature at the inlet of side A of denitration, and the flue gas temperature at the inlet of side B of denitration, and incorporate the concentrations of the flue gas at the outlets of side A of SCR and side B of SCR output by the mechanism model into the variable set.

6. The method for modeling a mechanism hybrid prediction model of a Transformer-based SCR denitration system according to claim 5, wherein, For multi-input systems such as SCR denitration, there is redundancy between some information. The partial mutual information method is used to decouple the input variables.

7. The modeling method of the mechanism hybrid prediction model of the SCR denitration system based on Transformer according to claim 5, characterized in that, There are different degrees of time delays in the input data monitored and recorded by the sensors for the output quantity, that is, the concentration at the outlet of the tail flue. Using the method of mutual information, calculate the MI value between each input variable and the output variable, find the optimal time delay of each variable, and perform data alignment processing.

Images