Sintering flue gas control method based on parameter coupling total differential constraint

Abstract

The application provides a sintering flue gas control method based on parameter coupling full differential constraint, relates to the nitrogen oxide emission reduction regulation and control technical field of steel sintering process flue gas, and the method comprises the following steps: S1, pretreating sintering historical process data, performing variable decoupling and feature generation; S2, training a cascaded differentiable digital twin model for a nitrogen oxide emission process; S3, constructing a nitrogen oxide composite objective function, performing full differential constraint online optimization, and determining an active operation variable regulation scheme; S4, generating a sintering flue gas nitrogen oxide emission control instruction, and performing sintering flue gas nitrogen oxide emission effect verification. The application can predict the nitrogen oxide emission situation in advance and intervene through four-layer process feature variable generation, can construct a nitrogen oxide composite objective function, can output an optimized control variable, can improve the nitrogen oxide emission control precision, and can control the nitrogen oxide emission situation in flue gas.

Description

Technical Field

[0001] This invention relates to the field of nitrogen oxide emission reduction and industrial control technology in the sintering process of steel, specifically to a sintering flue gas control method based on parameter coupling total differential constraint. Background Technology

[0002] The sintering process is one of the most concentrated high-temperature reactions in the steel industry. Changes in the proportions of fuel ore, quicklime, blended ore, return ore, and dust removal ash not only affect the combustion heat release, liquid phase generation, and permeability inside the sintering layer, but also further affect state parameters such as wind box temperature, wind box negative pressure, main exhaust fan inlet pressure, and large flue temperature, ultimately determining the nitrogen oxide emission level in the sintering flue gas.

[0003] Existing methods for predicting nitrogen oxides (NOx) typically employ empirical models, single regression models, or tree models. These methods suffer from three main shortcomings. First, they lack a rigorous distinction between actively controllable variables and passively responding variables, making it difficult to directly apply optimization suggestions to the field operation level. Second, the models usually only focus on the current input, failing to incorporate the NOx evolution trend from previous moments and the lagged effects of key process parameters into the calculation. Third, in the online optimization phase, the prediction model and optimizer are often used separately, making it difficult to integrate boundary constraints, locked equipment, state stability, batching ratio deviations, and physical lower limits into a unified, continuously differentiable optimization framework.

[0004] Therefore, it is necessary to propose an emission reduction method that can utilize the batching coupling mechanism and time-related information, and can directly output an executable adjustment scheme under the conditions of on-site equipment locking, boundary constraints and limitations. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention aims to provide a sintering flue gas control method based on parameter-coupled total differential constraints. This method employs a four-layer feature generation approach, utilizing active operational variables, passive state variables, batching coupling correlation vectors, and historical features to predict nitrogen oxide emissions in advance and intervene proactively. By leveraging data sample quality weights, cascaded consistency joint training, and three-variable integrated prediction, the method improves robustness to control noise, missing data, and operational fluctuations during sintering operations, balancing production stability with nitrogen oxide emissions control in the flue gas. Furthermore, it integrates batching ratio stability, state stability, boundary constraints, and physical lower limit constraints into the optimization function, outputting optimized control variables to improve the accuracy of nitrogen oxide emission control.

[0006] Specifically, the present invention provides a sintering flue gas control method based on parameter coupling total differential constraint, which includes the following steps: S1: Acquire sintering process data, perform preprocessing, variable decoupling and feature generation, and extract active operation variables. Passive state variables Ingredient coupling correlation vector Historical characteristic vector of nitrogen oxide emissions The input feature vector of the sintering process The normalized feature vectors of the sintering process parameters are then obtained by normalization. ; S2: Use Training a cascaded differentiable digital twin model for nitrogen oxide emission processes; training the cascaded dynamic mapping function G and the three-variable holographic prediction function ensemble F to obtain the predicted values ​​of passive state variables after linkage simulation. ;Will Input F to obtain the predicted value of nitrogen oxides at the sintering flue gas inlet. Establish a joint training loss function for sintered nitrogen oxides. Output cascade dynamics mapping parameters Parameters of the three-variable holographic prediction function , , Normalized mapping parameters, eigenvector parameters, and passive state sensitivity weights ; S3: Generate the locking state of the active operation variable and active operation of unbounded optimization variables By establishing a continuous transfer relationship between the active operation variable adjustment, the batching coupling change, the thermal state response, and the nitrogen oxide prediction results, a composite objective function for nitrogen oxides is constructed. Perform online optimization with total differential constraints to obtain the solution. Minimum emission reduction adjustment parameters; Emission reduction adjustment parameters include: actively manipulated variables , sintering thermal passive state variables Ingredient coupling correlation vector Predicted values ​​of nitrogen oxides at the inlet of sintering flue gas after optimization .

[0007] Preferably, S1 specifically includes: S11: Acquire sintering process data and extract active operation variables. and passive state variables Determine the fuel ratio parameters, flux ratio parameters, recycled feed ratio parameters, blending ore distribution deviation parameters, and combustion-melting coupling parameters to form a batching coupling correlation vector. ; S12: Screen out the process parameters associated with the nitrogen oxide concentration at the sintering flue gas inlet to form a historical characteristic vector of nitrogen oxide emissions. Output sintering process input feature vector ; S13: Input feature vector for the sintering process Perform normalization to obtain the normalized eigenvector of the j-th sintering process parameter at time t. .

[0008] Preferably, the ingredient coupling correlation vector in S11 for: ; ; ; ; ; ; in, This is the ingredient coupling correlation vector; Let be the fuel mix ratio parameter at time t; Let be the flux ratio parameter at time t; The circulating material proportioning parameters at time t; Let be the distribution deviation parameter of the homogenized ore at time t; Let be the combustion coupling parameters at time t; For denominator correction parameters; The target percentage of the i-th mixed ore bin obtained from the training set statistics; Let be the total flow rate of the mixing bin at time t; Let be the total flow rate of the fuel ore bin at time t; Let t be the total flow rate of the quicklime silo at time t; t represents the total flow rate of the cold return ore bin at time t; t is the time parameter; i is the mixing ore bin number; Let be the instantaneous flow rate of the i-th mixing bin at time t.

[0009] Preferably, S2 specifically comprises: S21: Train the cascaded dynamic mapping function G, inputting the normalized active operation variables. and the normalized ingredient coupling correlation vector The vector formed Output the predicted values ​​of passive state variables ; S22: Train the multivariate holographic prediction function ensemble F; perform joint consistency fine-tuning on the cascaded dynamics mapping function G and the multivariate holographic prediction function ensemble F to obtain the predicted values ​​of the passive state variables after linkage deduction. ;Will By inputting the multivariate holographic prediction function ensemble F, the predicted value of nitrogen oxides at the sintering flue gas inlet is obtained. ; S23: Determine the loss function for cascade dynamics mapping Direct prediction loss function Chain prediction loss function Establish a joint training loss function for sintered nitrogen oxides. .

[0010] Preferably, the predicted value of nitrogen oxides at the sintering flue gas inlet in S22 is: ; ; in, The predicted value of nitrogen oxides at the sintering flue gas inlet; Let m be the m-th holographic prediction function; For the m-th holographic prediction function Model parameters; For predicted values ​​of passive state variables; Number the holographic prediction function; These are the normalized active operation variables; This is the normalized ingredient coupling correlation vector; For normalized historical characteristics of nitrogen oxide emissions, G represents the cascade kinetic mapping function. These are the model parameters for the cascade dynamics mapping function.

[0011] Preferably, the joint training loss function for sintered nitrides in S23 is: ; ; ; ; in, For the joint training loss function of sintered nitrogen oxides; These are the weighting coefficients for the constraint cascade dynamics mapping function; To constrain the weighting coefficients of the holographic prediction function; The weighting coefficients are used to constrain the chain prediction loss function; The loss function is the cascade dynamics mapping function. The loss function is for direct prediction. The loss function is a chained prediction function. Weights for sample quality; Here, N is the loss function for the regression task; N is the number of training samples. The measured concentration of nitrogen oxides at the inlet of the sintering flue gas after normalization at time t; For predicted values ​​of passive state variables; These are normalized passive state variables; The predicted value of nitrogen oxides after averaging the variable functions; This is the predicted value of nitrogen oxides at the sintering flue gas inlet.

[0012] Preferably, S3 specifically comprises: S31: Generate the active operation variable locking state Actively manipulate unbounded optimization variables Map the data to the historical boundary of the actively manipulated variable, and then force the locked variable to be backfilled with the ground state value; S32: Set unbounded optimization variables for active operation The active operation variables are backfilled based on the unbounded optimization variables of the active operation; cascade kinetic mapping is performed to construct a composite objective function for nitrogen oxides. ; S33: Solve the objective function that causes nitrogen oxides to recombine. The minimum optimization variable yields the optimized active operation variable. , sintering thermal passive state variables , sintering batching coupling correlation vector Predicted values ​​of nitrogen oxides at the inlet of sintering flue gas after optimization .

[0013] Preferably, in S32, the active operation variables are backfilled based on the unbounded optimization variables of the active operation, specifically as follows: ; ; in, This is to backfill the j-th active operation variable after optimization; The j-th active operation variable is locked. These are the initial active operation variables; Let j be the candidate value for the active operation variable; To actively manipulate unbounded optimization variables; This is the historical lower bound of the j-th actively manipulated variable; The historical upper limit of the j-th actively manipulated variable; is the activation function; j is the active operation variable number.

[0014] Preferably, the objective function for nitrogen oxide recombination in S32 is as follows: ; in, The objective function is the composite objective function for nitrogen oxides; Adjust the penalty function for actively manipulated variables; For the passive state variable offset penalty function; This is a penalty function for ingredient ratio deviation. For boundary barrier penalty function; This is a penalty function for preventing penetration outside the domain. The weighting coefficients for the penalty terms of actively manipulated variables; These are the weighting coefficients for passive state variables; The weighting coefficients for the ingredient proportions; These are the weighting coefficients for the boundary barriers; To optimize the predicted value of nitrogen oxides at the inlet of the sintering flue gas.

[0015] Preferably, it also includes S4: verifying the effectiveness of the emission reduction adjustment parameters, specifically: S41: Calculate the adjustment ratio for each active variable to form the first-order active variable control parameters. This is used to quantify the change ratio of the normalized active operating variable relative to the current baseline operating condition after optimization. S42: Calculate the change ratio of each passive state variable to form the second-order state linkage prediction parameters. Output active operation variable adjustment items and passive state variables, verify the sintering condition transfer process, and use whether the linkage deduction value is within the corresponding historical boundary interval as the judgment basis. S43: Output ingredient coupling correlation vector from arrive The results of the changes are used to verify whether the nitrogen oxide emission reduction adjustment parameters of sintering flue gas are effective, and the effective judgment is based on the fact that the predicted value of nitrogen oxides at the inlet of sintering flue gas after optimization is lower than the predicted value of the ground state and the predicted emission reduction is greater than 0.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) The present invention adopts a four-layer feature generation method consisting of active operation variables, passive state variables, batching coupling correlation vector and historical features, which can simultaneously reflect the batching status, thermal response status and short-term evolution trend under the current sintering conditions, predict nitrogen oxide emissions in advance, and intervene in advance.

[0017] (2) This invention utilizes data sample quality weights, cascaded consistency joint training and three-variable integrated prediction to improve the robustness of on-site noise, missing data and operating condition fluctuation control in sintering conditions, making the production process more stable, and controlling nitrogen oxide emissions in flue gas while taking into account production stability.

[0018] (3) This invention integrates the stability of ingredient ratio, state stability, boundary constraints and physical lower limit constraints into a differentiable objective optimization function, and outputs the optimized control variables, avoiding the blindness caused by traditional experience adjustment or discrete search, and improving the accuracy of nitrogen oxide emission control. Attached Figure Description

[0019] Figure 1 The control block diagram is shown for a sintering flue gas control method based on parameter coupling total differential constraints. Figure 2 This is a flowchart of the sintering flue gas control method based on parameter coupling total differential constraint of the present invention; Figure 3 This is a chain diagram of the sintering flue gas control technology of the present invention; Figure 4 This is a comparison chart of the predicted and actual values ​​of the present invention on the test set. Detailed Implementation

[0020] Hereinafter, embodiments of the present invention will be described with reference to the accompanying drawings.

[0021] This invention proposes a sintering flue gas control method based on parameter coupling total differential constraints, such as... Figure 1 As shown, the sintering historical process data is preprocessed, and variable decoupling and feature generation are performed; a cascaded differentiable digital twin model for the nitrogen oxide emission process is trained; a composite objective function for nitrogen oxides is constructed, and online optimization with total differential constraints is performed to determine the active variable control scheme; nitrogen oxide emission control commands for sintering flue gas are generated, and the emission effect of nitrogen oxides in sintering flue gas is verified; such as... Figure 2 The diagram shows the flowchart of the sintering flue gas control method based on parameter coupling total differential constraint of the present invention; it specifically includes the following steps: S1: Preprocessing, variable decoupling, and feature generation of sintering history data. The sintering history data is cleaned, aggregated, missing data is filled in, and variables are decoupled to construct a 15-dimensional vector of active operational variables. 46-dimensional passive state variable vector 5D ingredient coupling correlation vector and 29-dimensional historical feature vector We established sintering process data sample quality weights, historical boundary libraries, and normalized mappings; and performed sintering historical process data preprocessing, variable decoupling, and feature generation.

[0022] The input to S1 is a continuous historical process data table of the sintering machine. It should be noted that the input for this step is the raw field data collected on-site; this raw data is processed through S11 to S13 to form a 95-dimensional feature vector consisting of 15 active operating variables, 46 passive state variables, 5-dimensional batching coupling and correlation vector variables, and 29 historical features. The raw field data includes at least the timestamp, the instantaneous flow rate of the adjustable silo corresponding to the 15 active operating variables listed in S11, the wind box temperature, the wind box negative pressure, the main exhaust fan inlet pressure, the main flue temperature, the flow rate of the secondary mixing hot water supply pipe, the sintering circulating fan inlet pressure detection, and the measured value of nitrogen oxides at the sintering flue gas inlet. S1 first sorts the data according to the timestamp, performs outlier truncation, physical synonym aggregation, and missing value imputation on the numerical columns to prevent abnormal sampling from directly entering the subsequent model.

[0023] S11: The original sintering history data is processed as follows: sorted chronologically, IQR outlier truncation is performed on numerical columns, and mean aggregation is performed on variables with similar physical meanings, such as the temperature of flues #1 and #2; missing values ​​are imputed using the median of the training set. At the variable decoupling level, 15-dimensional actively manipulated variables are output. and 46-dimensional passive state variables .

[0024] 15-dimensional active operation variables in the embodiment The instantaneous flow rates of the specific material silos are as follows: 13# dust removal ash silo #2, 11# carbonaceous dust silo, 1# blended ore silo #1, 6# fuel ore silo #1, 4# blended ore silo #4, 5# blended ore silo #5, 12# dust removal ash silo #1, 2# blended ore silo #2, 9# cold return ore silo #1, 16# quicklime silo #3, 3# blended ore silo #3, 15# quicklime silo #2, 7# fuel ore silo #2, 8# blast furnace return ore silo, and 14# quicklime silo #1.

[0025] 46-dimensional passive state variables in the embodiment At least the following parameters are included: furnace negative pressure of ignition furnace, inlet pressure detection of sintering circulating fan, inlet flow detection of sintering circulating fan, inlet temperature of sintering circulating fan, temperature of each air box, negative pressure of each north side air box, inlet pressure of main exhaust fan, temperature of main flue, flow rate of primary mixing water, flow rate of secondary mixing hot water pipe, temperature of secondary mixing hot water pipe, preheating steam flow rate of mixing tank, setting of sintering machine speed L1, and setting of roller feeder speed ratio.

[0026] 5D ingredient coupling correlation vector in the embodiment The calculation process consists of fuel proportioning parameters, flux proportioning parameters, circulating material proportioning parameters, blended ore distribution deviation parameters, and combustion-flux coupling parameters. Specifically, the instantaneous flow rates of each type of silo are grouped and summed to obtain the total flow rate of the blended ore, the total flow rate of fuel and carbonaceous dust, the total flow rate of quicklime, and the total flow rate of returned ore and dust collector ash. Then, using the total flow rate of the blended ore as a benchmark, the proportioning strength of each material relative to the blended ore is calculated, and the blended ore distribution deviation and fuel-flux coupling strength are further calculated. Specifically: ; ; ; ; in, Let be the total flow rate of the mixing bin at time t; Let be the total flow rate of the fuel ore bin at time t; Let t be the total flow rate of the quicklime silo at time t; Let t be the total flow rate of the cold return ore bin; Let be the instantaneous flow rate of the first mixed ore bin at time t; Let be the instantaneous flow rate of the second mixing bin at time t; Let be the instantaneous flow rate of the third mixing bin at time t; Let be the instantaneous flow rate of the fourth mixing bin at time t; Let be the instantaneous flow rate of the fifth mixing bin at time t; Let be the instantaneous flow rate of the first fuel ore bin at time t; Let be the instantaneous flow rate of the second fuel ore bin at time t; Let be the instantaneous flow rate of the carbon-containing dust bin at time t; Let be the instantaneous flow rate of the first quicklime silo at time t; Let be the instantaneous flow rate of the second quicklime silo at time t; Let be the instantaneous flow rate of the third quicklime silo at time t; The instantaneous flow rate of the blast furnace return ore bin at time t; The instantaneous flow rate of the cold return ore bin at time t; The instantaneous flow rate of the first dust collection silo at time t; Let t be the instantaneous flow rate of the second dust collection silo.

[0027] The above formula is used to classify and summarize the 15-dimensional active operating variables according to the sintering batching properties, providing a basic quantity for subsequent calculation of the batching structure relationship.

[0028] The specific parameters for determining the fuel proportioning, flux proportioning, recycled feed proportioning, blended ore distribution deviation, and combustion-melting coupling are as follows: ; ; ; ; ; in, The fuel ratio parameter at time t is used to characterize the fuel ore and carbon-containing dust input intensity corresponding to a unit of blended ore; The flux ratio parameter at time t is used to characterize the quicklime input intensity corresponding to a unit of blended ore; The circulating material proportioning parameter at time t is used to characterize the reflux intensity of the return ore and dust relative to the mixed ore. The parameter representing the distribution deviation of the mixed ore at time t is used to characterize the degree of deviation of the instantaneous flow rate ratio of the five mixed ore bins from the target ratio. Let be the combustion-melting coupling parameter at time t, used to characterize the coordinated change relationship between fuel input intensity and flux input intensity; For denominator correction parameters; The target percentage of the i-th mixed ore bin is obtained from the statistics of the training set.

[0029] Composition of 5-dimensional batching coupling correlation vector for: ; in, The 5-dimensional batching coupling correlation vector at time t is one of the outputs of S1. In S2, it is input into the digital twin model along with the active operating variables, passive state variables, and historical features. In S3, it is recalculated with the optimization results of the active operating variables to constrain and explain the changes in the batching structure during the optimization process of nitrogen oxide emissions from sintering flue gas.

[0030] The fuel ratio parameters in the above formula The parameters reflect the fuel input intensity corresponding to a unit of blended ore, the flux ratio parameter reflects the quicklime input intensity corresponding to a unit of blended ore, the circulating material ratio parameter reflects the return ore and dust ash return intensity, the blended ore distribution deviation parameter reflects the dispersion of ore blending in the five blended ore bins, and the combustion-melting coupling parameter reflects the intensity of the combined effect of fuel and flux.

[0031] S12: From the current 66-dimensional process characteristics, the 12 process parameters with the highest absolute correlation coefficient with the nitrogen oxide concentration at the sintering flue gas inlet are selected, and their first-order and second-order hysteresis values ​​are constructed; at the same time, the nitrogen oxide concentration values ​​at the sintering flue gas inlet of the first 3 sampling times, as well as the moving average values ​​of nitrogen oxides at the first 3 and first 5 points, are added to form a 29-dimensional historical feature vector of nitrogen oxide emissions. .

[0032] The 29-dimensional nitrogen oxide emission history feature vector in the example This historical feature vector is used to characterize the short-term evolution trend of nitrogen oxide concentration at the sintering flue gas inlet before the current operating conditions, as well as the lagged effects of highly correlated process parameters on nitrogen oxide emissions. The vector consists of nitrogen oxide concentration values ​​at the first three sampling times, the moving averages of nitrogen oxides at the first three and five sampling times, and first- and second-order lag terms for 12 highly correlated process parameters.

[0033] Twelve highly relevant process parameters were selected from the following variables, ranked according to their absolute correlation coefficients with the nitrogen oxide concentration at the sintering flue gas inlet: instantaneous flow rate of silo #2 for cold return ore (10#), inlet pressure detection of the sintering circulating fan, instantaneous flow rate of silo #4 for blending ore (4#), instantaneous flow rate of silo #1 for cold return ore (9#), instantaneous flow rate of silo #1 for dust collector ash (12#), instantaneous flow rate of silo #3 for blending ore (3#), temperature of the secondary mixing hot water supply pipe, temperature of the #17 wind box, preheating steam flow rate of the mixing tank, instantaneous flow rate of silo #2 for blending ore (2#), flux ratio parameters, and circulating material ratio parameters. The expression for the historical characteristic vector of nitrogen oxide emissions is: ; ; in, Let be the historical characteristic vector of nitrogen oxide emissions at time t; The concentration of nitrogen oxides at the inlet of the sintering flue gas at the previous sampling time; The concentration of nitrogen oxides at the inlet of the sintering flue gas is the concentration at the first two sampling times. The concentration of nitrogen oxides at the inlet of the sintering flue gas is the value at the first three sampling times. and , respectively, are the lag values ​​of the j-th selected process parameter at the first and second sampling times. The selected process parameters are the top 12 process parameters selected in descending order of their absolute correlation coefficient with the nitrogen oxide concentration at the sintering flue gas inlet, j=1,2,...,12; w is the width of the moving average window.

[0034] Therefore, the final output of S1 is the 95-dimensional sintering process input feature vector. for: ; in, Input the feature vector for the 95-dimensional sintering process at time t; Let be the active variable at time t; Let be the passive state variable at time t; Let be the ingredient coupling correlation vector at time t.

[0035] In this embodiment, the sintering process inputs a feature vector. The dimension is 95, of which 15 are active operation variables, 46 are passive state variables, 5 are ingredient coupling correlation vectors, and 29 are historical feature vectors.

[0036] To improve the performance of identifying low-quality samples during the training phase, a quality weight is calculated for each sintering process input feature vector sample, specifically as follows: ; ; ; ; in, The median of the training set for the j-th process parameter; Let be the interquartile range of the training set for the j-th process parameter; Let be the standard deviation of the training set for the j-th process parameter; This is a truncation function; The degree of missing input features in the sintering process at time t; The degree of deviation of the input features of the sintering process at time t from the statistical distribution of the training set; The transition degree of the input feature of the sintering process at time t relative to the previous sampling time; This is an indicator function for determining the absence of sintering process features; This is a function that takes the maximum value of multiple operating condition deviation indicators; Let be the sample quality weight at time t.

[0037] Output sample quality weights It is used in S2 for weighted training loss.

[0038] To control the physical boundaries during the online optimization phase, a historical boundary database B for each process parameter is established as follows: ; in, This is the historical boundary library for the j-th process parameter; The 5th percentile of the j-th process parameter; It is the 95th percentile of the j-th process parameter.

[0039] The output boundary library B will be used in S3 to constrain the optimization range of active operation variables and passive state variables.

[0040] S13: Input feature vector for the 95-dimensional sintering process Perform Min-Max normalization, specifically as follows: ; in, Let be the normalized eigenvector of the process parameters for the j-th sintering process at time t; The maximum value of all training samples for the j-th sintering process parameter; The minimum value of all training samples for the j-th sintering process parameter; The output of this step is: a 95-dimensional sintering process input feature vector. Normalized eigenvectors Sintering history process sample quality weighting Target share of mineral blending And the historical boundary library of process parameters B; among which, the normalized feature vector and sample quality weights It will be used in S2 for training cascaded differentiable digital twin models; target share of ore blending. The historical boundary library of process parameters, B, will be used for online optimization in S3.

[0041] S2: Training a cascaded differentiable digital twin model for NOx emission processes. The 95-dimensional sintering process output from S1 is input into the feature vector. The input includes: a 15-dimensional vector of active operation variables. 46-dimensional passive state variable vector 5. Ingredient Coupling Relationship Vector and 29-dimensional historical feature vector The training of cascaded kinetic mapping functions and holographic prediction functions is related to the NOx emissions from sintering flue gas as follows: after the normalized active operating variables are changed, they do not directly affect NOx as isolated variables. Instead, they first change the feed coupling structure and the 46-dimensional passive state, and then affect the combustion intensity, redox reaction conditions, and NOx generation, migration, and emission through wind box temperature, negative pressure, exhaust pressure, and moisture conditions. A differentiable digital twin model is obtained for online simulation, and cascaded digital twin model training is performed.

[0042] S21: First, train the cascaded dynamic mapping function G, whose input is a 20-dimensional vector. The output is the predicted value of the 46-dimensional passive state variable. The role of the cascaded dynamic mapping function G is to first transform the normalized active operating variables and the coupled changes in batching into state changes such as the air box temperature, air box negative pressure, and main exhaust fan inlet pressure.

[0043] S22: Training the ensemble of three-variable holographic prediction functions F. The input to each variable function is a 95-dimensional vector. The output is a one-dimensional predicted value of nitrogen oxides; the average of the three variables is used as the final prediction result. In this way, this step utilizes both the current normalized active operation variables and passive state variables, as well as historical evolution information, specifically: ; ; in, For the first A holographic prediction function; For the first Model parameters for a holographic prediction function; These are the normalized active operation variables, which are 15-dimensional variables. The passive state variables are normalized and are 46-dimensional variables; The normalized ingredient coupling correlation vector is a 5-dimensional variable. The normalized historical characteristic variables of nitrogen oxide emissions are 29-dimensional variables. For the first Predicted value of nitrogen oxides at the sintering flue gas inlet output by a function with multiple variables; The predicted value of nitrogen oxides after averaging the variable functions; Number the holographic prediction function.

[0044] Joint consistency fine-tuning is performed on the cascaded dynamic mapping function G and the holographic prediction function ensemble F. The normalized active operation variable output by S1 is then used. Coupled with normalized ingredient correlation vector By inputting the cascade dynamic mapping function G, the passive state variables obtained after the linkage derivation are: ; Where G is the cascade dynamics mapping function; These are the model parameters for the cascade dynamics mapping function; The passive state variable prediction value is a 46-dimensional passive state variable prediction value derived from the normalized active operating variable and the batching coupling correlation vector. This variable corresponds to the changes in thermal state such as wind box temperature, wind box negative pressure, and main exhaust fan inlet pressure during the sintering process. It is used to characterize the intermediate state that affects the nitrogen oxide emission process of sintering flue gas after the adjustment of the normalized active operating variable.

[0045] Normalize active operation variables Predicted values ​​of passive state variables Normalized ingredient coupling correlation vector and normalized historical characteristic variables of nitrogen oxide emissions By inputting the holographic prediction function integration F, the predicted value of nitrogen oxides at the sintering flue gas inlet is obtained as follows: ; in, The predicted value of nitrogen oxides at the sintering flue gas inlet is represented by the chain deduction of active manipulated variables - passive state variables - predicted nitrogen oxides value; The m-th holographic prediction function is used to predict the nitrogen oxide concentration at the sintering flue gas inlet based on active operating variables, passive state variables, batching coupling variables, and historical characteristics. For the m-th holographic prediction function The model parameters are m=1,2,3.

[0046] Therefore, the passive state variables normalized by the output of the cascade dynamic mapping function G It is explicitly used as the input to the subsequent holographic prediction function F, rather than remaining as an independent intermediate quantity.

[0047] S23: Construct the joint training loss function. Determine the cascaded dynamics mapping loss function, the direct prediction loss function, and the chained prediction loss function as follows: ; ; ; in, This is the loss function for the cascaded dynamics mapping, used to constrain the accuracy of the cascaded dynamics mapping function in extrapolating the 46-dimensional passive state variables; The direct prediction loss function is used to constrain the prediction accuracy of the holographic prediction function for nitrogen oxides under real passive state variable inputs. This is a chain prediction loss function used to constrain the chain consistency of predicting nitrogen oxides after deriving passive state variables from active operating variables. The sample quality weights obtained in step S1; The loss function for the regression task is the Huber loss function; N is the number of training samples. The measured concentration of nitrogen oxides at the inlet of the sintering flue gas after normalization at time t.

[0048] The final joint training loss function for sintered nitride oxides is: ; in, For the joint training loss function of sintered nitrogen oxides; To constrain the weighting coefficients of the cascade dynamics mapping function, the value in this embodiment is 0.65; To constrain the weighting coefficients of the holographic prediction function, the value in this embodiment is 0.25; The weighting coefficients for the chain prediction loss function are constrained; in this example, the value is 0.90.

[0049] Through the above steps, S2 forms a normalized active operation variable. To sintering thermal passive state variables Then, through chain-like deduction, the predicted value of nitrogen oxides at the sintering flue gas inlet is calculated. Continuously differentiable computational chains, such as Figure 3 The diagram shown is a chain diagram of the sintering flue gas control technology of the present invention; it will be used as a digital twin model for online optimization in S3.

[0050] In this embodiment, the network hierarchy of the cascaded dynamic mapping function G is 20->96->128->96->46, and the univariate network hierarchy of F is 95->192->160->96->1. Each hidden layer uses the Mish activation function and LayerNorm layers. The pre-training rounds are 220, and the joint consistency fine-tuning rounds are 140.

[0051] The output of this step is: cascade dynamics mapping parameters. Parameters of the three-variable holographic prediction function , , Normalized mapping parameters, eigenvector parameters, and passive state sensitivity weights ,in, , , , and All of these are model parameters generated in step S2, and they correspond one-to-one with the model parameter symbols with corresponding subscripts in the formula above. These parameters will be used for online optimization in step S3.

[0052] S3: Through online optimization with full differential constraints, the current sintering condition is transformed into an executable normalized active manipulated variable control scheme. This step does not perform mathematical optimization solely on the predicted nitrogen oxide value, but rather uses the differentiable digital twin model obtained in S2 as the computational chain to link changes in the normalized active manipulated variable, changes in the batching coupling structure, changes in the 46-dimensional passive state, and changes in the nitrogen oxide concentration at the sintering flue gas inlet. Specifically, the ground-state active manipulated variable under the current condition... Represents the current instantaneous flow rate of each adjustable silo, and the ground-state passive state variable. This represents the current thermal states, including the bellows temperature, bellows negative pressure, main exhaust fan inlet pressure, main flue temperature, and secondary mixing hot water supply pipe flow rate. The ground-state correlation vector represents these parameters. This represents the current fuel ratio, flux ratio, recycled feed ratio, mixed ore distribution deviation, and combustion-melting coupling relationship; historical feature vector. This indicates the short-term trend of NOx emissions and the lagged effects of highly relevant process parameters. These variables collectively determine the current combustion intensity, permeability, ventilation status, and flue gas generation environment of the sintering bed, and further influence the NOx concentration at the sintering flue gas inlet.

[0053] In S3, the initial 15-dimensional active operation variable vector under the current working condition is... Initial 46-dimensional passive state variable vector Initial 5-dimensional ingredient coupling correlation vector and the initial 29-dimensional historical feature vector The differentiable digital twin model obtained from input S2 is used for constrained total differential online optimization, combining active operation unbounded optimization variables, historical boundary libraries, and composite objective functions. During the optimization process, the optimizer first adjusts the permissible active operation variables, then recalculates the batching coupling correlation vector based on the adjusted active operation variables, and then uses a cascaded dynamic mapping function to deduce the linked changes of passive state variables such as wind box temperature, wind box negative pressure, and main exhaust fan inlet pressure. Finally, the optimized predicted value of sintering flue gas inlet NOx is obtained through a holographic prediction function. Thus, S3 connects the adjustment of normalized active operation variables, changes in batching structure, thermal state response, and changes in NOx emissions into a continuously differentiable online optimization process. The input is the current operating condition ground state. , , , Historical boundary library B, cascade dynamics mapping parameters Holographic prediction parameters Control cost weight and state sensitivity weights We perform online optimization using restricted total differentials.

[0054] S31: First, generate the active operation variable locking state based on the on-site locking device. If the hopper or equipment does not allow adjustment, then lock the active operation variable. If adjustment is allowed, then lock the active operation variable. Then, the unbounded optimization variables will be actively manipulated. Mapping to the historical boundary of the actively manipulated variable, the locked variable is then forcibly refilled with its ground state value. Therefore, the j-th actively manipulated variable is output. , It will continue to be used, and the output will be backfilled with optimized active operation variables. The data was reused during the meeting to calculate the new ingredient coupling correlation vector. .

[0055] After obtaining the active operation variables after backfill optimization and Then, step S3 utilizes the calculation of linkage state variables. Then, the calculated and optimized predicted values ​​of nitrogen oxides are used. Here, we must particularly emphasize: historical characteristics. The time series information observed before the optimization occurred remains unchanged in S3, serving only as the exogenous background input for the current optimization; while the actively manipulated variables... Ingredient coupling correlation vector and linked state variables The optimizer then updates the settings in real time.

[0056] S32: Construct the composite objective function for nitrogen oxides. Establish an online optimization computation chain. First, for each adjustable active operation unbounded optimization variable... The Sigmoid activation function is used to map the data to the historical range. This ensures that the instantaneous flow rate of the silo generated during the optimization process does not exceed the historical boundary. Specifically: ; in, For optimization variables The candidate value of the j-th active operation variable obtained through mapping; To actively manipulate unbounded optimization variables; This is the historical lower bound of the j-th actively manipulated variable; The historical upper limit of the j-th actively manipulated variable; This is the activation function.

[0057] The active operation variables are backfilled based on the unbounded optimization variables of the active operation, specifically as follows: ; in, This is to backfill the j-th active operation variable after optimization; For the j-th active operation variable, the locking state is when When the value is locked, it indicates that the variable is locked and will retain its base state value after optimization. ,when When the value is set to 0, it indicates that the variable is adjustable, and the optimized result is obtained using boundary mapping. ; These are the initial active operation variables.

[0058] The optimizer will not change silos that are prohibited from being adjusted on the production floor, utilizing the active operation variables optimized by backfilling. The recalculated ingredient coupling correlation vector is as follows: ; in, This is the process of recalculating the fuel ratio parameters, flux ratio parameters, circulating material ratio parameters, mixed ore distribution deviation parameters, and combustion-melting coupling parameters using the formula for calculating the batching coupling correlation vector in step S1. This is the ingredient coupling correlation vector; This is to backfill the optimized active operation variables.

[0059] Backfill the optimized active operation variables Coupled with the updated ingredient association vector Input the cascade dynamics mapping function obtained in step S2, and deduce the passive state variables as follows: ; in, The passive state variables of sintering thermal performance under sintering conditions at time t include the on-site measured thermal parameters such as the temperature of each air box of the sintering machine, the negative pressure of the air box, and the thickness of the material layer. This is a sintering batching, i.e., a thermal cascade kinetic mapping model; The complete set of process model parameters corresponding to the cascade dynamics mapping model G; For time t, the normalized active operation variable; Let t be the feature vector of the batching coupling correlation vector.

[0060] Finally, normalize the active operation variables. , sintering thermal passive state variables Ingredient coupling correlation vector and historical feature vectors Input the holographic prediction function to obtain the optimized nitrogen oxide prediction value, specifically: ; in, The optimized predicted value of nitrogen oxides at the sintering flue gas inlet; For holographic prediction functions; These are the parameters of the holographic prediction function; The feature vector of historical sintering conditions at time t is used to store historical monitoring data of sintering machine thermal performance, material batching, and flue gas over multiple time periods. The passive state variables of sintering thermal at time t include measured thermal parameters such as wind box temperature, material layer thickness, and negative pressure at the tail of the machine.

[0061] Therefore, the active operation variable adjustment, batching coupling changes, thermal state response, and NOx prediction results form a continuous transmission relationship, specifically: ; in, Let be the objective function for the composite of nitrogen oxides, and the optimization objective is to make Minimum This represents the optimized predicted value of nitrogen oxides, used to reflect emission reduction targets; Adjust the penalty function for actively manipulated variables; For the passive state variable offset penalty function; This is a penalty function for ingredient ratio deviation. For boundary barrier penalty function; This is a penalty function for preventing penetration outside the domain. The weighting coefficients for the penalty terms of actively manipulated variables; This is the weighting coefficient for the passive state variable; in this example, it is set to 0.12. This is the weighting coefficient for the ingredient ratio; in this example, it is set to 0.10. This is the weighting coefficient for the boundary barrier; in this example, it is set to 0.015.

[0062] The active operation variable adjustment penalty function is used to limit excessively large instantaneous flow rate adjustments in the silo, making the control scheme feasible on-site. Specifically: ; in, The control cost weight for the j-th active operation variable; Let be the current ground state value of the j-th active operation variable.

[0063] The passive state variable offset penalty function is used to limit excessive jumps in thermal states such as bellows temperature, bellows negative pressure, and main exhaust fan inlet pressure. Specifically: ; in, The sensitivity weight of the k-th passive state variable to nitrogen oxide concentration; This represents the current ground state value of the k-th passive state variable; This is the historical lower bound of the k-th passive state variable; This represents the historical upper limit of the k-th passive state variable.

[0064] The ingredient ratio offset penalty function is: ; in, This is the optimized fuel ratio and feeder coupling correlation vector; The fuel blending ratio is the current ground-state feed coupling correlation vector; The optimized flux ratio and batching coupling correlation vector; The flux ratio is the current ground-state batching coupling correlation vector; The optimized cyclic material proportioning and batching coupling correlation vector; The current ground state batching coupling correlation vector for the cyclic material proportion; The optimized batching coupling correlation vector represents the distribution deviation of the blended ore. The current ground-state batching coupling correlation vector represents the distribution deviation of the blended ore. The optimized combustion coupling strength batching coupling correlation vector; The current ground-state batching coupling correlation vector represents the combustion-melting coupling strength.

[0065] The boundary barrier penalty function is used to prevent the optimization result from getting too close to the boundary and to prevent the control command from being in an extreme state. Specifically: ; ; in, This represents the normalized position of the j-th active operation variable within the historical interval; It is the natural logarithm function.

[0066] The out-of-domain anti-penetration penalty function is used to prevent the model from giving nitrogen oxide predictions below the physically reasonable range. Specifically: ; Where k is the extraterritorial penalty coefficient; It is a linear rectified function; The physical lower limit for predicted nitrogen oxide values ​​can be determined by taking the historical minimum nitrogen oxide value and 90 mg / m³. 3 The larger of the two.

[0067] S33: Solving the objective function that causes nitrogen oxides to recombine using the quasi-Newton L-BFGS algorithm. Minimum optimization variable The final optimized version is obtained. , , and L-BFGS, short for Limited-memory Broyden-Fletcher-Goldfarb-Shanno, is a finite-memory version of the Quasi-Newton method, primarily used for unconstrained optimization problems of high-dimensional smooth functions. It utilizes automatic differentiation to determine the gradient of the current iteration, specifically: ; in, For the k-th iteration, the gradient vector obtained by solving the composite objective function for controlling nitrogen oxide emissions from sintering flue gas with respect to the normalized active operating variables is used to characterize the degree of influence of each batching control parameter on the NOx emission reduction target. For partial derivative solving operators, it represents taking the partial derivative with respect to the sintering batch flow rate optimization variable θ, which is used to quantify the change in the denitrification target corresponding to the adjustment range of each set of normalized active operation variables. It is a composite objective function; Let be the optimization variable for the k-th iteration.

[0068] The descent direction is determined and the optimization variables are updated based on the approximate inverse Hessian matrix using the quasi-Newton L-BFGS algorithm. ; in, Let be the gradient vector of the normalized active operation variable optimization variable corresponding to the objective function of the composite regulation of nitrogen oxides in sintering flue gas under the k-th iteration, which characterizes the degree of influence of the adjustment of the flow rate of each ingredient on the NOx emission index; To adapt to the online optimization scenario of sintering batching, the L-BFGS algorithm approximates the inverse Hessian matrix, which is used to quickly solve the normalized active operation variable to control the optimal iteration direction; The optimization direction for the parameters of each normalized active operation variable in this iteration of sintering is determined; The linear search step size is used to normalize the single adjustment magnitude of the active operation variable.

[0069] The inverse Hessian matrix is ​​updated based on information from two consecutive iterations: ; ; in, The difference vector of the active operation variable optimization variable before and after the k-th iteration is the sintering normalization before and after the k-th iteration; This is the gradient difference vector of the denitrification composite objective function before and after the k-th iteration; It is the identity matrix; To adapt to the iterative scaling factor for online optimization of sintering batching; This is the parameter for correcting the denominator.

[0070] The optimal normalized active operation variable and optimization variable are obtained through the above iterations. And further, the optimized normalized active operation variables are obtained. , sintering thermal passive state variables Ingredient coupling correlation vector Predicted values ​​of nitrogen oxides at the inlet of sintering flue gas after optimization The output of S3 is then fed into S4 for further analysis.

[0071] This invention uses the first embodiment to illustrate the model training and prediction effects of S2 and S3. Taking the continuous production data of a sintering machine in a steel enterprise as an example, after performing S1 on the original data, 95-dimensional input features are obtained, including 15 dimensions of active operation variables, 46 dimensions of passive state variables, 5 dimensions of batching coupling correlation vector, and 29 dimensions of historical features. The model is trained and evaluated using training, validation, and test sets. The results after training are as follows: Direct prediction result: R 2 =0.8465, MAE=10.8760mg / m 3 RMSE = 14.5433 mg / m 3 .

[0072] Cascaded chain deduction results: R 2 =0.8434, MAE=11.0772mg / m 3 RMSE = 14.6884 mg / m³ 3 .

[0073] The validation loss for the cascade dynamics model is 0.009899, ​​the validation loss for the holographic prediction model is 0.002271, and the validation loss for the joint consistency fine-tuning is 0.008599.

[0074] The above results show that the ingredient coupling variables and historical features output in step S1 are effectively utilized in step S2; the linkage state variables output in step S2 are then used for chain optimization in step S3; the entire technical solution forms a closed loop that runs through the training and optimization layers.

[0075] S4: Generate nitrogen oxide emission control commands for sintering flue gas and verify the effectiveness of nitrogen oxide emission control for sintering flue gas; based on the output of S3... , , and The system generates first-order active operation variable control parameters, second-order state linkage prediction parameters, related index change results, and emission reduction results to guide the online emission reduction control of the sintering machine; the input of S4 is the output of S3. , , and .

[0076] S41: Calculate the adjustment ratio for each active operated variable, remove terms with an absolute adjustment ratio not exceeding 0.1%, and sort them in descending order of absolute adjustment ratio to form the first-order active operated variable control parameters. The first-order active operated variable control parameters are calculated according to the following formula: ; in, is the control parameter for the j-th first-order active operating variable, used to quantize the change ratio of the normalized active operating variable relative to the current baseline operating condition.

[0077] Table 1 shows the first-order instruction output results when 15 feature columns of data are input in this invention.

[0078] S42: Calculate the change ratios of 46 passive state variables to form second-order state linkage prediction parameters. The historical boundary intervals are also given for each state variable in this table. The current ground state value, the linked projection value, and the change ratio are used to verify the sintering condition transition process. The second-order state linkage prediction parameters are calculated according to the following formulas: ; in, , is the second-order state linkage prediction parameter of the kth term, used to characterize the linkage change magnitude of the passive state variable of sintering thermal operation relative to the baseline condition after optimization. Let k be the passive state variable of sintering thermal operation after optimization at time t; Let k be the passive state variable of sintering thermal operation under the current reference operating conditions at time t.

[0079] Table 2 shows the actual data of the second-order derivation of total differential optimization when 15 feature columns are input in this invention.

[0080] And only output The active operation variable adjustment items, all 46 passive state variables, are related to their corresponding historical intervals. They are output together for operational condition verification.

[0081] S43: Output 5 ingredient coupling correlation vectors from arrive The results of these changes are used to verify the effectiveness of the nitrogen oxide emission reduction adjustment parameters in sintering flue gas. Therefore, S4 outputs not only how the parameters are adjusted, but also how the feed structure changes and how the state chain responds after the adjustment.

[0082] In this embodiment, the verification criteria for the emission reduction adjustment parameters include: the optimized predicted value of nitrogen oxides at the sintering flue gas inlet is lower than the predicted value of nitrogen oxides in the ground state, and the predicted emission reduction is greater than 0; simultaneously, all passive state variables corresponding to the second-order state linkage prediction parameters are within the historical boundary range, and the change in the batching coupling correlation vector does not trigger the proportion offset penalty constraint. When the above conditions are met, it is determined that the emission reduction adjustment parameters are effective.

[0083] The second embodiment of this invention proposes an online optimization example for high-emission scenarios. A high-emission operating condition sample A from historical data is selected as the example. For example... Figure 4 The figure shown is a comparison of the predicted and actual values ​​of the present invention on the test set. The measured value of nitrogen oxides at the sintering flue gas inlet of this data sample is 305.156 mg / m³. 3 Before inputting the current operating conditions of sample A into step S3, step S1 first constructs historical features based on the NOx concentrations of the previous three time points and the hysteresis values ​​of 12 highly relevant process quantities for that sample. Step S1 calculates the ground-state batching coupling correlation vector based on the current 15 active operating variables. Under this operating condition, the model's forward calculation yielded a predicted ground-state nitrogen oxide value of 304.08 mg / m³. 3Subsequently, constrained total differential optimization was performed in step S3, yielding an optimized predicted value of 139.29 mg / m³ for nitrogen oxides. 3 The model predicts an emission reduction of 164.79 mg / m³. 3 For the first-order active operating variable parameters corresponding to sample A, the core adjustment recommendations for the first five items are as follows: 1.9# Cold Return Ore 1# Warehouse Instantaneous Flow Rate: Adjusted from 8.43t / h to 16.00t / h, an adjustment of +89.83%.

[0084] 2.11# Instantaneous flow rate of carbon dust silo: adjusted from 1.23t / h to 1.06t / h, with an adjustment ratio of -13.67%.

[0085] Instantaneous flow rate of No. 5 bin for 3.5# blended ore: adjusted from 45.84t / h to 40.18t / h, an adjustment ratio of -12.36%.

[0086] Instantaneous flow rate of ore return bin for blast furnace #4.8: adjusted from 20.98t / h to 18.57t / h, an adjustment of -11.48%.

[0087] 5.13# Dust Collector 2# Silo Instantaneous Flow Rate: Adjusted from 4.11t / h to 3.71t / h, with an adjustment ratio of -9.66%.

[0088] The changes in the five ingredient coupling correlation vectors corresponding to sample A are as follows: 1. Fuel ratio parameter: adjusted from 0.0475 to 0.0486.

[0089] 2. Flux ratio parameter: adjusted from 0.0708 to 0.0731.

[0090] 3. The ratio of recycled material is adjusted from 0.1155 to 0.1333.

[0091] 4. Mixed ore distribution deviation parameter: adjusted from 0.0122 to 0.0183.

[0092] 5. Combustion coupling parameter: adjusted from 0.0034 to 0.0036.

[0093] The following are the results of the linkage deduction of some second-order passive state variables corresponding to sample A: 1. Sintering circulating fan inlet pressure detection: adjusted from -1.48 to -4.10.

[0094] 2. The flow rate of the hot water supply pipe for the two mixing tanks has been adjusted from 2.47 to 3.26.

[0095] 3.18# North side air box negative pressure: adjusted from -12.38 to -8.28.

[0096] 4.17# North side air box negative pressure: adjusted from -15.74 to -11.31.

[0097] 5.5# bellows temperature: adjusted from 86.60 to 104.62.

[0098] In this embodiment, the output of step S3 is Recalculated using the formula And got Finally obtained There is a transitive relationship between the ingredient coupling variables, linked state variables, and the final emission reduction results, rather than multiple isolated models. Meanwhile, because S3 simultaneously enabled active operational variable boundary mapping, state boundary truncation, proportion offset penalty, and out-of-domain penetration prevention penalty during the optimization process, all 46 passive state variables of sample A ultimately remained within their corresponding historical intervals.

[0099] The beneficial effects of this invention are as follows: This invention employs a four-layer feature generation method, consisting of active operational variables, passive state variables, batching coupling correlation vectors, and historical features, which can simultaneously reflect the current batching state, thermal response state, and short-term evolution trend. By utilizing sample quality weights, cascaded consistency joint training, and three-variable ensemble prediction, robustness to on-site noise, missing data, and operating condition fluctuations is improved. Furthermore, the stability of batching ratios, state stability, boundary constraints, and physical lower bound constraints are uniformly incorporated into a differentiable objective function, avoiding the blindness caused by traditional empirical adjustment or discrete search. On the dataset corresponding to this embodiment, the model's direct prediction result reaches R0. 2 =0.8465, MAE=10.8760mg / m 3 RMSE = 14.5433 mg / m 3 The cascaded chain deduction results reached R 2 =0.8434, MAE=11.0772mg / m 3 RMSE = 14.6884 mg / m 3 .

[0100] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A sintering flue gas control method based on parameter coupling total differential constraint, characterized in that: It includes: S1: Acquire sintering process data, perform preprocessing, variable decoupling and feature generation, and extract active operation variables. Passive state variables Ingredient coupling correlation vector Historical characteristic vector of nitrogen oxide emissions The input feature vector of the sintering process The normalized feature vectors of the sintering process parameters are then obtained by normalization. ; S2: Use Training a cascaded differentiable digital twin model for nitrogen oxide emission processes; training the cascaded dynamic mapping function G and the three-variable holographic prediction function ensemble F to obtain the predicted values ​​of passive state variables after linkage simulation. ;Will Input F to obtain the predicted value of nitrogen oxides at the sintering flue gas inlet. ; Establish a joint training loss function for sintered nitrous oxide ; Output cascade dynamics mapping parameters Parameters of the three-variable holographic prediction function , , Normalized mapping parameters, eigenvector parameters, and passive state sensitivity weights ; S3: Generate the locking state of the active operation variable and active operation of unbounded optimization variables By establishing a continuous transfer relationship between the active operation variable adjustment, the batching coupling change, the thermal state response, and the nitrogen oxide prediction results, a composite objective function for nitrogen oxides is constructed. Perform online optimization with total differential constraints to obtain the solution. Minimum emission reduction adjustment parameters; Emission reduction adjustment parameters include: actively manipulated variables , sintering thermal passive state variables Ingredient coupling correlation vector Predicted values ​​of nitrogen oxides at the inlet of sintering flue gas after optimization .

2. The sintering flue gas control method based on parameter coupling total differential constraint according to claim 1, characterized in that: Specifically, S1 is as follows: S11: Acquire sintering process data and extract active operation variables. and passive state variables ; Determine the fuel ratio parameters, flux ratio parameters, recycled feed ratio parameters, blended ore distribution deviation parameters, and combustion-melting coupling parameters to form a batching coupling correlation vector. ; S12: Screen out the process parameters associated with the nitrogen oxide concentration at the sintering flue gas inlet to form a historical characteristic vector of nitrogen oxide emissions. Output sintering process input feature vector ; S13: Input feature vector for the sintering process Perform normalization to obtain the normalized eigenvector of the j-th sintering process parameter at time t. .

3. The sintering flue gas control method based on parameter coupling total differential constraint according to claim 2, characterized in that: S11 Ingredient Coupling Relationship Vector for: ； ； ； ； ； ； in, This is the ingredient coupling correlation vector; Let be the fuel mix ratio parameter at time t; Let be the flux ratio parameter at time t; The circulating material proportioning parameters at time t; Let be the distribution deviation parameter of the homogenized ore at time t; Let be the combustion-fuselage coupling parameters at time t; For denominator correction parameters; The target percentage of the i-th mixed ore bin obtained from the training set statistics; Let be the total flow rate of the mixing bin at time t; Let be the total flow rate of the fuel ore bin at time t; Let t be the total flow rate of the quicklime silo at time t; t represents the total flow rate of the cold return ore bin at time t; t is the time parameter; i is the mixing ore bin number; Let be the instantaneous flow rate of the i-th mixing bin at time t.

4. The sintering flue gas control method based on parameter coupling total differential constraint according to claim 1, characterized in that: S2 specifically refers to: S21: Train the cascaded dynamic mapping function G, inputting the normalized active operation variables. and the normalized ingredient coupling correlation vector The vector formed Output the predicted values ​​of passive state variables ; S22: Train the multivariate holographic prediction function ensemble F; perform joint consistency fine-tuning on the cascaded dynamics mapping function G and the multivariate holographic prediction function ensemble F to obtain the predicted values ​​of the passive state variables after linkage deduction. ;Will By inputting the multivariate holographic prediction function ensemble F, the predicted value of nitrogen oxides at the sintering flue gas inlet is obtained. ; S23: Determine the loss function for cascade dynamics mapping Direct prediction loss function Chain prediction loss function Establish a joint training loss function for sintered nitrogen oxides. .

5. The sintering flue gas control method based on parameter coupling total differential constraint according to claim 4, characterized in that: The predicted value of nitrogen oxides at the sintering flue gas inlet in S22 is: ； ； in, The predicted value of nitrogen oxides at the sintering flue gas inlet; Let m be the m-th holographic prediction function; For the m-th holographic prediction function Model parameters; For predicted values ​​of passive state variables; Number the holographic prediction function; These are the normalized active operation variables; This is the normalized ingredient coupling correlation vector; For normalized historical characteristics of nitrogen oxide emissions, G represents the cascade kinetic mapping function. These are the model parameters for the cascade dynamics mapping function.

6. The sintering flue gas control method based on parameter coupling total differential constraint according to claim 4, characterized in that: The joint training loss function for sintered nitrides in S23 is: ； ； ； ； in, For the joint training loss function of sintered nitrogen oxides; These are the weighting coefficients for the constraint cascade dynamics mapping function; To constrain the weighting coefficients of the holographic prediction function; The weighting coefficients are used to constrain the chain prediction loss function; The loss function is the cascade dynamics mapping function. The loss function is for direct prediction; The loss function is a chained prediction function. Weights for sample quality; Here, N is the loss function for the regression task; N is the number of training samples. The measured concentration of nitrogen oxides at the inlet of the sintering flue gas after normalization at time t; For predicted values ​​of passive state variables; These are normalized passive state variables; The predicted value of nitrogen oxides after averaging the variable functions; This is the predicted value of nitrogen oxides at the sintering flue gas inlet.

7. The sintering flue gas control method based on parameter coupling total differential constraint according to claim 1, characterized in that: S3 specifically refers to: S31: Generate the active operation variable locking state Actively manipulate unbounded optimization variables Map the data to the historical boundary of the actively manipulated variable, and then force the locked variable to be backfilled with the ground state value; S32: Set unbounded optimization variables for active operation The active operation variables are backfilled based on the unbounded optimization variables of the active operation; cascade kinetic mapping is performed to construct a composite objective function for nitrogen oxides. ; S33: Solve the objective function that causes nitrogen oxides to recombine. The minimum optimization variable yields the optimized active operation variable. , sintering thermal passive state variables , sintering batching coupling correlation vector Predicted values ​​of nitrogen oxides at the inlet of sintering flue gas after optimization .

8. The sintering flue gas control method based on parameter coupling total differential constraint according to claim 7, characterized in that: In S32, the active operation variables are backfilled based on the unbounded optimization variables of the active operation, specifically as follows: ； ； in, This is to backfill the j-th active operation variable after optimization; The j-th active operation variable is locked. These are the initial active operation variables; Let j be the candidate value for the active operation variable; To actively manipulate unbounded optimization variables; This is the historical lower bound of the j-th actively manipulated variable; The historical upper limit of the j-th actively manipulated variable; is the activation function; j is the active operation variable number.

9. The sintering flue gas control method based on parameter coupling total differential constraint according to claim 7, characterized in that: The specific objective function for the recombination of nitrogen oxides in S32 is as follows: ； in, The objective function is the composite objective function for nitrogen oxides; Adjust the penalty function for actively manipulated variables; For the passive state variable offset penalty function; This is a penalty function for ingredient ratio deviation. For boundary barrier penalty function; This is a penalty function for preventing penetration outside the domain. The weighting coefficients for the penalty terms of actively manipulated variables; These are the weighting coefficients for passive state variables; The weighting coefficients for the ingredient proportions; These are the weighting coefficients for the boundary barriers; To optimize the predicted value of nitrogen oxides at the inlet of the sintering flue gas.

10. The sintering flue gas control method based on parameter coupling total differential constraint according to claim 1, characterized in that: It also includes S4: verifying the effectiveness of emission reduction adjustment parameters, specifically: S41: Calculate the adjustment ratio for each active variable to form the first-order active variable control parameters. This is used to quantify the change ratio of the normalized active operating variable relative to the current baseline operating condition after optimization. S42: Calculate the change ratio of each passive state variable to form the second-order state linkage prediction parameters. Output active operation variable adjustment items and passive state variables, verify the sintering condition transfer process, and use whether the linkage deduction value is within the corresponding historical boundary interval as the judgment basis. S43: Output ingredient coupling correlation vector from arrive The results of the changes are used to verify whether the nitrogen oxide emission reduction adjustment parameters of sintering flue gas are effective, and the effective judgment is based on the fact that the predicted value of nitrogen oxides at the inlet of sintering flue gas after optimization is lower than the predicted value of the ground state and the predicted emission reduction is greater than 0.

Images