SNCR denitration prediction method based on fusion multi-neural network
By integrating multiple neural networks to train and optimize model parameters, the problem of inaccurate prediction of NOx output in SNCR denitrification systems was solved, achieving accurate prediction of NOx output and improving the control precision of the denitrification system.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-03-06
- Publication Date
- 2026-04-07
AI Technical Summary
Existing single neural network models cannot accurately predict NOx output in SNCR denitrification systems, resulting in large fluctuations in NO concentration at the denitrification system outlet and an inability to quickly adjust the ammonia injection rate.
By employing a multi-neural network approach, the relevance and accuracy of predicted values are optimized through training an initial model, screening subsequent models, and adjusting model parameters, resulting in a final model that achieves accurate predictions.
It improves the prediction accuracy of NOx output content, enables accurate prediction of NOx output content, and enhances the dynamic control capability of the denitrification system.
Smart Images
Figure CN116167523B_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the field of automatic ammonia injection technology for the power industry, and particularly relates to an SNCR denitration prediction method based on a multi-neural network. BACKGROUND
[0002] During the operation of various types of combustor power generating units, a large amount of NOx generated by combustion will pollute the environment. In order to reduce NOx emissions, the implementation and optimization of denitration technology have become a key problem for reducing pollution of power generating units. In the selective non-catalytic reduction (SNCR) denitration control system of a power generating unit, the traditional method is to control the denitration system by using a PID algorithm, but this will cause the NO concentration at the outlet of the denitration system to fluctuate greatly. At the same time, due to the delay of the chemical reaction time of sampling detection, the NOx output cannot be accurately predicted to quickly adjust the amount of ammonia injection. In order to accurately predict the real-time NOx output content and thus dynamically control the content of the reaction reducing agent, an accurate denitration prediction model needs to be established. At present, in the prior art, there are denitration prediction models established by using BP neural network algorithms, multi-layer perceptron neural networks, deep recurrent neural networks DRNN-LSTM, etc. to predict the NOx output content. However, in the actual production environment, due to the different conditions of each unit, the parameters affecting the denitration system are complex, which leads to defects in the use of existing single neural network models, and the NOx output content in the SNCR denitration system cannot be accurately predicted. SUMMARY
[0003] In view of the deficiencies in the prior art, the purpose of the present application is to provide an SNCR denitration prediction method based on a multi-neural network, which can improve the prediction accuracy of the NOx output content.
[0004] To achieve the above-mentioned purpose, the present application provides the following technical scheme: an SNCR denitration prediction method based on a multi-neural network, comprising:
[0005] Step S1, a plurality of different initial models are obtained by training a preset data set, the predicted output values of each initial model are compared with the true output values in the data set, and each initial model is screened according to the comparison result to obtain a plurality of screened models;
[0006] Step S2, each screened model is sequentially run according to the comparison result to output different screened prediction values, and the correlation between each screened prediction value and the true NOx emission under the current state is calculated to obtain the correlation between each screened prediction value and the prediction accuracy of each screened prediction value;
[0007] Step S3: Based on the correlation between the predicted values after screening and the prediction accuracy of the predicted values after screening, the weight of each input parameter of each post-screening model is adjusted to obtain several optimized models. Each optimized model is used to output optimized predicted values.
[0008] Step S4: Calculate the correlation between each of the optimized predicted values and the prediction accuracy of each of the optimized predicted values based on the relationship between each of the optimized predicted values and the actual NOx emissions.
[0009] Step S5: Adjust the proportion of each input parameter of each optimization model according to the correlation between each of the optimized predicted values and the prediction accuracy of each of the optimized predicted values to obtain the adjusted optimization model.
[0010] Step S6: Calculate the prediction accuracy of each optimized prediction value based on the relationship between each optimized prediction value and the actual NOx emissions; determine whether the difference in prediction accuracy between the optimized model with the highest prediction accuracy after the most recent adjustment and the optimized model with the highest prediction accuracy after the previous adjustment is lower than a preset threshold.
[0011] If so, the optimized model with the highest prediction accuracy after the most recent adjustment will be output as the final model.
[0012] If not, proceed to step S5.
[0013] Further, step S1 includes:
[0014] Step S11: Divide the data in the dataset into a training set and a validation set according to a preset ratio;
[0015] Step S12: Train each of the initial models using the data in the training set. The initial models include Bi-LSTM network model, SVM network model, DNN neural network model, CNN neural network model, and unidirectional LSTM network model.
[0016] Step S13: Input several input parameters from the validation set into each of the initial models to obtain different predicted output values. The validation set also includes the true output values corresponding to each of the output parameters.
[0017] Step S14 involves comparing each predicted output value with the true output value to obtain the comparison result, and using mean squared error and goodness of fit to filter each initial model corresponding to each predicted output value in the comparison result to obtain each filtered post-model. The filtered post-model includes the Bi-LSTM network model, the CNN neural network model, and the unidirectional LSTM network model. Step S14 includes:
[0018] Step 141: Calculate the mean square error and goodness of fit of each predicted output value relative to the true output value in the comparison results;
[0019] Step 142: Substitute each of the mean square error values and the goodness-of-fit values into the preset sorting formula to obtain the sorting priority value;
[0020] Step 143: Sort and filter the initial models according to the sorting priority value from high to low to obtain the filtered models. The sorting priority value of the Bi-LSTM network model is higher than that of the unidirectional LSTM network model, and the sorting priority value of the unidirectional LSTM network model is higher than that of the CNN neural network model.
[0021] The sorting formula is configured as follows:
[0022]
[0023] Wherein, Sf represents the sorting priority value;
[0024] MSE is used to represent the mean square error value;
[0025] R 2 Used to represent the goodness-of-fit value;
[0026] k1 and k2 are used to represent the preset first and second coefficients, respectively, k1,k2∈(0,1).
[0027] Further, step S2 includes:
[0028] Step S21: Run the Bi-LSTM network model, the unidirectional LSTM network model, and the CNN neural network model in descending order of the sorting priority value, so that the Bi-LSTM network model, the unidirectional LSTM network model, and the CNN neural network model output different predicted values after screening respectively;
[0029] Step S22: Calculate the absolute value of the error between each of the predicted values after screening and the actual NOx emissions, and process each of the absolute values of error to obtain the prediction accuracy of each of the predicted values after screening.
[0030] Step S23: Perform correlation analysis on each of the post-screening predicted values, and calculate the first correlation coefficient of the remaining post-screening predicted values with respect to the post-screening predicted value with the highest prediction accuracy.
[0031] Further, step S3 includes:
[0032] Step S31: Adjust the weight parameters between the input layer and the hidden layer in the corresponding post-screening model according to each of the first correlation coefficients and the prediction accuracy of each of the post-screening predicted values.
[0033] Step S32: Retrain and adjust each of the screened models to obtain the optimized model.
[0034] Further, in step S31, the first correlation coefficient and the prediction accuracy of each of the screened predicted values are input into a weight parameter adjustment formula to adjust the weight parameters between the input layer and the hidden layer in the screened model. The weight parameter adjustment formula is configured as follows:
[0035]
[0036] W is used to represent the adjusted weight parameter;
[0037] W0 is used to represent the preset initial weight parameters;
[0038] P a Used to represent the prediction accuracy of the post-screening predicted values;
[0039] C1 is used to represent the first correlation coefficient;
[0040] b1 and b2 are used to represent the preset first adjustment coefficient and the second adjustment coefficient, respectively.
[0041] Furthermore, both the post-screening predicted value and the optimized predicted value are NOx output content. The input parameters of each post-screening model and each optimized model include several core parameters and several non-core parameters. The core parameters of each post-screening model are the same, and the core parameters include total coal quantity, total primary air volume, and total secondary air volume.
[0042] The non-core parameters of each of the screened models are different. The non-core parameters of the Bi-LSTM network model include actual power generation, total air volume, right furnace bed pressure, left furnace bed pressure, average bed temperature on the right side of the furnace front wall, average bed temperature on the left side of the furnace front wall, average bed temperature on the right side of the furnace rear wall, average bed temperature on the left side of the furnace rear wall, right furnace outlet temperature, left furnace outlet temperature, furnace upper pressure 3, and furnace upper pressure 1.
[0043] The non-core parameters of the unidirectional LSTM network model include unit load, burner swing angle, pressure difference between the wind box and the furnace, oxygen content at the furnace outlet, bed temperature, ammonia injection rate, and oxygen content at the air compressor inlet.
[0044] The non-core parameters of the CNN neural network model include unit load, primary air pressure, secondary air damper opening at each level, burnout damper opening at each level, boiler outlet flue gas flow rate, boiler outlet flue gas temperature, boiler outlet flue gas oxygen content, boiler outlet NOx concentration, boiler inlet NOx concentration, SNCR inlet flue gas pressure, SNCR outlet flue gas pressure, power supply coal consumption, power generation coal consumption, and boiler efficiency.
[0045] Then step S31 includes:
[0046] Step S311: Perform correlation analysis between each of the non-core parameters and the corresponding post-screening predicted value, and calculate the second correlation coefficient of each of the non-core parameters with respect to the corresponding post-screening predicted value;
[0047] Step S312: Adjust the weight parameters in the sieved model corresponding to the non-core parameters according to each of the second correlation coefficients;
[0048] In the weight parameter adjustment formula, when the weight parameters in the screened model correspond to non-core parameters, the first adjustment coefficient and the second adjustment coefficient are configured as follows: Where b1' and b2' represent the initial values of the first adjustment coefficient and the second adjustment coefficient, respectively;
[0049] C2 is used to represent the second correlation coefficient;
[0050] m is used to represent the number of the non-core parameters;
[0051] When the weight parameters in the screened model correspond to the core parameters, the first adjustment coefficient and the second adjustment coefficient are configured as follows:
[0052] Where n represents the number of the core parameters.
[0053] Furthermore, the core parameters and non-core parameters of each of the post-screening models and each of the optimized models are all obtained in real time by several sensors installed in the SNCR denitrification system;
[0054] The non-core parameters include vulnerable sensor parameters. The vulnerable sensor parameters of the Bi-LSTM network model include the right furnace bed pressure, the left furnace bed pressure, the furnace upper pressure 3, and the furnace upper pressure 1. The vulnerable sensor parameters of the unidirectional LSTM network model include the air compressor inlet oxygen quantity. The vulnerable sensor parameters of the CNN neural network model include the boiler outlet flue gas flow rate.
[0055] The step between step S311 and step S312 further includes step S31A, which determines whether each of the non-core parameters includes the vulnerable sensor parameter:
[0056] If so, a corresponding sensor activation command will be generated based on the vulnerable sensor parameters, and the weight parameters in the sieved model corresponding to the vulnerable sensor parameters will be adjusted according to each of the second correlation coefficients, and then the process will proceed to step S312.
[0057] If not, then proceed to step S312;
[0058] In the weight parameter adjustment formula, when the weight parameters in the sieved model correspond to the vulnerable sensor parameters, the first adjustment coefficient and the second adjustment coefficient are configured as follows: Where p represents the number of parameters of the vulnerable sensor;
[0059] Before running each of the optimization models, the pressure sensor located in the furnace, the temperature sensor located at the air compressor inlet, or the flow sensor located at the boiler outlet are adjusted to their on / off states for a preset period of time.
[0060] Furthermore, the step between step S4 and step S5 includes:
[0061] Step S4A: Determine whether the difference in prediction accuracy between the optimized model with the highest prediction accuracy and the screened model with the highest prediction accuracy is lower than a preset threshold.
[0062] If so, the optimized model with the highest prediction accuracy will be output as the final model.
[0063] If not, proceed to step S5.
[0064] Further, step S4 includes:
[0065] Step S41: Calculate the absolute value of the error between each of the optimized predicted values and the actual NOx emissions, and process each of the absolute values of error to obtain the prediction accuracy of each of the optimized predicted values.
[0066] Step S42: Perform correlation analysis on each of the optimized prediction values, and calculate the third correlation coefficient of the remaining optimized prediction values with respect to the optimized prediction value with the highest prediction accuracy;
[0067] Then step S5 includes:
[0068] Step S51: Adjust the weight parameters between the input layer and the hidden layer in the corresponding optimization model according to the prediction accuracy of each of the third correlation coefficients and each of the optimized prediction values.
[0069] Step S52: Retrain each of the optimization models to obtain the adjusted optimization model.
[0070] Further, step S51 includes:
[0071] Step S511: Perform correlation analysis between each non-core parameter and the corresponding optimized predicted value, and calculate the fourth correlation coefficient of each non-core parameter with respect to the corresponding optimized predicted value.
[0072] Step S512: Adjust the weight parameters in the optimization model corresponding to the non-core parameters according to each fourth correlation coefficient;
[0073] In the weight parameter adjustment formula, when the weight parameters in the optimization model correspond to non-core parameters, the first and second adjustment coefficients are configured as follows: C4 represents the fourth correlation coefficient;
[0074] When the weight parameters in the optimization model correspond to the core parameters, the first and second adjustment coefficients are configured as follows:
[0075] The step between step S511 and step S512 further includes step S51A, which determines whether each of the non-core parameters includes the vulnerable sensor parameter:
[0076] If so, a corresponding sensor activation command will be generated based on the vulnerable sensor parameters, and the weight parameters in the optimization model corresponding to the vulnerable sensor parameters will be adjusted according to each of the second correlation coefficients.
[0077] If not, proceed to step S512;
[0078] In the weight parameter adjustment formula, when the weight parameters in the optimization model correspond to vulnerable sensor parameters, the first adjustment coefficient and the second adjustment coefficient are configured as follows:
[0079] Before each optimized model is run and adjusted, the pressure sensor located in the furnace, the temperature sensor located at the air compressor inlet, or the flow sensor located at the boiler outlet is adjusted in advance for the preset time period to adjust the opening and closing status.
[0080] The beneficial effects of this invention are:
[0081] This invention calculates the correlation and prediction accuracy between each post-screening predicted value and the actual NOx emission based on the relationship between each post-screening predicted value and the actual NOx emission. The proportion of each input parameter of each post-screening model is adjusted to obtain each optimized model. Compared with the original post-screening model, the prediction accuracy has been initially improved.
[0082] Then, based on the correlation between the optimized predicted values and the prediction accuracy, the weight of each input parameter of each optimized model is adjusted to obtain the adjusted optimized model. By iteratively adjusting the weight parameters of each optimized model, the prediction accuracy of each optimized model is continuously improved until the difference in prediction accuracy is lower than the preset threshold, and the final model is output. This final model can achieve accurate prediction of NOx output content, thereby effectively improving the prediction accuracy of NOx output content. Attached Figure Description
[0083] Figure 1 This is a flowchart of the steps of the SNCR denitrification prediction method integrating multiple neural networks in this invention;
[0084] Figure 2 This is a process diagram of the SNCR denitrification system in this invention;
[0085] Figure 3 This is a diagram showing the mean square error values of each initial model in this invention;
[0086] Figure 4 This is a graph showing the goodness-of-fit values of each initial model in this invention;
[0087] Figure 5 This is a flowchart of step 1 in this invention;
[0088] Figure 6 This is a flowchart of step 14 in this invention;
[0089] Figure 7 This is a flowchart of step 2 in this invention;
[0090] Figure 8 This is a flowchart of step 3 in this invention;
[0091] Figure 9 This is a flowchart of step 31 in this invention;
[0092] Figure 10This is a flowchart of step 4A in this invention;
[0093] Figure 11 This is a flowchart of step 4 in this invention;
[0094] Figure 12 This is a flowchart of step 5 in this invention;
[0095] Figure 13 This is a flowchart of step 51 in this invention. Detailed Implementation
[0096] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Identical components are denoted by the same reference numerals. It should be noted that the terms "front," "rear," "left," "right," "upper," and "lower" used in the following description refer to directions in the accompanying drawings, and the terms "bottom surface," "top surface," "inner," and "outer" refer to directions toward or away from the geometric center of a specific component, respectively.
[0097] like Figure 1 As shown in this embodiment, an SNCR denitrification prediction method integrating multiple neural networks includes:
[0098] Step S1: Train several different initial models using a pre-set dataset, compare the predicted output value of each initial model with the actual output value in the dataset, and filter each initial model based on the comparison results to obtain several filtered models.
[0099] Step S2: Run each post-screening model sequentially according to the comparison results to output different post-screening predicted values. Calculate the correlation between each post-screening predicted value and the accuracy of each post-screening predicted value based on the relationship between each post-screening predicted value and the actual NOx emissions under the current state.
[0100] Step S3: Based on the correlation between the predicted values after screening and the prediction accuracy of each predicted value after screening, adjust the proportion of each input parameter of each screening model to obtain several optimized models. Each optimized model is used to output optimized predicted values.
[0101] Step S4: Calculate the correlation between each optimized prediction value and the prediction accuracy of each optimized prediction value based on the relationship between each optimized prediction value and the actual NOx emissions.
[0102] Step S5: Adjust the proportion of each input parameter of each optimization model according to the correlation between each optimized predicted value and the prediction accuracy of each optimized predicted value to obtain the adjusted optimization model.
[0103] Step S6: Calculate the prediction accuracy of each optimized prediction value based on the relationship between each optimized prediction value and the actual NOx emissions. Determine whether the difference in prediction accuracy between the optimized model with the highest prediction accuracy after the most recent adjustment and the optimized model with the highest prediction accuracy after the previous adjustment is lower than a preset threshold.
[0104] If so, the optimized model with the highest prediction accuracy after the most recent adjustment will be used as the final model output;
[0105] If not, proceed to step S5.
[0106] Specifically, this embodiment is applied to an SNCR denitrification system to accurately predict the NOx output content in the flue gas after boiler combustion in the SNCR denitrification system. The boiler used in the SNCR denitrification system is a DG1089 / 17.4-II1 type circulating fluidized bed boiler manufactured by Dongfang Boiler Co., Ltd., with a power of 300MW and a superheated steam flow rate of 1088.48t / H. The SNCR denitrification system can achieve denitrification by reacting urea or liquid ammonia as a reducing agent with nitric oxide in the SNCR flue gas. In this embodiment, the SNCR denitrification system uses urea as a reducing agent for reduction reaction denitrification. The denitrification process flow of the SNCR denitrification system is as follows: Figure 2 As shown.
[0107] Preferred, such as Figure 5 As shown, step S1 includes:
[0108] Step S11: Divide the data in the dataset into a training set and a validation set according to a preset ratio;
[0109] Step S12: Train each initial model using the data in the training set. The initial models include Bi-LSTM network model, SVM network model, DNN neural network model, CNN neural network model and unidirectional LSTM network model.
[0110] Step S13: Input several input parameters from the validation set into each initial model to obtain different predicted output values. The validation set also contains the true output values corresponding to each output parameter.
[0111] Step S14: Compare each predicted output value with the actual output value to obtain the comparison result. Use the mean squared error and goodness of fit to screen each initial model corresponding to each predicted output value in the result to obtain each screened model. The screened models include Bi-LSTM network model, CNN neural network model and unidirectional LSTM network model.
[0112] Specifically, in this embodiment, the data in the dataset includes standard input parameters and actual output values. The standard input parameters are those collected under the current conditions in the SNCR denitrification system, including actual power generation, total coal consumption, total air volume, total primary air volume, total secondary air volume, right-side furnace bed pressure, left-side furnace bed pressure, average bed temperature on the right side of the furnace front wall, average bed temperature on the left side of the furnace front wall, average bed temperature on the right side of the furnace rear wall, average bed temperature on the left side of the furnace rear wall, right-side furnace outlet temperature, left-side furnace outlet temperature, furnace upper pressure 3, and furnace upper pressure 1. The actual output values are the actual data corresponding to the standard input parameters, including the NOx content in the original flue gas and the urea flow rate in the furnace area urea solution. The NOx output content can be calculated using the original flue gas NOx content and the furnace area urea flow rate. The ratio of training set to validation set can be 6:1. Five initial models—Bi-LSTM, SVM, DNN, CNN, and unidirectional LSTM—are trained using the standard input parameters and corresponding true output values from the training set. During the training of these initial models, data fusion and missing value handling are performed on the parameters in the training set. Then, through feature derivation, supervised processing, and dimensionless transformation, statistical features of various factors and temporal features of time-series data are integrated, resulting in more cross-term and higher-order features. Next, empirical methods and stepwise trial-and-error are used to determine the hidden layer dimensions. Finally, the models are trained. After the five initial models are trained, the standard input parameters from the validation set are input into each of the five initial models, causing them to output different predicted output values. These predicted output values are then compared with the true output values. Mean squared error and goodness-of-fit comparisons are used to filter the initial models corresponding to each predicted output value, selecting the three best initial models as the final models for subsequent steps.
[0113] Preferred, such as Figure 6 As shown, step 14 includes:
[0114] Step 141: Calculate the mean square error and goodness of fit of each predicted output value relative to the true output value in the comparison results;
[0115] Step 142: Substitute each mean square error value and goodness-of-fit value into the preset sorting formula to obtain the sorting priority value;
[0116] Step 143: Sort and filter the initial models according to their priority values from high to low to obtain the filtered models. The priority value of the Bi-LSTM network model is higher than that of the unidirectional LSTM network model, and the priority value of the unidirectional LSTM network model is higher than that of the CNN neural network model.
[0117] Specifically, in this embodiment, the mean squared error and goodness-of-fit values corresponding to the Bi-LSTM network model, SVM network model, DNN neural network model, CNN neural network model, and unidirectional LSTM network model are respectively as follows: Figure 3 and Figure 4 As shown, after calculating the mean squared error and goodness-of-fit values of each initial model, a ranking priority value is generated using a ranking formula. The top three initial models with the highest ranking priority values are then selected as the screening models, completing the model screening process. This screening process not only considers the impact of mean squared error and goodness-of-fit values on the accuracy of the initial models, but also utilizes a ranking formula to automatically rank the initial models, simplifying the model screening process. Furthermore, the screened models have higher prediction accuracy compared to other initial models with lower ranking priority values, laying the foundation for further improvements in model accuracy.
[0118] Preferably, the sorting formula is configured as follows:
[0119]
[0120] Where Sf represents the sorting priority value;
[0121] MSE is used to represent the mean square error value;
[0122] R 2 Used to represent goodness-of-fit values;
[0123] k1 and k2 are used to represent the preset first and second coefficients, respectively, k1,k2∈(0,1).
[0124] Specifically, in this embodiment, k1 can be 2 / 5 and k2 can be 3 / 5. The values of k1 and k2 can be adaptively adjusted when the type of the initial model changes.
[0125] Preferred, such as Figure 7 As shown, step S2 includes:
[0126] Step S21: Run the Bi-LSTM network model, the one-way LSTM network model, and the CNN neural network model in descending order of sorting priority, so that the Bi-LSTM network model, the one-way LSTM network model, and the CNN neural network model output different sieved prediction values respectively.
[0127] Step S22: Calculate the absolute value of the error between each post-screening predicted value and the actual NOx emission, and process each absolute value of error to obtain the prediction accuracy of each post-screening predicted value.
[0128] Step S23: Perform correlation analysis on the predicted values after screening and calculate the first correlation coefficient of the remaining predicted values with respect to the predicted value with the highest prediction accuracy.
[0129] Specifically, in this embodiment, the core logic of this technical solution is to switch to the model with the highest current prediction accuracy. Throughout all steps of this technical solution, the system switches to the model with the highest current prediction accuracy to maintain the prediction accuracy at all times. During the switching operation of each post-screening model, the input parameters of each post-screening model include several core parameters and several non-core parameters. The core parameters are the same for different post-screening models, while the non-core parameters differ. Because the input parameters of each post-screening model are not entirely the same, and the algorithms of the post-screening models are different, the output post-screening prediction values differ in magnitude but are of the same type, all used to represent the NOx output content.
[0130] In step 143, the ranking priority of the Bi-LSTM network model was calculated to be the highest, followed by the unidirectional LSTM network model, and then the CNN neural network model. The Bi-LSTM network model, LSTM network model, and CNN neural network model were run sequentially according to their ranking priority, resulting in different post-screening predicted values output by the three models. The absolute value of the error was obtained by subtracting the absolute value of each post-screening predicted value from the actual NOx emissions. The prediction accuracy of the post-screening predicted value was then calculated using 1 - absolute value of error / post-screening predicted value. After calculating the prediction accuracy of each post-screening predicted value, the prediction value with the highest prediction accuracy was used as a benchmark. The first correlation coefficient between the remaining post-screening predicted values and the prediction value with the highest prediction accuracy was calculated using the sample covariance formula, laying the foundation for subsequent calculations.
[0131] Preferred, such as Figure 8 As shown, step S3 includes:
[0132] Step S31: Adjust the weight parameters between the input layer and the hidden layer in the corresponding post-screening model according to each first correlation coefficient and the prediction accuracy of each post-screening prediction value.
[0133] Step S32: Retrain and adjust each screened model to obtain the optimized model.
[0134] Specifically, in this embodiment, the weight parameters between the input layer and the hidden layer in the post-screening model are adjusted by using the first correlation coefficient and the prediction accuracy of each post-screening prediction value. This achieves parameter tuning of the post-screening model, improves the prediction accuracy of each post-screening model, and obtains optimized models. Compared with the original post-screening model, the prediction accuracy of each optimized model has been initially improved.
[0135] Preferably, in step S31, the first correlation coefficient and the prediction accuracy of each post-screening predicted value are input into the weight parameter adjustment formula to adjust the weight parameters between the input layer and the hidden layer in the post-screening model. The weight parameter adjustment formula is configured as follows:
[0136]
[0137] W is used to represent the adjusted weight parameters;
[0138] W0 is used to represent the preset initial weight parameters;
[0139] P a Used to represent the prediction accuracy of the predicted values after screening;
[0140] C1 is used to represent the first correlation coefficient;
[0141] b1 and b2 are used to represent the preset first adjustment coefficient and the second adjustment coefficient, respectively.
[0142] Preferably, both the post-screening predicted value and the optimized predicted value are NOx output content. The input parameters of each post-screening model include several core parameters and several non-core parameters. The core parameters of each post-screening model are the same, including total coal quantity, total primary air volume, and total secondary air volume.
[0143] The non-core parameters of each screened model are different. The non-core parameters of the Bi-LSTM network model include actual power generation, total air volume, right furnace bed pressure, left furnace bed pressure, average bed temperature on the right side of the furnace front wall, average bed temperature on the left side of the furnace front wall, average bed temperature on the right side of the furnace rear wall, average bed temperature on the left side of the furnace rear wall, right furnace outlet temperature, left furnace outlet temperature, furnace upper pressure 3, and furnace upper pressure 1.
[0144] Non-core parameters of the unidirectional LSTM network model include unit load, burner sway angle, pressure difference between the wind box and the furnace, oxygen content at the furnace outlet, bed temperature, ammonia injection rate, and oxygen content at the air compressor inlet.
[0145] Non-core parameters of the CNN neural network model include unit load, primary air pressure, secondary air damper opening at each level, burnout damper opening at each level, boiler outlet flue gas flow rate, boiler outlet flue gas temperature, boiler outlet flue gas oxygen content, boiler outlet NOx concentration, boiler inlet NOx concentration, SNCR inlet flue gas pressure, SNCR outlet flue gas pressure, power supply coal consumption, power generation coal consumption, and boiler efficiency.
[0146] like Figure 9 As shown, step S31 includes:
[0147] Step S311: Perform correlation analysis between each non-core parameter and the corresponding post-screening predicted value, and calculate the second correlation coefficient of each non-core parameter with respect to the corresponding post-screening predicted value.
[0148] Step S312: Adjust the weight parameters in the sieved model corresponding to the non-core parameters according to each second correlation coefficient;
[0149] In the weight parameter adjustment formula, when the weight parameters in the screened model correspond to non-core parameters, the first and second adjustment coefficients are configured as follows: Where b1' and b2' represent the initial values of the first adjustment coefficient and the second adjustment coefficient, respectively;
[0150] C2 is used to represent the second correlation coefficient;
[0151] m is used to represent the number of non-core parameters;
[0152] When the weight parameters in the screened model correspond to the core parameters, the first and second adjustment coefficients are configured as follows:
[0153] Here, n represents the number of core parameters.
[0154] Specifically, in this embodiment, the non-core parameters of each post-screening model are different. Different non-core parameters will affect the post-screening predicted values output by each post-screening model, thereby affecting the prediction accuracy of each post-screening model. In order to reduce the impact of different non-core parameters on the prediction accuracy of each post-screening model and improve the prediction accuracy of each post-screening model, it is necessary to perform correlation analysis between each non-core parameter and the corresponding post-screening predicted value. The second correlation coefficient of each non-core parameter with respect to the corresponding post-screening predicted value is calculated using the sample covariance calculation formula. When the second correlation coefficient is large, it indicates that the non-core parameter has a large impact on the post-screening predicted value, and the weight parameter of the non-core parameter needs to be adjusted to adjust the post-screening predicted value. The initial values of the first adjustment coefficient and the second adjustment coefficient are determined, which can be 2 and 3, respectively. In this embodiment, the number of core parameters is 3. Therefore, when the weight parameter in the post-screening model corresponds to the core parameter, the first adjustment coefficient and the second adjustment coefficient are constant values. When the weight parameters in the post-screening model correspond to non-core parameters, the second correlation coefficient needs to be calculated based on each non-core parameter and the post-screening prediction value. Since the number of non-core parameters varies in different post-screening models, the weight parameters corresponding to the non-core parameters will change dynamically with the changes in the non-core parameters. The second correlation coefficient achieves dynamic optimization of the weight parameters by dynamically adjusting the first and second adjustment coefficients in the weight parameter adjustment formula, thereby improving the prediction accuracy of the post-screening model.
[0155] Preferably, all core parameters and non-core parameters of each sieved model and each optimized model are obtained in real time by several sensors installed in the SNCR denitrification system;
[0156] The non-core parameters include vulnerable sensor parameters. The vulnerable sensor parameters of the Bi-LSTM network model include the right furnace bed pressure, the left furnace bed pressure, the furnace upper pressure 3, and the furnace upper pressure 1. The vulnerable sensor parameters of the unidirectional LSTM network model include the air compressor inlet oxygen quantity. The vulnerable sensor parameters of the CNN neural network model include the boiler outlet flue gas flow rate.
[0157] Between steps S311 and S312, there is also step S31A, which determines whether the non-core parameters include vulnerable sensor parameters:
[0158] If so, a corresponding sensor activation command will be generated based on the vulnerable sensor parameters, and the weight parameters in the sieved model corresponding to the vulnerable sensor parameters will be adjusted according to each second correlation coefficient, and then the process will proceed to step S312.
[0159] If not, proceed to step S312;
[0160] In the weight parameter adjustment formula, when the weight parameters in the screened model correspond to the parameters of vulnerable sensors, the first adjustment coefficient and the second adjustment coefficient are configured as follows: Where p represents the number of vulnerable sensor parameters;
[0161] Before running each optimization model, the pressure sensor located in the furnace, the temperature sensor located at the air compressor inlet, or the flow sensor located at the boiler outlet should be adjusted to adjust their opening and closing status for a preset period of time.
[0162] Specifically, in this embodiment, different post-screening models require different non-core parameters for prediction. In traditional model prediction, sensors that collect input parameters of the SNCR denitrification system need to be kept on at all times for real-time data acquisition. This not only causes unnecessary energy consumption but also accelerates sensor damage and reduces sensor lifespan. In this technical solution, the post-screening model with the highest prediction accuracy is prioritized. While this model is running, the easily damaged sensors in the other two post-screening models can be turned off to extend their lifespan. As the weight parameters of the post-screening models are adjusted to obtain an optimized model, the optimized model with the highest prediction accuracy after adjustment may not be the same as the original post-screening model with the highest prediction accuracy. Therefore, before running the optimized model, it is determined whether the non-core parameters of the optimized model contain parameters of easily damaged sensors to generate corresponding sensor activation commands. This allows easily damaged sensors to be activated in advance, so that they can be in a closed state before the model stops running. Data acquisition begins upon receiving the sensor activation command, and the sensors automatically shut down after the model stops running for a period of time. In this embodiment, the vulnerable sensors include a pressure sensor located in the furnace, a temperature sensor located at the air compressor inlet, and a flow sensor located at the boiler outlet.
[0163] The non-core parameters in step S312 are the remaining non-core parameters that do not include vulnerable sensor parameters. When the weight parameters in the sieved model correspond to vulnerable sensor parameters, the following is adopted: Calculate the first adjustment coefficient and the second adjustment coefficient. When the weight parameters in the screened model correspond to the remaining non-core parameters that do not include vulnerable sensor parameters, the following adjustment is adopted: The first and second adjustment coefficients are calculated by using different coefficient configuration formulas. This enables precise differentiation of whether vulnerable sensor parameters are included among the non-core parameters when adjusting the weight parameters in the post-screening model. Different first and second adjustment coefficients can be calculated based on whether vulnerable sensor parameters are included among the non-core parameters, thereby achieving dynamic optimization of the weight parameters in the post-screening model and further improving the prediction accuracy of the post-screening model.
[0164] Preferred, such as Figure 10 As shown, the step between S4 and S5 also includes:
[0165] Step S4A: Determine whether the difference in prediction accuracy between the optimized model with the highest prediction accuracy and the screened model with the highest prediction accuracy is lower than a preset threshold.
[0166] If so, the optimized model with the highest prediction accuracy will be output as the final model.
[0167] If not, proceed to step S5.
[0168] Specifically, in this embodiment, the preset threshold can be 0.01%. When the difference in prediction accuracy between the optimized model with the highest prediction accuracy and the screened model with the highest prediction accuracy is less than 0.01%, it indicates that the prediction accuracy of the optimized model with the highest prediction accuracy has met the requirements and can be used as the final model output. Otherwise, the weight parameters of each optimized model need to be continuously adjusted to optimize the prediction accuracy of the model.
[0169] Preferred, such as Figure 11 As shown, step S4 includes:
[0170] Step S41: Calculate the absolute value of the error between each optimized prediction value and the actual NOx emission, and process each absolute value of error to obtain the prediction accuracy of each optimized prediction value.
[0171] Step S42: Perform correlation analysis on each optimized prediction value and calculate the third correlation coefficient of the remaining optimized prediction values with respect to the optimized prediction value with the highest prediction accuracy.
[0172] like Figure 12 As shown, step S5 includes:
[0173] Step S51: Adjust the weight parameters between the input layer and the hidden layer in the corresponding optimization model according to the prediction accuracy of each third correlation coefficient and each optimized prediction value.
[0174] Step S52: Retrain each optimization model to obtain the adjusted optimization model.
[0175] Specifically, in this embodiment, step S42 uses the sample covariance calculation formula to calculate the third correlation coefficient of the remaining optimized predicted values with respect to the optimized predicted value with the highest prediction accuracy. Step S51 adjusts the weight parameters between the input layer and hidden layer in the optimization model in the same way as in step S31: the third correlation coefficient and the prediction accuracy of each optimized predicted value are input into the weight parameter adjustment formula to adjust the weight parameters between the input layer and hidden layer in the optimization model. During the calculation of the weight parameters, the third correlation coefficient replaces the first correlation coefficient in the original weight parameter adjustment formula, while the rest remain unchanged. The updated weight parameter adjustment formula is then: C3 is used to represent the third correlation coefficient. The input parameters of the optimization model are the same as those of the sieved model, including core parameters and non-core parameters. The core parameters are the same for different optimization models, while the non-core parameters are different.
[0176] like Figure 13 As shown, step S51 includes:
[0177] Step S511: Perform correlation analysis between each non-core parameter and the corresponding optimized predicted value, and calculate the fourth correlation coefficient of each non-core parameter with respect to the corresponding optimized predicted value.
[0178] Step S512: Adjust the weight parameters in the optimization model corresponding to the non-core parameters according to each fourth correlation coefficient;
[0179] In the weight parameter adjustment formula, when the weight parameters in the optimization model correspond to non-core parameters, the first and second adjustment coefficients are configured as follows: C4 represents the fourth correlation coefficient;
[0180] When the weight parameters in the optimization model correspond to the core parameters, the first and second adjustment coefficients are configured as follows:
[0181] Between steps S511 and S512, there is also step S51A, which determines whether the non-core parameters include vulnerable sensor parameters:
[0182] If so, a corresponding sensor activation command will be generated based on the vulnerable sensor parameters, and the weight parameters in the optimization model corresponding to the vulnerable sensor parameters will be adjusted according to each second correlation coefficient.
[0183] If not, proceed to step S512;
[0184] In the weight parameter adjustment formula, when the weight parameters in the optimization model correspond to the parameters of vulnerable sensors, the first adjustment coefficient and the second adjustment coefficient are configured as follows:
[0185] Before each optimized model is run and adjusted, the pressure sensor located in the furnace, the temperature sensor located at the air compressor inlet, or the flow sensor located at the boiler outlet are adjusted to their open / closed states in advance for a preset time period.
[0186] Specifically, in this embodiment, the dynamic adjustment process of the first and second adjustment coefficients in step S51 is the same as in step S31, except that the second correlation coefficient is replaced by the fourth correlation coefficient. The weight parameters corresponding to the non-core parameters will change dynamically as the non-core parameters change. The fourth correlation coefficient achieves dynamic optimization of the weight parameters by dynamically adjusting the first and second adjustment coefficients in the weight parameter adjustment formula, thereby improving the prediction accuracy of the optimized model.
[0187] The dynamic adjustment process of the first adjustment coefficient and the second adjustment coefficient in step S51A is the same as that in step S31A. The only difference is that the second correlation coefficient is replaced by the fourth correlation coefficient. The weight parameters corresponding to the vulnerable sensor parameters will change dynamically with the changes in the vulnerable sensor parameters.
[0188] Different optimization models require different non-core parameters for prediction. In traditional model prediction, sensors collecting input parameters of the SNCR denitrification system need to be kept on at all times for real-time data acquisition. This not only causes unnecessary energy consumption but also accelerates sensor damage and reduces sensor lifespan. In this technical solution, the optimization model with the highest prediction accuracy is prioritized. While this optimization model is running, easily damaged sensors in the other two optimization models can be turned off to extend their lifespan. As the weight parameters of the optimization model are adjusted, the optimized model with the highest prediction accuracy after adjustment may not be the same as the original model with the highest prediction accuracy. Therefore, before running the optimization model, it is determined whether the non-core parameters of the optimization model contain parameters of vulnerable sensors to generate corresponding sensor activation commands. This allows vulnerable sensors to be activated in advance, so that they can be in a closed state before the model stops running. Data acquisition begins upon receiving the sensor activation command, and the sensors automatically shut down after the model stops running for a period of time. In this embodiment, vulnerable sensors include a pressure sensor located in the furnace, a temperature sensor located at the air compressor inlet, and a flow sensor located at the boiler outlet.
[0189] The non-core parameters in step S512 are the remaining non-core parameters that do not include vulnerable sensor parameters. When the weight parameters in the optimization model correspond to vulnerable sensor parameters, the following is adopted: Calculate the first and second adjustment coefficients. When the weight parameters in the optimization model correspond to the remaining non-core parameters that do not include vulnerable sensor parameters, the following adjustment coefficients are used. The calculation of the first and second adjustment coefficients, using different coefficient configuration formulas, enables precise differentiation of whether vulnerable sensor parameters are included among the non-core parameters when adjusting the weight parameters in the optimization model. This allows for the calculation of different first and second adjustment coefficients based on whether vulnerable sensor parameters are included among the non-core parameters, thereby achieving dynamic optimization of the weight parameters in the optimization model and further improving the prediction accuracy of the optimization model.
[0190] Step S6 calculates the absolute value of the error between each optimized predicted value and the actual NOx emission, and processes these absolute values to obtain the prediction accuracy of each optimized predicted value. If the difference in prediction accuracy between the optimized model with the highest prediction accuracy after the most recent adjustment and the optimized model with the highest prediction accuracy after the previous adjustment is less than 0.01%, it indicates that the prediction accuracy of the optimized model with the highest prediction accuracy after the most recent adjustment meets the requirements and can be used as the final model output. Otherwise, it indicates that the optimized model after the most recent adjustment still has room for further optimization, and the adjustment process in Step 5 needs to be repeated. Through several adjustments to each optimized model, the difference in prediction accuracy gradually decreases, while the prediction accuracy of each optimized model continuously improves until the difference in prediction accuracy is less than 0.01%, thus obtaining the final model. By iteratively adjusting the optimized models, the prediction accuracy of the optimized models is continuously improved until the final model is obtained. The final model can accurately predict NOx output content, effectively improving the prediction accuracy of NOx output content.
[0191] The above are merely preferred embodiments of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principle of the present invention should also be considered within the scope of protection of the present invention.
Claims
1. A method for predicting SNCR denitrification by integrating multiple neural networks, characterized in that, include: Step S1: Train several different initial models using a preset dataset, compare the predicted output value of each initial model with the actual output value in the dataset, and filter each initial model based on the comparison results to obtain several filtered models. Step S2: Run each of the post-screening models sequentially according to the comparison results to output different post-screening predicted values. Calculate the correlation between each post-screening predicted value and the prediction accuracy of each post-screening predicted value based on the relationship between each post-screening predicted value and the actual NOx emissions under the current state. Step S3: Based on the correlation between the predicted values after screening and the prediction accuracy of the predicted values after screening, adjust the proportion of each input parameter of each post-screening model to obtain several optimized models. Each optimized model is used to output optimized predicted values. Step S4: Calculate the correlation between each of the optimized predicted values and the prediction accuracy of each of the optimized predicted values based on the relationship between each of the optimized predicted values and the actual NOx emissions. Step S5: Adjust the proportion of each input parameter of each optimization model according to the correlation between each of the optimized predicted values and the prediction accuracy of each of the optimized predicted values to obtain the adjusted optimization model. Step S6: Calculate the prediction accuracy of each optimized prediction value based on the relationship between each optimized prediction value and the actual NOx emissions; determine whether the difference in prediction accuracy between the optimized model with the highest prediction accuracy after the most recent adjustment and the optimized model with the highest prediction accuracy after the previous adjustment is lower than a preset threshold. If so, the optimized model with the highest prediction accuracy after the most recent adjustment will be output as the final model. If not, proceed to step S5.
2. The SNCR denitrification prediction method fused with multiple neural networks according to claim 1, characterized in that, Step S1 includes: Step S11: Divide the data in the dataset into a training set and a validation set according to a preset ratio; Step S12: Train each of the initial models using the data in the training set. The initial models include Bi-LSTM network model, SVM network model, DNN neural network model, CNN neural network model, and unidirectional LSTM network model. Step S13: Input several input parameters from the validation set into each of the initial models to obtain different predicted output values. The validation set also includes the true output values corresponding to each input parameter. Step S14: Compare each of the predicted output values with the true output values to obtain the comparison results. Use the mean squared error and goodness of fit to filter each of the initial models corresponding to each of the predicted output values in the comparison results to obtain each of the filtered post-models. The filtered post-models include the Bi-LSTM network model, the CNN neural network model, and the unidirectional LSTM network model. Step S14 includes: Step 141: Calculate the mean square error and goodness of fit of each predicted output value relative to the true output value in the comparison results; Step 142: Substitute each of the mean square error values and the goodness-of-fit values into the preset sorting formula to obtain the sorting priority value; Step 143: Sort and filter the initial models according to the sorting priority value from high to low to obtain the filtered models. The sorting priority value of the Bi-LSTM network model is higher than that of the unidirectional LSTM network model, and the sorting priority value of the unidirectional LSTM network model is higher than that of the CNN neural network model. The sorting formula is configured as follows: ； in, Used to represent the sorting priority value; Used to represent the mean square error value; Used to represent the goodness-of-fit value; , These are used to represent the preset first and second coefficients, .
3. The SNCR denitrification prediction method integrating multiple neural networks according to claim 2, characterized in that, Step S2 includes: Step S21: Run the Bi-LSTM network model, the unidirectional LSTM network model, and the CNN neural network model in descending order of the sorting priority value, so that the Bi-LSTM network model, the unidirectional LSTM network model, and the CNN neural network model output different predicted values after screening respectively; Step S22: Calculate the absolute value of the error between each of the predicted values after screening and the actual NOx emissions, and process each of the absolute values of error to obtain the prediction accuracy of each of the predicted values after screening. Step S23: Perform correlation analysis on each of the post-screening predicted values, and calculate the first correlation coefficient of the remaining post-screening predicted values with respect to the post-screening predicted value with the highest prediction accuracy.
4. The SNCR denitrification prediction method integrating multiple neural networks according to claim 3, characterized in that, Step S3 includes: Step S31: Adjust the weight parameters between the input layer and the hidden layer in the corresponding post-screening model according to each of the first correlation coefficients and the prediction accuracy of each of the post-screening predicted values. Step S32: Retrain and adjust each of the screened models to obtain the optimized model.
5. The SNCR denitrification prediction method fused with multiple neural networks according to claim 4, characterized in that: In step S31, the first correlation coefficient and the prediction accuracy of each of the screened predicted values are input into a weight parameter adjustment formula to adjust the weight parameters between the input layer and the hidden layer in the screened model. The weight parameter adjustment formula is configured as follows: ； Used to represent the adjusted weight parameters; Used to represent the preset initial weight parameters; Used to represent the prediction accuracy of the post-screening predicted values; Used to represent the first correlation coefficient; , These are used to represent the preset first adjustment coefficient and the second adjustment coefficient, respectively.
6. The SNCR denitrification prediction method integrating multiple neural networks according to claim 5, characterized in that, Both the post-screening predicted value and the optimized predicted value are NOx output content. The input parameters of each post-screening model and each optimized model include several core parameters and several non-core parameters. The core parameters of each post-screening model are the same, and the core parameters include total coal quantity, total primary air volume, and total secondary air volume. The non-core parameters of each of the screened models are different. The non-core parameters of the Bi-LSTM network model include actual power generation, total air volume, right furnace bed pressure, left furnace bed pressure, average bed temperature on the right side of the furnace front wall, average bed temperature on the left side of the furnace front wall, average bed temperature on the right side of the furnace rear wall, average bed temperature on the left side of the furnace rear wall, right furnace outlet temperature, left furnace outlet temperature, furnace upper pressure 3, and furnace upper pressure 1. The non-core parameters of the unidirectional LSTM network model include unit load, burner swing angle, pressure difference between the wind box and the furnace, oxygen content at the furnace outlet, bed temperature, ammonia injection rate, and oxygen content at the air compressor inlet. The non-core parameters of the CNN neural network model include unit load, primary air pressure, secondary air damper opening at each level, burnout damper opening at each level, boiler outlet flue gas flow rate, boiler outlet flue gas temperature, boiler outlet flue gas oxygen content, boiler outlet NOx concentration, boiler inlet NOx concentration, SNCR inlet flue gas pressure, SNCR outlet flue gas pressure, power supply coal consumption, power generation coal consumption, and boiler efficiency. Then step S31 includes: Step S311: Perform correlation analysis between each of the non-core parameters and the corresponding post-screening predicted value, and calculate the second correlation coefficient of each of the non-core parameters with respect to the corresponding post-screening predicted value; Step S312: Adjust the weight parameters in the sieved model corresponding to the non-core parameters according to each of the second correlation coefficients; In the weight parameter adjustment formula, when the weight parameters in the screened model correspond to non-core parameters, the first adjustment coefficient and the second adjustment coefficient are configured as follows: in, , These are used to represent the initial values of the first adjustment coefficient and the second adjustment coefficient, respectively; Used to represent the second correlation coefficient; Used to indicate the number of the non-core parameters; When the weight parameters in the screened model correspond to the core parameters, the first adjustment coefficient and the second adjustment coefficient are configured as follows: in, This is used to indicate the number of the core parameters.
7. The SNCR denitrification prediction method fused with multiple neural networks according to claim 6, characterized in that, The core parameters and non-core parameters of each of the post-screening models and each of the optimized models are all obtained in real time by several sensors installed in the SNCR denitrification system; The non-core parameters include vulnerable sensor parameters. The vulnerable sensor parameters of the Bi-LSTM network model include the right furnace bed pressure, the left furnace bed pressure, the furnace upper pressure 3, and the furnace upper pressure 1. The vulnerable sensor parameters of the unidirectional LSTM network model include the air compressor inlet oxygen quantity. The vulnerable sensor parameters of the CNN neural network model include the boiler outlet flue gas flow rate. The step between step S311 and step S312 further includes step S31A, which determines whether each of the non-core parameters includes the vulnerable sensor parameter: If so, a corresponding sensor activation command will be generated based on the vulnerable sensor parameters, and the weight parameters in the sieved model corresponding to the vulnerable sensor parameters will be adjusted according to each of the second correlation coefficients, and then the process will proceed to step S312. If not, proceed to step S312; In the weight parameter adjustment formula, when the weight parameters in the sieved model correspond to the vulnerable sensor parameters, the first adjustment coefficient and the second adjustment coefficient are configured as follows: in, Used to indicate the number of parameters of the vulnerable sensor; Before running each of the optimization models, the pressure sensor located in the furnace, the temperature sensor located at the air compressor inlet, or the flow sensor located at the boiler outlet are adjusted to their on / off states for a preset period of time.
8. The SNCR denitrification prediction method fused with multiple neural networks according to claim 1, characterized in that, The step between step S4 and step S5 also includes: Step S4A: Determine whether the difference in prediction accuracy between the optimized model with the highest prediction accuracy and the screened model with the highest prediction accuracy is lower than a preset threshold. If so, the optimized model with the highest prediction accuracy will be output as the final model. If not, proceed to step S5.
9. The SNCR denitrification prediction method fused with multiple neural networks according to claim 7, characterized in that, Step S4 includes: Step S41: Calculate the absolute value of the error between each of the optimized predicted values and the actual NOx emissions, and process each of the absolute values of error to obtain the prediction accuracy of each of the optimized predicted values. Step S42: Perform correlation analysis on each of the optimized prediction values, and calculate the third correlation coefficient of the remaining optimized prediction values with respect to the optimized prediction value with the highest prediction accuracy; Then step S5 includes: Step S51: Adjust the weight parameters between the input layer and the hidden layer in the corresponding optimization model according to the prediction accuracy of each of the third correlation coefficients and each of the optimized prediction values. Step S52: Retrain each of the optimization models to obtain the adjusted optimization model.
10. The SNCR denitrification prediction method fused with multiple neural networks according to claim 9, characterized in that, Step S51 includes: Step S511: Perform correlation analysis between each non-core parameter and the corresponding optimized predicted value, and calculate the fourth correlation coefficient of each non-core parameter with respect to the corresponding optimized predicted value. Step S512: Adjust the weight parameters in the optimization model corresponding to the non-core parameters according to each fourth correlation coefficient; In the weight parameter adjustment formula, when the weight parameters in the optimization model correspond to non-core parameters, the first and second adjustment coefficients are configured as follows: in, Used to represent the fourth correlation coefficient; When the weight parameters in the optimization model correspond to the core parameters, the first and second adjustment coefficients are configured as follows: ; The step between step S511 and step S512 further includes step S51A, which determines whether each of the non-core parameters includes the vulnerable sensor parameter: If so, a corresponding sensor activation command will be generated based on the vulnerable sensor parameters, and the weight parameters in the optimization model corresponding to the vulnerable sensor parameters will be adjusted according to each of the second correlation coefficients, and then the process will proceed to step S512. If not, proceed to step S512; In the weight parameter adjustment formula, when the weight parameters in the optimization model correspond to vulnerable sensor parameters, the first adjustment coefficient and the second adjustment coefficient are configured as follows: Before each optimized model is run and adjusted, the pressure sensor located in the furnace, the temperature sensor located at the air compressor inlet, or the flow sensor located at the boiler outlet is adjusted in advance for the preset time period to adjust the opening and closing status.