Multi-objective combustion optimization method based on economic predictive control

Abstract

The invention provides a multi-target combustion optimization method based on economic predictive control, and aims to solve the problems that the power generation load and the in-furnace coal quality are frequently changed, conventional DCS control has defects in the aspects of air quantity and oxygen quantity optimization, boiler efficiency balancing, NOx emission, steam temperature control and the like, part of boiler parameters are difficult to accurately measure, and the existing combustion optimization technology is limited in applicability. The method comprises the following steps: collecting and preprocessing historical operation data, classifying and screening according to steady-state working conditions, and constructing a correlation model to perform preliminary combustion optimization; a load and coal quality change prediction model and a combustion process dynamic relation model are established according to unit characteristics and historical data; then, in the online secondary optimization stage, an initial parameter combination is selected according to the current state and the prediction state, and the combustion parameter adjustment amount meeting the multiple constraints is calculated through an economic prediction control algorithm and applied and corrected. Experiments show that the method can improve the boiler efficiency, reduce NOx emission, optimize steam temperature control, relieve the labor intensity of personnel and effectively improve the combustion performance.

Description

Technical Field

[0001] The invention relates to the field of thermal power equipment, and in particular to a multi-objective combustion optimization method based on economic forecasting control. Background Art

[0002] In the process of thermal power generation, boiler combustion faces many challenges due to frequent changes in power generation load and coal quality. Conventional DCS control has obvious shortcomings. The optimization of air volume and oxygen content is not fully considered in combustion control, and it is difficult to effectively balance boiler efficiency and NOx emissions. In addition, the quality of steam temperature control is low, and the amount of water spray for cooling is too large. At the same time, boiler-related parameters such as coal quality entering the furnace, primary air powder concentration, furnace temperature distribution, etc. are unmeasurable or inaccurate, making it difficult to apply combustion adjustment test results and foreign combustion optimization technologies. Most combustion optimization systems can only be used for open-loop guidance or steady-state conditions, and cannot meet actual needs. These problems have seriously affected the safety, economy and environmental protection of boiler combustion, and a more effective combustion optimization method is urgently needed to solve them. Summary of the invention

[0003] The main purpose of the present invention is to provide a multi-objective combustion optimization method based on economic forecasting control In order to solve the above technical problems, the technical solution adopted by the present invention is: a multi-objective combustion optimization method based on economic forecasting control, the method comprising: S1. Collect historical operation data from the DCS system and perform data preprocessing on the collected historical operation data; S2. Classify and screen the historical operation data according to steady-state operation conditions, and establish a correlation model between boiler efficiency and NOx emission and secondary air damper opening, burnout air damper opening, and flue gas oxygen content for each type of steady-state condition. The correlation model can be constructed using multiple linear regression or neural network; S3. Based on the unit operation characteristics and historical data, a prediction model for unit load and coal quality changes is established using time series analysis, grey prediction or machine learning methods; S4. Based on real-time data and prediction models, adjust combustion parameters online to ensure that the boiler is always in the optimal state under dynamically changing operating conditions; S5. Continuously monitor the optimization effect, adjust the optimization algorithm and model according to the actual situation, and ensure the long-term stability and optimization performance of the system.

[0004] In the preferred embodiment, in step S1: historical operation data is collected, and the historical operation data includes unit load, parameters related to coal quality entering the furnace, primary air powder concentration, furnace temperature distribution, flue gas oxygen content, secondary air volume of each layer, burnout air volume, coal feed volume of each layer burner, burner organization, primary air volume, boiler efficiency, NOx emission concentration, main steam temperature and reheat steam temperature; The data preprocessing step includes cleaning the collected data to remove outliers and erroneous data, and using data interpolation or regression analysis methods to estimate and supplement unmeasurable or uncertain parameters to ensure data accuracy and reliability.

[0005] In the preferred embodiment, the data preprocessing step in step S1 is specifically as follows: Use the set reasonable data range and 1-3 times standard deviation method to screen and remove outliers and erroneous data; for the furnace temperature distribution, if some position sensors fail or the measurement is inaccurate: for the furnace temperature distribution, use the Kriging interpolation method, based on the Gaussian model ,in is the temperature of the point to be estimated, is the temperature of the surrounding known points, is the weight coefficient, is the number of surrounding known points; The Gaussian model is used to estimate the unmeasurable or inaccurate position, and the partial least squares regression method is used for the parameters related to the coal quality entering the furnace. Make an estimated supplement, where is the coal quality parameter, are other measurable parameters, is the regression coefficient vector, It is an error term, which ensures the accuracy and reliability of data and provides a data basis for subsequent combustion optimization analysis.

[0006] In the preferred embodiment, in step S2: the association model is constructed using multivariate linear regression or neural network, and the association model optimization problem is solved by linear programming or nonlinear programming algorithm to obtain the optimal secondary air damper opening, burnout air damper opening and flue gas oxygen content combination that takes into account boiler efficiency and NOx emissions under different steady-state conditions and record the results as preliminary combustion optimization instructions.

[0007] In the preferred embodiment, the preliminary combustion optimization method in step S2 is: clustering the unit load in the historical operation data using a density-based spatial clustering algorithm, wherein for the unit load data point , by calculating the radius Whether the number of data points in the neighborhood of exceeds the minimum number of points MinPts, the core points are determined and clustered, and then the corresponding data subsets such as coal quality parameters, primary air powder concentration, furnace temperature distribution, flue gas oxygen content, secondary air volume of each layer, burnout air volume, coal feed volume of each layer burner, burner organization, boiler efficiency, NOx emission concentration, main steam temperature and reheat steam temperature under the steady-state operation conditions of each cluster are screened out; The correlation model building module uses a deep neural network DNN with multiple hidden layers to build a correlation model for each data subset of steady-state conditions. The input layer node is the secondary air door opening. , Burnout air damper opening , Flue gas oxygen content , the output layer node is the boiler efficiency and NOx emission concentration , by defining the loss function during training , and according to the back propagation algorithm according to the formula and Update the weights and biases of the neural network; in, and is the boiler efficiency and NOx emission concentration predicted by the neural network, and is the actual boiler efficiency and NOx emission concentration, is the weight coefficient, ranging from 0 to 1, determined by experiment; It is Layer to The weight of the layer, It is Bias of the layer; Optimization problem solving module, building objective function And the secondary air door opening satisfy , Burnout air damper opening satisfy , Flue gas oxygen content satisfy Constraints of in, is the weight coefficient, which is determined according to the actual requirements for boiler efficiency and environmental protection; A genetic algorithm is used to solve the optimization problem by performing selection operations through roulette selection method, crossover operations through single-point crossover, and mutation operations by changing individual gene values ​​with a certain mutation probability, so as to obtain the optimal secondary air damper opening, burnout air damper opening, and flue gas oxygen content combination that takes into account boiler efficiency and NOx emissions under different steady-state operating conditions, and the results are recorded as preliminary combustion optimization instructions.

[0008] In the preferred embodiment, step S3 further includes an economic forecasting model construction step, which uses a combination of mechanism analysis and data-driven methods to construct a dynamic relationship model between key parameters of the combustion process and boiler efficiency, NOx emissions, and steam temperature, wherein partial differential equations are used to describe the physical and chemical changes in the combustion process and actual operating data are used to perform parameter identification and model correction.

[0009] In the preferred solution, the steps of constructing the prediction model of unit load and coal quality change in step S3 are as follows: It includes a unit load forecasting unit, which is constructed using a long short-term memory network LSTM combined with a Kalman filter. The LSTM unit is based on the formula , , , , , Perform operations; in, They are input gate, forget gate, candidate memory unit, memory unit, output gate and hidden state. and are the corresponding weights and biases, is the sigmoid function, Represents element-wise multiplication; The Kalman filter unit is based on the state equation and Modify the LSTM output to predict the unit load; in, is the state vector, is the state transition matrix, is the control input matrix, is the control input vector, is the process noise vector, is the observation vector, is the observation matrix, is the observation noise vector. By continuously updating the Kalman gain, the prediction value is adjusted to improve the prediction accuracy; The coal quality change prediction unit adopts the grey Markov chain model. First, according to the formula Carry out grey prediction, and then use the statistically obtained transfer probability matrix The grey prediction results are corrected by Markov chain to realize the prediction of coal quality change; in, is the predicted value of the sequence generated by one accumulation, is the first value of the original sequence, and are the parameters obtained by solving the grey differential equation by the least square method; It has a dynamic relationship model construction unit for the combustion process, based on the mechanism through the mass conservation equation , energy conservation equation The system of partial differential equations describes the combustion process; Among them, the mass conservation equation is the density of the fuel, It's time. is the velocity vector, is the source term of fuel, indicating the rate at which fuel is generated or consumed; Energy conservation equation is the density of the mixture, is the total energy, It's pressure. is the thermal conductivity, is the temperature, is the heat source term; And use the least squares support vector machine to optimize the problem , Identify partial differential equation parameters, modify models, and construct dynamic relationship models between key parameters of the combustion process and boiler efficiency, NOx emissions, and steam temperature; in, is the weight vector, is the bias, is the error term, is the penalty parameter, is the kernel function that maps the input data to a high-dimensional space. is the output data. By solving the optimization problem, the estimated values ​​of the parameters in the partial differential equation are obtained, thus completing the model correction.

[0010] In the preferred embodiment, in step S4: According to the actual operating status of the current unit and the predicted future load and coal quality change trends, the initial optimization parameter combination is selected from the preliminary combustion optimization results.

[0011] In each control cycle, the economic prediction control algorithm is used for online optimization calculation; in each control cycle, according to the current operating parameters and predicted changes, the optimization model based on dynamic programming or model predictive control (MPC) is used to calculate the combustion parameter adjustment under the following constraints: NOx emission upper limit constraint, reheat steam temperature upper and lower limit constraint, each control quantity upper and lower limit constraint and rate constraint; Apply the calculated combustion parameter adjustment to the actual combustion control system, and use the feedback correction mechanism to collect actual combustion data for comparison and analysis with the predicted values; According to the deviation, the prediction model is corrected and the model parameters are optimized online to achieve closed-loop dynamic combustion optimization.

[0012] In the preferred embodiment, the online secondary optimization step of the multi-objective combustion optimization in step S4 is: The initial optimization parameter combination selection module uses Gaussian process regression GPR to establish the mapping relationship between the current unit operation status and the preliminary combustion optimization results. Its covariance function uses the radial basis function as follows: ; in, is the signal variance, which controls the smoothness of the function; is the length scale, which determines the decay rate of the correlation in the input space; is the noise variance, is the Kronecker function; Estimating hyperparameters by maximum likelihood estimation (MLE) method\Estimating hyperparameters by maximum likelihood estimation (MLE) method , and according to the predicted mean and variance, the initial optimization parameter combination is selected using a probability-based selection strategy; The online optimization calculation module builds an optimization model based on model predictive control (MPC), and the objective function is ; in, , The weight coefficients for boiler efficiency and NOx emission concentration at different time steps can be adjusted dynamically according to actual needs. For example, as time goes by, more emphasis is placed on the recent optimization effect, and the weight coefficient can be gradually increased; It is the penalty coefficient for the change of the control quantity to prevent the control quantity from changing drastically; For the The combustion parameters are in the time step The amount of change; and are the predicted boiler efficiency and NOx emission concentration at the time step The value of Constraints include: NOx upper limit constraint: ,in is the maximum allowed NOx emission concentration; Reheat steam temperature upper and lower limit constraints: ,in and The time steps The lower and upper limits of the reheat steam temperature; Upper and lower limit constraints of each control quantity: ,in and For the The lower and upper limits of the combustion parameters; Rate constraints of each control variable: ,in For the The maximum variation allowed for a combustion parameter per unit time; The above optimization problem is solved by using the dynamic programming algorithm. In the dynamic programming, the state variables are defined as the current unit operation state and the combustion parameter combination, and the decision variables are the adjustment amount of the combustion parameters at the next moment. By constructing the value function , which means from the time step Start in state Take decision The optimal target value after , its recursive formula is: ; in, is the objective function value of the current time step, Based on the system dynamic model and current decision Get the state at the next moment; start the reverse solution from the last time step to obtain the optimal combustion parameter adjustment strategy; Feedback correction: Collect actual combustion data, including actual boiler efficiency , NOx emission concentration , Reheat steam temperature And the actual values ​​of combustion parameters ; Calculate the deviation between the predicted value and the actual value: ; ; ; ; The prediction model and optimization model are corrected online using the deviation; for the prediction model, the new model parameters are obtained using the recursive least squares method; the prediction model is assumed to be a linear model ,in To predict the output, is the input data matrix, is the model parameter vector, and the update formula of RLS is: ; ; ; in, is the Kalman gain vector, is the covariance matrix, and the initial value can be set based on prior knowledge; For the optimization model, the weight coefficients and constraints in the objective function are adjusted according to the deviation to achieve closed-loop dynamic combustion optimization.

[0013] In the preferred solution, in step S5: continuously monitor and evaluate the operation effect of the combustion optimization system, including boiler efficiency improvement, NOx emission reduction, main steam temperature and reheat steam temperature control accuracy; Adjust and improve the optimization algorithm and model based on the evaluation results. When the optimization effect of a specific operating condition is not good, analyze the cause and take measures such as supplementing data, improving the model structure or adjusting the constraint range to improve the combustion optimization performance; The algorithm steps include: Operation effect monitoring and evaluation module, using hierarchical analysis method to construct judgment matrix , through the formula Solving for the largest eigenvalue and the eigenvector And normalize to get the weight vector , calculate the scores of each indicator such as boiler efficiency improvement score , NOx emission reduction score , Main steam temperature control accuracy score , Reheat steam temperature control accuracy score , comprehensive evaluation score ,when Trigger adjustment and improvement process when necessary; Adjust and improve the module, use principal component analysis (PCA) to calculate the covariance matrix The eigenvalue of and the eigenvector Perform dimensionality reduction processing. If it is a data problem, use the nearest neighbor interpolation method (in ) Supplementary data, if it is a model structure problem, use adaptive model adjustment algorithm and transfer learning method to improve it, if it is a constraint problem, adjust the constraint range according to the data distribution, such as when the calculated data is near the boundary When adjusting the constraint to (in , ), thereby improving combustion optimization performance.

[0014] The present invention provides a multi-objective combustion optimization method based on economic forecasting control, which has many beneficial effects. It can automatically adjust the combustion operation parameters according to the changes in the unit load and coal quality, and effectively improve the boiler efficiency on the basis of ensuring the safety of the unit operation, which can be increased by about 0.5% on average and reduce coal consumption by 1.5-2g; at the same time, it significantly reduces the flue gas NOx emission, with an average reduction of about 10%-20%, and can control the reheat steam temperature within the required range, improving the operating level of the main steam temperature and the reheat steam temperature. In addition, the method reduces the arbitrariness of combustion adjustment and the labor intensity of the operating personnel. Through closed-loop optimization control, no manual intervention is required, and the model has strong adaptive ability, can adapt well to the changes in load and coal types, is suitable for dynamic optimization of variable load, and provides strong support for the efficient, clean and stable operation of thermal power generation. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] The present invention will be further described below in conjunction with the accompanying drawings and embodiments: Figure 1 It is the main structural diagram of the cleaning process of the present invention; DETAILED DESCRIPTION Example 1 like Figure 1 As shown, a multi-objective combustion optimization method based on economic forecasting control is characterized in that the method includes: S1. Collect historical operation data from the DCS system and perform data preprocessing on the collected historical operation data; S2. Classify and screen the historical operation data according to steady-state operation conditions, and establish a correlation model between boiler efficiency and NOx emission and secondary air damper opening, burnout air damper opening, and flue gas oxygen content for each type of steady-state condition. The correlation model can be constructed using multiple linear regression or neural network; S3. Based on the unit operation characteristics and historical data, a prediction model for unit load and coal quality changes is established using time series analysis, grey prediction or machine learning methods; S4. Based on real-time data and prediction models, adjust combustion parameters online to ensure that the boiler is always in the optimal state under dynamically changing conditions; S5. Continuously monitor the optimization effect, adjust the optimization algorithm and model according to the actual situation, and ensure the long-term stability and optimization performance of the system.

[0016] Data collection: The system collects historical operation data in real time. The historical operation data covers but is not limited to unit load, parameters related to coal quality entering the furnace (such as coal type, calorific value, volatile matter, etc.), primary air powder concentration, furnace temperature distribution, flue gas oxygen content, secondary air volume of each layer, burnout air volume, coal supply to each layer burner, burner organization (milling and shutdown combination), primary air volume, boiler efficiency, NOx emission concentration, main steam temperature and reheat steam temperature.

[0017] Data preprocessing: Clean the collected data, remove outliers and erroneous data, and ensure data quality. For parameters that cannot be measured or cannot be measured accurately, use data interpolation or regression analysis methods to estimate and supplement to ensure the accuracy and reliability of the data.

[0018] Steady-state operating condition classification screening: historical operation data is classified and screened according to steady-state operating conditions, and a correlation model between boiler efficiency and NOx emissions and secondary air damper opening, burnout air damper opening, and flue gas oxygen content is established for each type of steady-state operating condition. The correlation model can be constructed using multiple linear regression or neural network.

[0019] Optimization solution: Solve the optimization problem of the associated model through linear programming or nonlinear programming algorithm to obtain the best combination of secondary air damper opening, burnout air damper opening and flue gas oxygen content that takes into account boiler efficiency and NOx emissions under different steady-state conditions, and record the results as preliminary combustion optimization instructions.

[0020] Prediction model construction: Based on the unit operating characteristics and historical data, time series analysis, grey prediction or machine learning methods are used to establish a prediction model for unit load and coal quality changes.

[0021] Dynamic relationship model construction: By combining mechanism analysis and data-driven methods, a dynamic relationship model between key parameters of the combustion process and boiler efficiency, NOx emissions, and steam temperature is constructed. Partial differential equations are used to describe the physical and chemical changes in the combustion process, and actual operating data is combined to perform parameter identification and model correction to ensure the accuracy and adaptability of the model.

[0022] Initial parameter selection: According to the actual operating status of the current unit and the predicted future load and coal quality change trend, the initial optimization parameter combination is selected from the preliminary combustion optimization results.

[0023] Online optimization calculation: In each control cycle, the economic prediction control algorithm is used for online optimization calculation. In each control cycle, according to the current operating parameters and predicted changes, the optimization model based on dynamic programming or model predictive control (MPC) is used to calculate the combustion parameter adjustment under the following constraints: NOx emission upper limit constraints; Upper and lower limits of reheat steam temperature; Upper and lower limit constraints and rate constraints for each control variable; Closed-loop dynamic optimization: Apply the calculated combustion parameter adjustment to the actual combustion control system, and use the feedback correction mechanism to collect actual combustion data and compare and analyze the predicted values. According to the deviation, the prediction model is corrected online and the model parameters are optimized to achieve closed-loop dynamic combustion optimization.

[0024] Continuous monitoring and evaluation: Continuously monitor and evaluate the operating effect of the combustion optimization system, including boiler efficiency improvement, NOx emission reduction, main steam temperature and reheat steam temperature control accuracy, etc.

[0025] Optimization and adjustment: According to the evaluation results, the optimization algorithm and model are adjusted and improved. When the optimization effect under specific working conditions is not good, the reasons are analyzed and measures such as supplementing data, improving model structure or adjusting constraint range are taken to improve combustion optimization performance.

[0026] Example 2 Further illustrate with reference to Example 1, Figure 1 In the structure shown, in step SI: historical operation data is collected, and the historical operation data includes unit load, parameters related to coal quality entering the furnace, primary air powder concentration, furnace temperature distribution, flue gas oxygen content, secondary air volume of each layer, burnout air volume, coal feed volume of each layer burner, burner organization, primary air volume, boiler efficiency, NOx emission concentration, main steam temperature and reheat steam temperature; The data preprocessing step includes cleaning the collected data to remove outliers and erroneous data, and using data interpolation or regression analysis methods to estimate and supplement unmeasurable or uncertain parameters to ensure data accuracy and reliability.

[0027] In the preferred embodiment, the data preprocessing step in step S1 is specifically as follows: Use the set reasonable data range and 1-3 times standard deviation method to screen and remove outliers and erroneous data; for the furnace temperature distribution, if some position sensors fail or the measurement is inaccurate: for the furnace temperature distribution, use the Kriging interpolation method, based on the Gaussian model ,in is the temperature of the point to be estimated, is the temperature of the surrounding known points, is the weight coefficient, is the number of surrounding known points; The Gaussian model is used to estimate the unmeasurable or inaccurate position, and the partial least squares regression method is used for the parameters related to the coal quality entering the furnace. Make an estimated supplement, where is the coal quality parameter, are other measurable parameters, is the regression coefficient vector, It is an error term, which ensures the accuracy and reliability of data and provides a data basis for subsequent combustion optimization analysis.

[0028] Establish a data connection interface with the DCS system, read and store historical operation data from the DCS system at a predetermined time interval (for example, once every 5 minutes). Ensure data integrity, that is, each acquisition covers unit load, coal quality parameters (such as low calorific value of coal), , volatile matter , ash content etc.), primary air powder concentration , furnace temperature distribution (where $x, y, z$ are the coordinates of different positions in the furnace), flue gas oxygen content , Secondary air volume of each floor ( Indicates different layers), burnout air volume , Coal supply to each layer burner , burner organization (using binary code to represent the combination of grinding, starting and stopping, such as 0 for stopping and 1 for starting), primary air volume , boiler efficiency , NOx emission concentration , Main steam temperature and reheat steam temperature .

[0029] Set a reasonable range for the data, for example, the unit load should be within a certain percentage range of the rated load (such as 20%-120%), the flue gas oxygen content is generally 3%-8%, etc. Data points outside these ranges are considered outliers and deleted.

[0030] Calculate the mean of each parameter using statistical methods, such as the 1-3 standard deviation method and standard deviation , for the data point ,like , then the data point is considered as an outlier and deleted.

[0031] Supplementary information for estimation of unmeasurable or inaccurate parameters: For furnace temperature distribution, if some position sensors fail or the measurement is inaccurate, Kriging interpolation is used for estimation. Assuming that the furnace temperature distribution conforms to a certain spatial correlation model, such as the Gaussian model, the formula is: ; in, It is the point to be estimated The temperature, is the temperature of the surrounding known points, is the weight coefficient obtained by solving a system of linear equations based on the variogram. is the number of known points around. The variogram is used to describe the change of the correlation of temperature values ​​between two points in space with distance.

[0032] For parameters related to coal quality entering the furnace, if there are large errors or omissions in the measurement, the partial least squares regression (PLS regression) method is used for estimation. (such as low heat) and other measurable parameters (such as primary air powder concentration, flue gas oxygen content, etc.) has a linear relationship, and its model is: ; in, is the regression coefficient vector, by minimizing the residual sum of squares get, is the predicted value, is the error term. In PLS regression, and Decomposition is performed and latent variables are extracted to establish a regression model to improve the accuracy and stability of the model.

[0033] Through the above steps, the process of collecting historical operating data from the DCS system and preprocessing it is completed, providing a reliable data basis for subsequent combustion optimization analysis.

[0034] Example 3 Further illustrate with reference to Example 1, Figure 1 In the structure shown, in step S2: the association model is constructed by multivariate linear regression or neural network, and the association model optimization problem is solved by linear programming or nonlinear programming algorithm to obtain the best secondary air damper opening, burnout air damper opening and flue gas oxygen content combination that takes into account boiler efficiency and NOx emissions under different steady-state conditions and record the results as preliminary combustion optimization instructions.

[0035] In the preferred embodiment, the preliminary combustion optimization method in step S2 is: clustering the unit load in the historical operation data using a density-based spatial clustering algorithm, wherein for the unit load data point , by calculating the radius Whether the number of data points in the neighborhood of exceeds the minimum number of points MinPts, the core points are determined and clustered, and then the corresponding data subsets such as coal quality parameters, primary air powder concentration, furnace temperature distribution, flue gas oxygen content, secondary air volume of each layer, burnout air volume, coal feed volume of each layer burner, burner organization, boiler efficiency, NOx emission concentration, main steam temperature and reheat steam temperature under the steady-state operation conditions of each cluster are screened out; The correlation model building module uses a deep neural network DNN with multiple hidden layers to build a correlation model for each data subset of steady-state conditions. The input layer node is the secondary air door opening. , Burnout air damper opening , Flue gas oxygen content , the output layer node is the boiler efficiency and NOx emission concentration , by defining the loss function during training , and according to the back propagation algorithm according to the formula and Update the weights and biases of the neural network; in, and is the boiler efficiency and NOx emission concentration predicted by the neural network, and is the actual boiler efficiency and NOx emission concentration, is the weight coefficient, ranging from 0 to 1, determined by experiment; It is Layer to The weight of the layer, It is Bias of the layer; Optimization problem solving module, building objective function And the secondary air door opening satisfy , Burnout air damper opening satisfy , Flue gas oxygen content satisfy Constraints of in, is the weight coefficient, which is determined according to the actual requirements for boiler efficiency and environmental protection; A genetic algorithm is used to solve the optimization problem by performing selection operations through roulette selection method, crossover operations through single-point crossover, and mutation operations by changing individual gene values ​​with a certain mutation probability, so as to obtain the optimal secondary air damper opening, burnout air damper opening, and flue gas oxygen content combination that takes into account boiler efficiency and NOx emissions under different steady-state operating conditions, and the results are recorded as preliminary combustion optimization instructions.

[0036] Steady-state operating condition classification screening: First, cluster analysis is performed on the unit load in the historical operating data. The density-based spatial clustering algorithm (DBSCAN) is used. Its core idea is based on the density of data points. If the density of data points in an area exceeds a certain threshold, these points are divided into a cluster. , calculate the density around each point, if the radius is If the number of data points in the neighborhood of exceeds the minimum number of points $MinPts$, then the point is a core point. By continuously expanding the neighborhood of the core point, the interconnected core points and their boundary points are divided into a cluster, representing a load range under a steady-state operating condition.

[0037] For each cluster (i.e., a steady-state operating condition), other relevant data belonging to this operating condition are screened out, including data subsets such as coal quality parameters entering the furnace, primary air powder concentration, furnace temperature distribution, flue gas oxygen content, secondary air volume of each layer, burnout air volume, coal feed amount of each layer burner, burner organization, boiler efficiency, NOx emission concentration, main steam temperature and reheat steam temperature.

[0038] Association model construction: For each data subset of steady-state conditions, a neural network is used to build an association model. Here, a deep neural network (DNN) with multiple hidden layers is selected, for example, 3 hidden layers, each with 50 neurons. The input layer node is the secondary air door opening , Burnout air damper opening , Flue gas oxygen content , the output layer node is the boiler efficiency and NOx emission concentration .

[0039] The training process of the neural network is as follows: the loss function is defined as a weighted combination of mean square error (MSE) and cross entropy (CE), and the formula is: ; in, and is the boiler efficiency and NOx emission concentration predicted by the neural network, and is the actual boiler efficiency and NOx emission concentration, is the weight coefficient, ranging from 0 to 1, and is determined through experiments.

[0040] Backpropagation is used to update the weights and biases of the neural network. In each iteration, the gradient of the loss function for each weight and bias is calculated, and then the gradient is updated according to a certain learning rate. Update weights and biases, the formula is as follows: ; ; in, It is Layer to The weight of the layer, It is The bias of the layer.

[0041] Optimization problem solving: The objective function of the optimization problem is constructed as the weighted sum of maximizing boiler efficiency and minimizing NOx emission concentration. The formula is: ; in, It is a weight coefficient, which is determined according to the actual requirements for boiler efficiency and environmental protection.

[0042] The constraints include the secondary air damper opening, the burnout air damper opening, and the upper and lower limits of the flue gas oxygen content, for example: ; ; ; The genetic algorithm is used to solve the optimization problem. First, the initial population is randomly generated, and each individual represents a combination of secondary air damper opening, burnout air damper opening, and flue gas oxygen content. Then, the fitness value of each individual is calculated, that is, the objective function The population is continuously evolved through selection, crossover and mutation operations until the termination condition is met (such as reaching the maximum number of iterations or the fitness value converges). The selection operation uses the roulette wheel selection method to allocate the selection probability according to the fitness value of the individual; the crossover operation uses a single-point crossover to randomly select a crossover point to exchange some genes of two individuals; the mutation operation changes the value of a gene of an individual with a certain mutation probability.

[0043] Finally, the optimal combination of secondary air damper opening, burnout air damper opening and flue gas oxygen content that takes into account both boiler efficiency and NOx emissions under different steady-state operating conditions is obtained, and the results are recorded as preliminary combustion optimization instructions.

[0044] Example 4 Further illustrate with reference to Example 1, Figure 1 In the structure shown, step S3 further includes an economic forecasting model construction step, which uses a combination of mechanism analysis and data-driven methods to construct a dynamic relationship model between key parameters of the combustion process and boiler efficiency, NOx emissions, and steam temperature, wherein partial differential equations are used to describe the physical and chemical changes in the combustion process and actual operating data are combined for parameter identification and model correction.

[0045] In the preferred solution, the steps of constructing the prediction model of unit load and coal quality change in step S3 are as follows: It includes a unit load forecasting unit, which is constructed using a long short-term memory network LSTM combined with a Kalman filter. The LSTM unit is based on the formula , , , , , Perform operations; in, They are input gate, forget gate, candidate memory unit, memory unit, output gate and hidden state. and are the corresponding weights and biases, is the sigmoid function, Represents element-wise multiplication; The Kalman filter unit is based on the state equation and Modify the LSTM output to predict the unit load; in, is the state vector, is the state transition matrix, is the control input matrix, is the control input vector, is the process noise vector, is the observation vector, is the observation matrix, is the observation noise vector. By continuously updating the Kalman gain, the prediction value is adjusted to improve the prediction accuracy; The coal quality change prediction unit adopts the grey Markov chain model. First, according to the formula Carry out grey prediction, and then use the statistically obtained transfer probability matrix The grey prediction results are corrected by Markov chain to realize the prediction of coal quality change; in, is the predicted value of the sequence generated by one accumulation, is the first value of the original sequence, and are the parameters obtained by solving the grey differential equation by the least square method; It has a dynamic relationship model construction unit for the combustion process, based on the mechanism through the mass conservation equation , energy conservation equation The system of partial differential equations describes the combustion process; Among them, the mass conservation equation is the density of the fuel, It's time. is the velocity vector, is the source term of fuel, indicating the rate at which fuel is generated or consumed; Energy conservation equation is the density of the mixture, is the total energy, It's pressure. is the thermal conductivity, is the temperature, is the heat source term; And use the least squares support vector machine to optimize the problem , Identify partial differential equation parameters, modify models, and construct dynamic relationship models between key parameters of the combustion process and boiler efficiency, NOx emissions, and steam temperature; in, is the weight vector, is the bias, is the error term, is the penalty parameter, is the kernel function that maps the input data to a high-dimensional space. is the output data. By solving the optimization problem, the estimated values ​​of the parameters in the partial differential equation are obtained, thus completing the model correction.

[0046] Construction of unit load and coal quality change prediction model: For unit load prediction, a method combining long short-term memory network (LSTM) and Kalman filter is used. LSTM is used to capture the long-term dependency of unit load time series, and its core formula is: ; ; ; ; ; ; in, They are input gate, forget gate, candidate memory unit, memory unit, output gate and hidden state. and are the corresponding weights and biases, is the sigmoid function, Represents element-wise multiplication.

[0047] Kalman filtering is used to correct the prediction results of LSTM, and its state equation is: ; ; in, is the state vector, is the state transition matrix, is the control input matrix, is the control input vector, is the process noise vector, is the observation vector, is the observation matrix, is the observation noise vector. By continuously updating the Kalman gain, the predicted value is adjusted to improve the prediction accuracy.

[0048] For the prediction of coal quality changes, the grey Markov chain model is used. First, the coal quality parameters (such as low calorific value, volatile matter, etc.) are grey predicted, and the basic formula is: ; in, is the predicted value of the sequence generated by one accumulation, is the first value of the original sequence, and It is the parameter obtained by solving the grey differential equation by the least square method.

[0049] Then, the Markov chain is used to correct the state transition of the grey prediction results. The state interval of coal quality parameters is defined, and the transition probability matrix between different states is statistically calculated based on historical data. , through the current state and transition probability matrix, the state interval of the coal quality parameters at the next moment is predicted, and the predicted value is further corrected.

[0050] Construction of dynamic relationship model between key parameters of combustion process and boiler efficiency, NOx emission and steam temperature Based on the mechanism analysis, the partial differential equations of the combustion process are established using the laws of conservation of mass, energy and momentum. For example, the mass conservation equation for fuel combustion is: ; in, is the density of the fuel, It's time. is the velocity vector, is the source term of fuel, representing the rate at which fuel is generated or consumed.

[0051] For the energy conservation equation: ; in, is the density of the mixture, is the total energy, It's pressure. is the thermal conductivity, is the temperature, is the heat source term.

[0052] The parameters in the above partial differential equations are identified using a data-driven approach. For example, the least squares support vector machine (LS SVM) method is used to construct a training model with actual operating data as input and unknown parameters in the equation as output. The optimization problem of LS SVM is: ; ; in, is the weight vector, is the bias, is the error term, is the penalty parameter, is the kernel function that maps the input data to a high-dimensional space. is the output data. By solving the optimization problem, the estimated values ​​of the parameters in the partial differential equation are obtained, thus completing the model correction.

[0053] Example 5 Further illustrate with reference to Example 1, Figure 1 In the structure shown, in step S4: according to the actual operating status of the current unit and the predicted future load and coal quality change trend, an initial optimization parameter combination is selected from the preliminary combustion optimization results.

[0054] In each control cycle, the economic prediction control algorithm is used for online optimization calculation; in each control cycle, according to the current operating parameters and predicted changes, the optimization model based on dynamic programming or model predictive control (MPC) is used to calculate the combustion parameter adjustment under the following constraints: NOx emission upper limit constraint, reheat steam temperature upper and lower limit constraint, each control quantity upper and lower limit constraint and rate constraint; Apply the calculated combustion parameter adjustment to the actual combustion control system, and use the feedback correction mechanism to collect actual combustion data for comparison and analysis with the predicted values; According to the deviation, the prediction model is corrected and the model parameters are optimized online to achieve closed-loop dynamic combustion optimization.

[0055] In the preferred embodiment, the online secondary optimization step of the multi-objective combustion optimization in step S4 is: The initial optimization parameter combination selection module uses Gaussian process regression GPR to establish the mapping relationship between the current unit operation status and the preliminary combustion optimization results. Its covariance function uses the radial basis function as follows: ; in, is the signal variance, which controls the smoothness of the function; is the length scale, which determines the decay rate of the correlation in the input space; is the noise variance, is the Kronecker function; Estimating hyperparameters by maximum likelihood estimation (MLE) method\Estimating hyperparameters by maximum likelihood estimation (MLE) method , and according to the predicted mean and variance, the initial optimization parameter combination is selected using a probability-based selection strategy; Initial optimization parameter combination selection: First, a mapping relationship model between the current unit operating state and the preliminary combustion optimization results is established. Using the Gaussian Process Regression (GPR) method, it is assumed that the current unit operating state vector is (in is the unit load, The coal entering the furnace has a low calorific value. For volatile matter, is the main steam temperature, is the reheat steam temperature, etc.), the parameter combination vector in the preliminary combustion optimization result is: (in is the secondary air door opening, is the burnout air damper opening, oxygen content in flue gas, etc.).

[0056] The covariance function of GPR uses the radial basis function (RBF): ; in, is the signal variance, which controls the smoothness of the function; is the length scale, which determines the decay rate of the correlation in the input space; is the noise variance, is the Kronecker function.

[0057] By training the GPR model, the hyperparameters are estimated using the Maximum Likelihood Estimation (MLE) method. In the prediction stage, for the current unit operating status , the predicted preliminary combustion optimization parameter combination The mean and variance of are: ; ; in, is the input matrix of training data, is the output vector of training data, is the covariance vector between the test point and the training point, is the covariance matrix between training points.

[0058] According to the predicted mean and variance, a probability-based selection strategy is adopted, such as selecting a preliminary combustion optimization parameter combination within a certain confidence interval (such as a 95% confidence interval) as the initial optimization parameter combination to balance the relationship between exploration and utilization.

[0059] The online optimization calculation module builds an optimization model based on model predictive control (MPC), and the objective function is ; in, , The weight coefficients for boiler efficiency and NOx emission concentration at different time steps can be adjusted dynamically according to actual needs. For example, as time goes by, more emphasis is placed on the recent optimization effect, and the weight coefficient can be gradually increased; It is the penalty coefficient for the change of the control quantity to prevent the control quantity from changing drastically; For the The combustion parameters are in the time step The amount of change; and are the predicted boiler efficiency and NOx emission concentration at the time step The value of Constraints include: NOx upper limit constraint: ,in is the maximum allowed NOx emission concentration; Reheat steam temperature upper and lower limit constraints: ,in and The time steps The lower and upper limits of the reheat steam temperature; Upper and lower limit constraints of each control quantity: ,in and For the The lower and upper limits of the combustion parameters; Rate constraints of each control variable: ,in For the The maximum variation allowed for a combustion parameter per unit time; The above optimization problem is solved by using the dynamic programming algorithm. In the dynamic programming, the state variables are defined as the current unit operation state and the combustion parameter combination, and the decision variables are the adjustment amount of the combustion parameters at the next moment. By constructing the value function , which means from the time step Start in state Take decision The optimal target value after , its recursive formula is: ; in, is the objective function value of the current time step, Based on the system dynamic model and current decision Get the state at the next moment; start the reverse solution from the last time step to obtain the optimal combustion parameter adjustment strategy; Feedback correction: Collect actual combustion data, including actual boiler efficiency , NOx emission concentration , Reheat steam temperature And the actual values ​​of combustion parameters ; Calculate the deviation between the predicted value and the actual value: ; ; ; ; The prediction model and optimization model are corrected online using the deviation; for the prediction model, the new model parameters are obtained using the recursive least squares method; the prediction model is assumed to be a linear model ,in To predict the output, is the input data matrix, is the model parameter vector, and the update formula of RLS is: ; ; ; in, is the Kalman gain vector, is the covariance matrix, and the initial value can be set based on prior knowledge; For the optimization model, the weight coefficients and constraints in the objective function are adjusted according to the deviation to achieve closed-loop dynamic combustion optimization.

[0060] Example 6 Further illustrate with reference to Example 1, Figure 1 In the structure shown, in step S5: continuously monitor and evaluate the operation effect of the combustion optimization system, including the improvement of boiler efficiency, the reduction of NOx emissions, and the control accuracy of the main steam temperature and the reheat steam temperature; Adjust and improve the optimization algorithm and model based on the evaluation results. When the optimization effect of a specific operating condition is not good, analyze the cause and take measures such as supplementing data, improving the model structure or adjusting the constraint range to improve the combustion optimization performance; The algorithm steps include: Operation effect monitoring and evaluation module, using hierarchical analysis method to construct judgment matrix , through the formula Solving for the largest eigenvalue and the eigenvector And normalize to get the weight vector , calculate the scores of each indicator such as boiler efficiency improvement score , NOx emission reduction score , Main steam temperature control accuracy score , Reheat steam temperature control accuracy score , comprehensive evaluation score ,when Trigger adjustment and improvement process when necessary; Adjust and improve the module, use principal component analysis (PCA) to calculate the covariance matrix The eigenvalue of and the eigenvector Perform dimensionality reduction processing. If it is a data problem, use the nearest neighbor interpolation method (in ) Supplementary data, if it is a model structure problem, use adaptive model adjustment algorithm and transfer learning method to improve it, if it is a constraint problem, adjust the constraint range according to the data distribution, such as when the calculated data is near the boundary When adjusting the constraint to (in , ), thereby improving combustion optimization performance.

[0061] Operation effect monitoring and evaluation Establish a multi-index comprehensive evaluation system and use the Analytic Hierarchy Process (AHP) to determine the weight of each indicator; first construct a judgment matrix ,element Indicator Relative to the indicator For example, for boiler efficiency improvement (Indicator 1) and NOx emission reduction (Indicator 2), if boiler efficiency improvement is considered to be slightly more important than NOx emission reduction, then , ; Calculate the maximum eigenvalue of the judgment matrix and the corresponding eigenvector , through the formula ( ) to solve for the eigenvector Normalize and get the weight vector of each indicator ; For each time interval (e.g. every hour), collect the actual boiler efficiency , NOx emissions , Main steam temperature and reheat steam temperature ; Calculate the score of each indicator, such as boiler efficiency improvement score (in is the baseline boiler efficiency), NOx emission reduction score (in is the baseline NOx emission); Main steam temperature control accuracy score (in is the set main steam temperature, is the allowable standard deviation of main steam temperature deviation), reheat steam temperature control accuracy score (in is the set reheat steam temperature, is the allowable standard deviation of reheat steam temperature); Comprehensive evaluation score ,in Assign scores to each indicator; set thresholds ,when When the combustion optimization system is considered to be operating poorly, the subsequent adjustment and improvement process is triggered; Adjustment and improvement: When the operation effect is not good, first analyze the cause; use the principal component analysis (PCA) method to reduce the dimension of the relevant data and find the main change trends and abnormal points in the data; assume that the collected data matrix is , whose covariance matrix is (in is the number of data samples), calculate the eigenvalues ​​of the covariance matrix and the eigenvector , and sort them according to the size of the eigenvalue; Before selection The eigenvectors corresponding to the eigenvalues ​​constitute the projection matrix , projecting the original data into a low-dimensional space ; By observing the data distribution and abnormal points in the low-dimensional space and combining them with the actual operation conditions, we can determine the possible reasons, such as whether there are missing or abnormal data under certain working conditions that lead to inaccurate models, or unreasonable constraints that limit the optimization effect, etc.; If it is a data problem, the data expansion algorithm is used to supplement the data; for example, for the missing coal quality data, the nearest neighbor interpolation method based on similar operating conditions is used to interpolate and supplement the data based on other operating conditions with similar unit load, combustion parameters and steam temperature; assuming that the missing data point is ,turn up Nearest neighbor data points , whose distance metric is Euclidean distance , the supplementary data is , where the weight ; If it is a model structure problem, an adaptive model adjustment algorithm is used; for example, for the dynamic relationship model of the combustion process, if it is found that its prediction error for certain working conditions is large, the number of hidden layers or nodes of the neural network is increased, or the topological structure of the model is changed; the training of the new model adopts the transfer learning method, using some parameters of the original model as initialization and fine-tuning on the new data set to speed up the training speed and improve the model performance; If it is a constraint problem, adjust the constraint range based on the statistical analysis of the actual operation data; for example, for the upper and lower limit constraints of the combustion parameters, if it is found that the actual operation often approaches or exceeds the constraint boundary, but it does not have a negative impact on the system operation, the constraint range can be appropriately relaxed; assuming that the original constraint is , by calculating the distribution of actual data in the interval, such as calculating the proportion of data near the boundary ,if (set threshold), then the adjustment constraint is ,in , , and Determined based on data distribution and actual experience.

[0062] The above-mentioned embodiments are only preferred technical solutions of the present invention and should not be regarded as limitations of the present invention. The protection scope of the present invention should be the technical solutions recorded in the claims, including equivalent replacement solutions of the technical features in the technical solutions recorded in the claims.

Claims

1. A multi-objective combustion optimization method based on economic forecasting control, characterized by: The method includes: S1. Collect historical operation data from the DCS system and perform data preprocessing on the collected historical operation data; S2. Classify and screen the historical operation data according to steady-state operation conditions, and establish a correlation model between boiler efficiency and NOx emission and secondary air damper opening, burnout air damper opening, and flue gas oxygen content for each type of steady-state condition. The correlation model can be constructed using multiple linear regression or neural network; S3. Based on the unit operation characteristics and historical data, a prediction model for unit load and coal quality changes is established using time series analysis, grey prediction or machine learning methods; S4. Based on real-time data and prediction models, adjust combustion parameters online to ensure that the boiler is always in the optimal state under dynamically changing operating conditions; S5. Continuously monitor the optimization effect, adjust the optimization algorithm and model according to the actual situation, and ensure the long-term stability and optimization performance of the system.

2. The multi-objective combustion optimization method based on economic forecasting control according to claim 1 is characterized in that: In step S1: Collect historical operation data, including unit load, coal quality parameters, primary air powder concentration, furnace temperature distribution, flue gas oxygen content, secondary air volume of each layer, burnout air volume, coal feed volume of each layer burner, burner organization, primary air volume, boiler efficiency, NOx emission concentration, main steam temperature and reheat steam temperature; The data preprocessing step includes cleaning the collected data to remove outliers and erroneous data, and using data interpolation or regression analysis methods to estimate and supplement unmeasurable or uncertain parameters to ensure data accuracy and reliability.

3. The multi-objective combustion optimization method based on economic forecasting control according to claim 2 is characterized in that: The data preprocessing step in step S1 is specifically as follows: Use the set reasonable data range and 1-3 times standard deviation method to screen and remove outliers and erroneous data; for the furnace temperature distribution, if some position sensors fail or the measurement is inaccurate: for the furnace temperature distribution, use the Kriging interpolation method, based on the Gaussian model ,in is the temperature of the point to be estimated, is the temperature of the surrounding known points, is the weight coefficient, is the number of surrounding known points; The Gaussian model is used to estimate the unmeasurable or inaccurate position, and the partial least squares regression method is used for the parameters related to the coal quality entering the furnace. Make an estimated supplement, where is the coal quality parameter, are other measurable parameters, is the regression coefficient vector, It is an error term, which ensures the accuracy and reliability of data and provides a data basis for subsequent combustion optimization analysis.

4. The multi-objective combustion optimization method based on economic forecasting control according to claim 1 is characterized in that: in step S2: The correlation model is constructed using multivariate linear regression or neural network, and the correlation model optimization problem is solved through linear programming or nonlinear programming algorithm to obtain the optimal secondary air damper opening, burnout air damper opening and flue gas oxygen content combination that takes into account boiler efficiency and NOx emissions under different steady-state conditions and record the results as preliminary combustion optimization instructions.

5. The multi-objective combustion optimization method based on economic forecasting control according to claim 4 is characterized in that: The preliminary combustion optimization method in step S2 is: clustering the unit load in the historical operation data using a density-based spatial clustering algorithm, wherein for the unit load data point , by calculating the radius Whether the number of data points in the neighborhood of exceeds the minimum number of points MinPts, the core points are determined and clustered, and then the corresponding data subsets such as coal quality parameters, primary air powder concentration, furnace temperature distribution, flue gas oxygen content, secondary air volume of each layer, burnout air volume, coal feed volume of each layer burner, burner organization, boiler efficiency, NOx emission concentration, main steam temperature and reheat steam temperature under the steady-state operation conditions of each cluster are screened out; The correlation model building module uses a deep neural network DNN with multiple hidden layers to build a correlation model for each data subset of steady-state conditions. The input layer node is the secondary air door opening. , Burnout air damper opening , Flue gas oxygen content , the output layer node is the boiler efficiency and NOx emission concentration , by defining the loss function during training , and according to the back propagation algorithm according to the formula and Update the weights and biases of the neural network; in, and is the boiler efficiency and NOx emission concentration predicted by the neural network, and is the actual boiler efficiency and NOx emission concentration, is the weight coefficient, ranging from 0 to 1, determined by experiment; It is Layer to The weight of the layer, It is Bias of the layer; Optimization problem solving module, building objective function And the secondary air door opening satisfy , Burnout air damper opening satisfy , Flue gas oxygen content satisfy Constraints of in, is the weight coefficient, which is determined according to the actual requirements for boiler efficiency and environmental protection; A genetic algorithm is used to solve the optimization problem by performing selection operations through roulette selection method, crossover operations through single-point crossover, and mutation operations by changing individual gene values ​​with a certain mutation probability, so as to obtain the optimal secondary air damper opening, burnout air damper opening, and flue gas oxygen content combination that takes into account boiler efficiency and NOx emissions under different steady-state operating conditions, and the results are recorded as preliminary combustion optimization instructions.

6. The multi-objective combustion optimization method based on economic forecasting control according to claim 1 is characterized in that: Step S3 further includes an economic forecasting model construction step, which uses a combination of mechanism analysis and data-driven methods to construct a dynamic relationship model between key parameters of the combustion process and boiler efficiency, NOx emissions, and steam temperature, in which partial differential equations are used to describe the physical and chemical changes in the combustion process and actual operating data are combined for parameter identification and model correction.

7. The multi-objective combustion optimization method based on economic forecasting control according to claim 6 is characterized by: Step S3: The steps for constructing the prediction model for unit load and coal quality changes are as follows: It includes a unit load forecasting unit, which is constructed using a long short-term memory network LSTM combined with a Kalman filter. The LSTM unit is based on the formula , , , , , Perform operations; in, They are input gate, forget gate, candidate memory unit, memory unit, output gate and hidden state. and are the corresponding weights and biases, is the sigmoid function, Represents element-wise multiplication; The Kalman filter unit is based on the state equation and Modify the LSTM output to predict the unit load; in, is the state vector, is the state transition matrix, is the control input matrix, is the control input vector, is the process noise vector, is the observation vector, is the observation matrix, is the observation noise vector, which adjusts the prediction value by continuously updating the Kalman gain to improve the prediction accuracy; The coal quality change prediction unit adopts the grey Markov chain model. First, according to the formula Carry out grey prediction, and then use the statistically obtained transfer probability matrix The grey prediction results are corrected by Markov chain to realize the prediction of coal quality change; in, is the predicted value of the sequence generated by one accumulation, is the first value of the original sequence, and are the parameters obtained by solving the grey differential equation by the least square method; It has a dynamic relationship model construction unit for the combustion process, based on the mechanism through the mass conservation equation , energy conservation equation The system of partial differential equations describes the combustion process; Among them, the mass conservation equation is the density of the fuel, It's time. is the velocity vector, is the source term of fuel, indicating the rate at which fuel is generated or consumed; Energy conservation equation is the density of the mixture, is the total energy, It's pressure. is the thermal conductivity, is the temperature, is the heat source term; And use the least squares support vector machine to optimize the problem , Identify partial differential equation parameters, modify models, and construct dynamic relationship models between key parameters of the combustion process and boiler efficiency, NOx emissions, and steam temperature; in, is the weight vector, is the bias, is the error term, is the penalty parameter, is the kernel function that maps the input data to a high-dimensional space. is the output data. By solving the optimization problem, the estimated values ​​of the parameters in the partial differential equation are obtained, thus completing the model correction.

8. The multi-objective combustion optimization method based on economic forecasting control according to claim 1 is characterized in that: in step S4: According to the actual operating status of the current unit and the predicted future load and coal quality change trend, the initial optimization parameter combination is selected from the preliminary combustion optimization results; In each control cycle, the economic prediction control algorithm is used for online optimization calculation; in each control cycle, according to the current operating parameters and predicted changes, the optimization model based on dynamic programming or model predictive control (MPC) is used to calculate the combustion parameter adjustment under the following constraints: NOx emission upper limit constraint, reheat steam temperature upper and lower limit constraint, each control quantity upper and lower limit constraint and rate constraint; Apply the calculated combustion parameter adjustment to the actual combustion control system, and use the feedback correction mechanism to collect actual combustion data for comparison and analysis with the predicted values; According to the deviation, the prediction model is corrected and the model parameters are optimized online to achieve closed-loop dynamic combustion optimization.

9. The multi-objective combustion optimization method based on economic forecasting control according to claim 8 is characterized by: The online secondary optimization steps of the multi-objective combustion optimization in step S4 are: The initial optimization parameter combination selection module uses Gaussian process regression GPR to establish the mapping relationship between the current unit operation status and the preliminary combustion optimization results. Its covariance function uses the radial basis function as follows: ； in, is the signal variance, which controls the smoothness of the function; is the length scale, which determines the decay rate of the correlation in the input space; is the noise variance, is the Kronecker function; Estimating hyperparameters by maximum likelihood estimation (MLE) method\Estimating hyperparameters by maximum likelihood estimation (MLE) method , and according to the predicted mean and variance, the initial optimization parameter combination is selected using a probability-based selection strategy; The online optimization calculation module builds an optimization model based on model predictive control (MPC), and the objective function is ; in, , The weight coefficients for boiler efficiency and NOx emission concentration at different time steps can be adjusted dynamically according to actual needs. For example, as time goes by, more emphasis is placed on the recent optimization effect, and the weight coefficient can be gradually increased; It is the penalty coefficient for the change of the control quantity to prevent the control quantity from changing drastically; For the The combustion parameters are in the time step The amount of change; and are the predicted boiler efficiency and NOx emission concentration at the time step The value of Constraints include: NOx upper limit constraint: ,in is the maximum allowed NOx emission concentration; Reheat steam temperature upper and lower limit constraints: ,in and The time steps The lower and upper limits of the lower reheat steam temperature; Upper and lower limit constraints of each control quantity: ,in and For the The lower and upper limits of the combustion parameters; Rate constraints of each control variable: ,in For the The maximum variation allowed for a combustion parameter per unit time; The above optimization problem is solved by using the dynamic programming algorithm. In the dynamic programming, the state variables are defined as the current unit operation state and the combustion parameter combination, and the decision variables are the adjustment amount of the combustion parameters at the next moment. By constructing the value function , which means from the time step Start in state Take decision The optimal target value after , its recursive formula is: ； in, is the objective function value of the current time step, Based on the system dynamic model and current decision Get the state at the next moment; start the reverse solution from the last time step to obtain the optimal combustion parameter adjustment strategy; Feedback correction: Collect actual combustion data, including actual boiler efficiency , NOx emission concentration , Reheat steam temperature And the actual values ​​of combustion parameters ; Calculate the deviation between the predicted value and the actual value: ； ； ； ； The prediction model and optimization model are corrected online using the deviation; for the prediction model, the new model parameters are obtained using the recursive least squares method; the prediction model is assumed to be a linear model ,in To predict the output, is the input data matrix, is the model parameter vector, and the update formula of RLS is: ； ； ； in, is the Kalman gain vector, is the covariance matrix, and the initial value can be set based on prior knowledge; For the optimization model, the weight coefficients and constraints in the objective function are adjusted according to the deviation to achieve closed-loop dynamic combustion optimization.

10. The multi-objective combustion optimization method based on economic forecasting control according to claim 1, Its characteristics are: in step S5: Continuously monitor and evaluate the operating effect of the combustion optimization system, including boiler efficiency improvement, NOx emission reduction, and main steam temperature and reheat steam temperature control accuracy; Adjust and improve the optimization algorithm and model based on the evaluation results. When the optimization effect of a specific operating condition is not good, analyze the cause and take measures such as supplementing data, improving the model structure or adjusting the constraint range to improve the combustion optimization performance; The algorithm steps include: Operation effect monitoring and evaluation module, using hierarchical analysis method to construct judgment matrix , through the formula Solving for the largest eigenvalue and the eigenvector And normalize to get the weight vector , calculate the scores of each indicator such as boiler efficiency improvement score , NOx emission reduction score , Main steam temperature control accuracy score , Reheat steam temperature control accuracy score , comprehensive evaluation score ,when Trigger adjustment and improvement process when necessary; Adjust and improve the module, use principal component analysis (PCA) to calculate the covariance matrix The eigenvalue of and the eigenvector Perform dimensionality reduction processing. If it is a data problem, use the nearest neighbor interpolation method (in ) Supplementary data, if it is a model structure problem, use adaptive model adjustment algorithm and transfer learning method to improve it, if it is a constraint problem, adjust the constraint range according to the data distribution, such as when the calculated data is near the boundary When adjusting the constraint to (in , ), thereby improving combustion optimization performance.

Images