A power station boiler combustion parameter self-optimizing method and system based on extreme random tree modeling

Abstract

The application discloses a power station boiler combustion parameter self-optimization method based on an extreme random tree modeling, which comprises the following steps: S1, system software and hardware arrangement and communication environment building and function debugging; S2, test data exception elimination, and selection of test boiler modeling parameter related data; S3, establishment of five prediction models based on an extreme random tree model, including NOx emission, CO emission, flue gas temperature, reheat steam temperature and oxygen content deviation; S4, introduction of furnace temperature data to improve the prediction effect of the five prediction models; S5, measurement of the accuracy of the five prediction models, and quantification of the training and test effect; S6, dynamic control optimization based on the boiler combustion characteristic model. The index prediction method based on the extreme random tree is used to construct the prediction models of five indexes, including NOx concentration, CO concentration, air preheater outlet flue gas temperature, reheat steam temperature and oxygen content deviation on the left and right sides of the economizer outlet, and the furnace temperature is introduced to further improve the prediction accuracy of the models.

Description

Technical Field

[0001] The present invention relates to the field of self-optimization of combustion parameters of power station boilers, and in particular to a method and system for self-optimization of combustion parameters of power station boilers based on extreme random tree modeling. Background Art

[0002] In my country, thermal power generation still occupies a core position in the power generation system. Reducing coal consumption, reducing pollutant emissions, and improving the safety and economic efficiency of unit operation have become key issues. Compared to hardware equipment modification, boiler combustion optimization can conveniently improve the economic and safe operation of boilers. Optimizing combustion control in thermal power plants primarily involves real-time optimization and adjustment of operating parameters such as coal feed control, air supply control, furnace temperature, exhaust temperature, and reheat steam temperature. This can maximize the actual operating efficiency of the boiler, achieve economical operation of the boiler, and ultimately save energy and improve power plant efficiency.

[0003] There are many technical studies on boiler combustion optimization at home and abroad. At present, the main method is to conduct combustion adjustment tests at several specific steady-state load points to obtain better combustion setting parameters and set the combustion parameter optimization operating curve. During the combustion control process, the air distribution, coal distribution and other parameters are adjusted according to the set operating curve to achieve combustion optimization adjustment under steady-state conditions.

[0004] First, due to limitations in test time and conditions, combustion adjustment tests can generally only be conducted at a limited number of load and coal type operating points. Furthermore, boiler equipment performance changes over time. Consequently, actual boiler operating conditions often differ significantly from test conditions, causing the original optimization test results to deviate from the optimal value or even become invalid. If operators perform combustion adjustments based on these results, significant operational deviations will result.

[0005] Second, under the current operation adjustment method, operators are affected by subjective factors during operation. When the coal quality, load and other operating conditions change, the boiler combustion adjustment operations performed based on operating experience have certain differences, making it difficult to simultaneously complete multiple-objective combustion optimization adjustments such as reducing NOx emissions and improving boiler efficiency.

[0006] Third, due to the lack of real-time monitoring equipment for coal quality and fly ash carbon content, the economic efficiency of the boiler cannot be calculated in real time, let alone adjusted in real time based on the economic efficiency.

[0007] Fourth, due to the variability of boundary conditions such as coal quality and unit load, the current combustion control method cannot meet the needs of dynamic combustion optimization adjustment. Summary of the Invention

[0008] The object of the present invention is to provide a method and system for self-optimization of power plant boiler combustion parameters based on extreme random tree modeling to solve the problems raised in the above background technology.

[0009] To solve the above technical problems, the present invention provides the following technical solution: a method for self-optimization of power plant boiler combustion parameters based on extreme random tree modeling, comprising the following steps:

[0010] S1. System software and hardware layout, communication environment construction and function debugging;

[0011] S2. Abnormal test data are eliminated and relevant data of test boiler modeling parameters are selected;

[0012] S3. Establish five prediction models for NOx emissions, CO emissions, flue gas temperature, reheat steam temperature, and oxygen deviation based on the extreme random tree model;

[0013] S4. Introducing furnace temperature data to improve the prediction results of the five prediction models;

[0014] S5. Measure the accuracy of the five prediction models and quantify their training and testing effects;

[0015] S6. Dynamic control optimization based on boiler combustion characteristic model.

[0016] Preferably, in step S1, the system software and hardware deployment includes hardware deployment such as servers, displays, communication cards, power supplies, communication cables, CO online detection devices, and the establishment of a software operating environment;

[0017] The communication environment construction mainly includes two parts: the communication between the combustion parameter self-optimization system and the DCS system, and the internal data processing of the DCS system. The data exchange between the combustion parameter self-optimization system and the DCS system is carried out through MODBUS communication, and finally the DCS system transmits the output control quantity signal to the corresponding actuator;

[0018] In step S1, the communication function is tested, including communication point verification, to confirm that the basic two-way communication functions between the self-optimization system and the DCS are complete, the data points are accurate, the stability is good, and there are no missing or incorrect reading / writing of required variable points. At the same time, the optimization instructions can be accurately sent to the DCS, and the heartbeat signal test is completed to ensure that the combustion parameter self-optimization system and the DCS heartbeat signal communicate normally. When a communication failure occurs, the heartbeat stops and the self-optimization system exits safely without affecting the normal operation of the unit.

[0019] Preferably, in step S2, after removing abnormal fluctuations and redundant data from the test data, a total of 4700 groups of operating data are screened out, 4200 groups are taken as training sets, and 500 groups are taken as test sets.

[0020] Preferably, in the step S3, the data related to the modeling parameters are mainly the selected modeling input parameters and model output parameters;

[0021] The modeling input parameters include: a total of 12-dimensional parameters of six coal feeders A / B / C / D / E / F related to the coal feeder; a total of 12-dimensional parameters of the secondary air volume; used to describe the influence of the burner air distribution mode on the NOx emission characteristics; a total of 4-dimensional parameters of the overfire air on both sides of A / B; the rest include the flue gas and feed water temperature at the economizer outlet, load, main steam pressure and temperature, furnace outlet flue gas temperature, total air volume, and total coal volume parameters. The total number of model input parameters is 60 dimensions, and the model output parameters are the NOx emissions, CO emissions, flue gas temperature, reheater steam temperature, and oxygen content deviation on both sides of A / B under the corresponding working conditions.

[0022] Preferably, the five prediction models are consistent in the modeling method. Taking the NOx emission prediction model as an example, the step S3 is as follows:

[0023] S31. Establish initial nodes, prepare a sample set Cx (x = 1...P), where P is the number of base classifiers. For each sample Cx, randomly and without replacement, m (m << M) feature attributes are extracted from M feature attributes to construct P initial decision tree models;

[0024] S32. Obtain the best splitting value under random features. For a single decision tree model, use the data set Cx for independent training. When splitting nodes, each feature attribute in the decision tree model randomly generates a splitting threshold, and then calculates the bifurcation value. According to the numerical characteristics of the model, the minimum mean square error is used as the bifurcation value. After comparing all feature bifurcation values, the optimal value is selected, and the node bifurcation is achieved with the feature corresponding to this optimal value;

[0025] S33. Repeat S22 to further divide the nodes until they cannot be divided any further. At this time, an EP model is obtained. Repeat P times to generate all EPs; S34. Integrate all EP models and use the mean method to obtain the prediction result of the regression model

[0026]

[0027] In the formula, f

[0030] , ,

[0028] , ,

[0029] , is the prediction result of the x-th decision tree.

[0028] Preferably, in the step S4, the furnace temperature data is obtained based on the developed acoustic temperature measurement system, using the relationship between the propagation speed of the acoustic signal in the gas medium and the temperature, and is calculated using the following method:

[0029]

[0030] Where c is the acoustic wave velocity, m / s; γ is the gas specific heat index; R is the gas constant, J / mol·K; M is the gas molar weight, kg / mol; T is the medium temperature, K; L is the acoustic wave propagation distance, m; t is the acoustic wave propagation time, s; and Z is the flue gas correlation coefficient. L can be determined after the acoustic wave transceiver is deployed. Then, the flight time t required for the acoustic wave to traverse the path can be measured in real time. The average linear temperature T on the acoustic wave propagation channel is calculated according to the above formula. Several temperature measurement lines are intersected to construct a linear temperature network. The furnace cross-section temperature can then be restored using the temperature field reconstruction algorithm.

[0031] Preferably, in step S5, the model accuracy is calculated using the root mean square error and the coefficient of determination as evaluation indicators using the following method:

[0032]

[0033] Where: y i is the actual output; is the predicted output; n is the number of samples.

[0034] Preferably, in step S6, the dynamic control optimization includes two parts: preliminary optimization and online closed-loop control. The preliminary optimization involves selecting the optimal operation combination for the operator from the collected historical data with the goal of minimizing NOx. Six reference curves, namely, coal feed rate-secondary air, load-ember air, and load-oxygen constant values, are calibrated as the basis for closed-loop parameter self-optimization. Then, based on the preliminary optimization, the optimization problem solution scope and multi-objective optimization function are constructed according to the requirements of amplitude and speed limits, and the mill downtime problem is taken into account, so that the system only optimizes the coal feed configuration between the operating mills without changing the total coal capacity of the boiler.

[0035] Preferably, the main steps of the preliminary optimization in step S6 are:

[0036] A1: Remove invalid data points caused by measurement and transmission anomalies;

[0037] A2: Select the steady-state operating section from the operating and test data;

[0038] A3: Select appropriate upper and lower limits and step size.

[0039] A4: Divide and classify the steady-state operating section according to the coal feeder coal quantity instruction or the total coal quantity instruction in a certain step length;

[0040] A5: Select several control quantity combinations with the highest boiler efficiency in each coal quantity instruction range;

[0041] A6: Use the average values ​​of several combined control variables as data points to connect and obtain the optimized guidance curve to obtain the preliminary optimization results.

[0042] The main steps of the online closed-loop control in step S6 are:

[0043] T1: Objective function:

[0044]

[0045] sMV min <MV<MV max (8)

[0046] ΔMV min <ΔMV<ΔMV max (9)

[0047] NO x (k+j|k)≤NO xmax ,j=1,2,…,N (10)

[0048] T rhmin ≤T rh (k+j|k)≤T rhmax ,j=1,2,…,N (11)

[0049] Where, η(k+j|k) is the predicted value of boiler efficiency at N moments in the future, obtained by the established nonlinear model of boiler efficiency. N is the number of prediction steps, which should be large enough to cover the entire period from the change of the control variable to the end of the transition process of the output variable. MV represents the control variables such as oxygen constant, burnout air door opening, secondary air door opening, and coal demand bias. Equations (8) and (9) are the upper and lower limit constraints and change rate constraints for each control variable. Equation (10) is the constraint on NOx emissions, NO x (k+j|k) is the predicted value of NOx at N moments in the future, obtained by the established NOx nonlinear model. xmax is the upper limit of NOx emissions, which can be set and modified by the operator according to the unit load; Equation (11) is the constraint on the reheat steam temperature, T rh (k+j|k) is the predicted value of the reheat steam temperature at the next N moments, which is obtained by the established nonlinear model of reheat steam temperature. rhmax is the upper limit of reheat steam temperature, T rhmin It is the lower limit of reheat steam temperature, which can be set and modified by the operator according to the unit load;

[0050] T2: Instructions for setting constraints:

[0051] For the O2 constant and each secondary air damper, the original DCS control logic already had an open-loop control curve configured for random group load changes, provided by the boiler plant or commissioning unit. Considering its rationality and safety, when setting the O2 constant and the upper and lower limits of each secondary air damper, this curve was used as a benchmark, with fluctuations of 1% and 10% as the optimization interval. The size of this interval was adjusted based on actual conditions. Furthermore, the impact of mill start-up and shutdown on secondary air damper control must be considered. When a mill is shut down, the corresponding secondary air damper must be closed to 5% open.

[0052] T3: Solving the optimization problem of Tyrannosaurus Rex optimization algorithm NSTROA based on non-dominated sorting:

[0053] T4: Handling grinding stoppages based on multiple intelligent optimization algorithms and equality constraints:

[0054] The control strategy for combustion optimization is based on a nonlinear dynamic model that transforms each manipulated variable into each optimized variable. It uses a multi-objective predictive control method to provide the best combination of manipulated variables while satisfying production constraints. The optimization bias is calculated based on the optimized combustion parameters and sent to the DCS for closed-loop optimization control, ultimately achieving dynamic optimization of boiler combustion. The problem can be expressed as:

[0055]

[0056] In formula (12), inequality constraint ① is obtained by transforming the control quantity increment constraint, and constraint ② is the upper and lower limit constraints of the control quantity.

[0057] The present invention also discloses a power plant boiler combustion parameter self-optimization system based on extreme random tree modeling, including hardware and software facilities, the hardware including a server, a display, a communication card, a power supply, a communication cable, and a CO online detection device, and the hardware is equipped with software for executing the above-mentioned power plant boiler combustion parameter self-optimization method based on extreme random tree modeling.

[0058] Compared with the prior art, the present invention has the following beneficial effects:

[0059] The present invention provides a method and system for self-optimization of power plant boiler combustion parameters based on extreme randomized tree modeling. The present invention proposes an indicator prediction method based on extreme randomized trees (ET) to construct a prediction model for five indicators: NOx concentration, CO concentration, flue gas temperature at the air preheater outlet, reheat steam temperature, and oxygen deviation on the left and right sides of the economizer outlet. The furnace temperature, an important information affecting the indicator, is introduced to further improve the prediction accuracy of the model. Based on the model prediction data and the unit historical data, an indicator reflecting the economic characteristics of boiler combustion is constructed. The self-optimization of boiler combustion adjustment parameters is achieved by combining multiple intelligent optimization algorithms and prediction strategies. Under the conditions of meeting the constraints, the coal feed rate of each coal feeder, the opening of the secondary wind box damper, the opening of the burnout damper, and the oxygen content of the flue gas are automatically adjusted. For complex disturbance conditions with variable load, variable coal quality, and variable optimization targets, the boiler combustion control has excellent perception, self-adaptation, and self-optimization capabilities. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] Figure 1 This is a schematic diagram of the overall conceptual framework of the present invention;

[0061] Figure 2 Schematic diagram of the dynamic control optimization steps of the present invention;

[0062] Figure 3 It is the DCS communication point number allocation table;

[0063] Figure 4 Detailed list of some DCS communication point numbers. DETAILED DESCRIPTION

[0064] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0065] This embodiment provides a prediction model and parameter self-optimization method and system based on extreme random trees, such as Figure 1 As shown, the following steps are included:

[0066] S1. System software and hardware layout, communication environment construction and function debugging;

[0067] S2. Abnormal test data are eliminated and relevant data of test boiler modeling parameters are selected;

[0068] S3. Establish five prediction models for NOx emissions, CO emissions, flue gas temperature, reheat steam temperature, and oxygen deviation based on the extreme random tree model;

[0069] S4. Introducing furnace temperature data to further improve the prediction results of the five models;

[0070] S5. Measure model accuracy and quantify its training and testing effects;

[0071] S6. Dynamic control optimization based on boiler combustion characteristic model.

[0072] Furthermore, in step S1, the system software and hardware deployment includes hardware deployment such as servers, displays, communication cards, power supplies, communication cables, and CO online detection devices, as well as the establishment of a software operating environment.

[0073] In step S1, the communication environment construction mainly includes two parts: communication between the combustion parameter self-optimization system and the DCS system, and internal data processing of the DCS system.

[0074] The data exchange between the combustion parameter self-optimization system and the DCS system is carried out through MODBUS communication, and the DCS system finally transmits the output control quantity signal to the corresponding actuator.

[0075] The combustion parameter self-optimization software realizes data communication and exchange with DCS based on the configuration point table. The configuration of the point table is as follows: Figure 3 As shown, some specific variable regulations and descriptions are as follows Figure 4 shown.

[0076] In step S1, the communication function is tested, including communication point verification, to confirm that the basic two-way communication functions between the self-optimization system and the DCS are complete, the data points are accurate, the stability is good, and there are no missing or incorrect reading / writing of required variable points. At the same time, the optimization instructions can be accurately sent to the DCS, and the heartbeat signal test is completed to ensure that the combustion parameter self-optimization system and the DCS heartbeat signal communicate normally. When a communication failure occurs, the heartbeat stops and the self-optimization system exits safely without affecting the normal operation of the unit.

[0077] In step S2, after removing abnormal fluctuations and redundant data, the test data is screened out to obtain 4700 groups of operating data, 4200 groups of which are used as training sets and 500 groups as test sets.

[0078] In the step S3, the data related to the modeling parameters mainly include the selected modeling input parameters and model output parameters. The modeling input parameters include: a total of 12-dimensional parameters for six coal feeders A / B / C / D / E / F related to the coal feeder; a total of 12-dimensional parameters for the secondary air volume; used to describe the influence of the burner air distribution mode on the NOx emission characteristics; a total of 4-dimensional parameters for the over-fire air on both sides of A / B; the remaining parameters include the flue gas and feed water temperature at the economizer outlet, load, main steam pressure and temperature, flue gas temperature at the furnace outlet, total air volume, total coal volume, etc. The model input parameters total 60 dimensions. The model output parameters are the NOx emissions, CO emissions, flue gas temperature, reheated steam temperature, and oxygen content deviation on the A / B sides corresponding to the working conditions.

[0079] In the step S3, the Extra Trees (ET) is a tree-based perturbation and combination algorithm, which can be used for nonlinear system modeling and regression prediction. Each tree in the Extra Trees is independently trained with all the original data, improving the utilization rate of the training samples. And when splitting nodes, the bifurcation values are randomly selected without following the criterion of selecting the best splitting threshold or feature, increasing the difference and randomness between decision trees. Therefore, the Extra Trees not only has the ability of the random forest to handle multi-dimensional data sets and automatically select features, but also reduces the sensitivity of the algorithm to noise and enhances the generalization ability.

[0080] Further, the five prediction models are consistent in the modeling method. Taking the NOx emission prediction model as an example, the step S3 is as follows:

[0081] S31. Establish an initial node, prepare a sample set Cx (x = ......P), where P is the number of base classifiers. For each sample Cx, randomly and without replacement, extract m (m << M) feature attributes from M feature attributes to construct P initial decision tree models;

[0082] S32. Obtain the best splitting value under random features. For a single decision tree model, use the data set Cx for independent training. When splitting nodes, a splitting threshold is randomly generated for each feature attribute in the decision tree model, and then the bifurcation value is calculated. Regarding the numerical characteristics of the model, the minimum mean square error is used as the bifurcation value. After comparing all the feature bifurcation values, the optimal value is selected, and the node is bifurcated with the feature corresponding to this optimal value;

[0083] S33. Repeat S22 to further divide the nodes until they cannot be divided any further. At this time, an EP model is obtained. Repeat P times to generate all EPs;

[0084] S34. Integrate all EP models and use the mean method to obtain the prediction result of the regression model

[0085] Where, f x is the prediction result of the x-th decision tree.

[0086] Furthermore, in step S4, the furnace temperature data is obtained based on the developed acoustic temperature measurement system, using the relationship between the propagation speed of the acoustic wave signal in the gas medium and the temperature, and is calculated using the following method:

[0087]

[0088] Where c is the speed of sound waves, m / s; γ is the specific heat index of gas; R is the gas constant, J / mol·K; M is the molar weight of gas, kg / mol; T is the medium temperature, K; L is the sound wave propagation distance, m; t is the sound wave propagation time, s; and Z is the flue gas correlation coefficient.

[0089] Specifically, after the acoustic wave transceiver is arranged, L can be determined, and then the flight time t required for the acoustic wave to traverse the path can be measured in real time. The average linear temperature T on the acoustic wave propagation channel can be calculated according to the above formula. Several temperature measurement lines are intersected to construct a linear temperature network. Then, the furnace cross-section temperature can be restored according to the temperature field reconstruction algorithm. The measurement point arrangement of the acoustic wave temperature measurement system and the furnace cross-section temperature division matrix are shown as follows: Figure 2 shown.

[0090] Furthermore, in step S5, the model accuracy is calculated using the root mean square error (RMSE) and the coefficient of determination (R2) as evaluation indicators using the following method:

[0091]

[0092] Where: y i is the actual output; is the predicted output; n is the number of samples.

[0093] Furthermore, if Figure 2 As shown, in step S6, the dynamic control optimization includes two parts: preliminary optimization and online closed-loop control. The preliminary optimization is to screen out the optimal operation combination of the operator from the collected historical data with the goal of minimizing NOx, calibrate the coal feed-secondary air (3 groups of upper, middle and lower), load-burning air (2 groups), and load-oxygen constant value (1 group), a total of 6 groups of reference curves, as the optimization basis for closed-loop parameter self-optimization. Then, based on the preliminary optimization, the optimization problem solution scope and multi-objective optimization function are constructed according to the requirements of amplitude limit and speed limit, and the mill stop problem is taken into account, so that the system only optimizes the coal feed configuration between the operating mills without changing the total coal capacity of the boiler.

[0094] Specifically, the main steps of the preliminary optimization in step S6 are:

[0095] A1: Remove invalid data points caused by measurement and transmission anomalies;

[0096] A2: Select the steady-state operating section from the operating and test data;

[0097] A3: Select appropriate upper and lower limits and step size.

[0098] A4: Divide and classify the steady-state operating section according to the coal feeder coal quantity instruction or the total coal quantity instruction in a certain step length;

[0099] A5: Select several control quantity combinations with the highest boiler efficiency in each coal quantity instruction range;

[0100] A6: Use the average values ​​of the control quantities of several combinations as data points to connect and run the guidance curve after optimization to obtain the preliminary optimization results.

[0101] Specifically, the main steps of the online closed-loop control in step S6 are:

[0102] T1: Objective function:

[0103]

[0104] sMV min <MV<MV max (8)

[0105] ΔMV min <ΔMV<ΔMV max (9)

[0106] NO x (k+j|k)≤NO xmax ,j=1,2,…,N (10)

[0107] T rhmin ≤T rh (k+j|k)≤T rhmax ,j=1,2,…,N (11)

[0108] Where, η(k+j|k) is the predicted value of boiler efficiency at N moments in the future, obtained by the established nonlinear model of boiler efficiency. N is the number of prediction steps, which should be large enough to cover the entire period from the change of the control variable to the end of the transition process of the output variable. MV represents the control variables such as oxygen constant, burnout air door opening, secondary air door opening, and coal demand bias. Equations (8) and (9) are the upper and lower limit constraints and change rate constraints for each control variable. Equation (10) is the constraint on NOx emissions, NO x (k+j|k) is the predicted value of NOx at N moments in the future, obtained by the established NOx nonlinear model. xmaxis the upper limit of NOx emissions, which can be set and modified by the operator according to the unit load; Equation (11) is the constraint on the reheat steam temperature, T rh (k+j|k) is the predicted value of the reheat steam temperature at the next N moments, which is obtained by the established nonlinear model of reheat steam temperature. rhmax is the upper limit of reheat steam temperature, T rhmin It is the lower limit of the reheat steam temperature and can be set and modified by the operating personnel according to the unit load.

[0109] The biggest difference between the objective function (7) and the dynamic performance index of deviation used in general parameter self-optimization is that it is an economic index without a fixed target value. The goal of solving the optimization problem of equations (7) to (11) is to solve a set of operating variables so that the average boiler efficiency in the future N-step optimization time domain is the highest while satisfying the NOx emission limit set by the operator and various control quantity constraints. In order to avoid the problem of large variance and small correlation between models when directly calculating the N-step prediction value, the present invention uses a hybrid prediction strategy of direct prediction + recursive prediction. N models are constructed according to N time steps, and the prediction value calculated by each prediction model is incorporated into the next model for training. At the same time, the prediction time domain rolling optimization method is adopted to quickly eliminate the influence of modeling errors and disturbances on the optimization results. Based on the hybrid prediction and time domain rolling optimization strategy, the boiler efficiency and emission optimization is carried out over a period of time in the future. On the one hand, it is beneficial to overcome the influence of measurement signal delays such as NOx, fly ash carbon content, and O2, and organize combustion in advance; on the other hand, it is beneficial to overcome the influence of measurement noise and make the operation quantity action more stable. It should be pointed out that the optimization result of this method does not depend on the absolute accuracy of the boiler efficiency calculation, but only requires the trend to be correct, thus reducing its calculation difficulty.

[0110] T2: Instructions for setting constraints:

[0111] For the O2 constant and each secondary air damper, the original DCS control logic already had an open-loop control curve configured for random group load fluctuations, provided by the boiler plant or commissioning unit. To ensure rationality and safety, the upper and lower limits for the O2 constant and each secondary air damper were set using this curve as a benchmark, with fluctuations of 1% and 10% between them as the optimal range. This range can be adjusted based on actual conditions. Furthermore, the impact of mill startup and shutdown on secondary air damper control must be considered. When a mill is shut down, the corresponding secondary air damper must be closed to 5%.

[0112] T3: Solving the optimization problem of Tyrannosaurus Rex optimization algorithm NSTROA based on non-dominated sorting:

[0113] The optimization problem in Equations (7)-(11) is a complex constrained nonlinear optimization problem, which is solved using the Non-Dominated Tyrannosaurus Optimization (NSTROA) algorithm. The NSTROA algorithm combines the Tyrannosaurus Optimization (TROA) algorithm with a non-dominated sorting strategy, offering advantages such as fast search speed and can quickly find optimal solutions to multiple objective functions.

[0114] T4: Handling grinding stoppages based on multiple intelligent optimization algorithms and equality constraints:

[0115] The combustion optimization control strategy is based on a nonlinear dynamic model that transforms various operating variables (oxygen content, coal bias, OFA damper opening, secondary damper opening, etc.) to various optimization variables (boiler efficiency, SCR inlet NOx concentration, reheat steam temperature, etc.). It uses a multi-objective predictive control method to provide the best combination of operating variables while meeting production constraints. The optimization bias is calculated based on the optimized combustion parameters and sent to the DCS for closed-loop optimization control, ultimately achieving dynamic optimization of boiler combustion. The problem can be expressed as:

[0116]

[0117] In formula (12), inequality constraint ① is obtained by transforming the control quantity increment constraint, and constraint ② is the upper and lower limit constraints of the control quantity.

[0118] The actual operation shows that the number of coal mills in operation changes with the load. According to different load conditions, the present invention adopts a genetic optimization algorithm and a mathematical dynamic programming algorithm to solve the problem.

[0119] When the load is high, at least one coal mill is operating on each floor. At this point, the operating variables have unequal upper and lower bounds (lb < ub), necessitating the search for an optimal solution set that maximizes the performance of each objective function. A multi-objective genetic algorithm (MGA) is an evolutionary algorithm used to analyze and solve multi-objective optimization problems. Its core goal is to coordinate the relationships between the various objective functions. Therefore, the parameter self-optimizer uses the Elitist Non-Dominated Sorting Genetic Algorithm (NSGA-II) with an elitist strategy to solve the problem. Among the numerous genetic algorithms for multi-objective optimization, the NSGA-II algorithm has become a fundamental algorithm for multi-objective optimization due to its simplicity, effectiveness, and significant advantages. This algorithm proposes a fast non-dominated sorting algorithm and introduces an elitist strategy to expand the sampling space, thereby reducing computational complexity and improving the accuracy of optimization results.

[0120] Under low load conditions, when all coal mills on a certain layer are stopped, the secondary air door opening of that layer should be closed to a certain minimum opening u minopen Therefore, by judging the operation status of each layer of coal mills in advance, adding equality constraints to the secondary air door opening and coal quantity bias of the non-milling layer, forcing the lb secair =ub secair =u minopen , lb millbias =ub millbias = 0 and canceling the corresponding incremental constraints prompts the parameter self-optimizer to choose a solution based on dynamic programming theory. The Bellman equation (BellmanDP) is selected to solve the dynamic programming optimal control problem. This method can obtain global and local optimal solutions, reflecting the evolution of the dynamic process. Incorporating practical knowledge and experience can improve solution efficiency. The Bellman equation expresses the optimal decision of the optimal strategy under a given current state and decomposes the problem into optimal decisions for subproblems. By calculating each manipulated variable and solving for the remaining manipulated variables, the remaining variables are reduced in computational effort, improved computational efficiency, and enhanced algorithmic flexibility.

[0121] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A method for self-optimization of power plant boiler combustion parameters based on extreme random tree modeling, characterized by: The steps include: S1. System software and hardware layout, communication environment construction and function debugging; S2. Abnormal test data are eliminated and relevant data of test boiler modeling parameters are selected; S3. Establish five prediction models for NOx emissions, CO emissions, flue gas temperature, reheat steam temperature, and oxygen deviation based on the extreme random tree model; S4. Introducing furnace temperature data to improve the prediction results of the five prediction models; S5. Measure the accuracy of the five prediction models and quantify their training and testing effects; S6. Dynamic control optimization based on boiler combustion characteristic model; In step S6, the dynamic control optimization includes two parts: preliminary optimization and online closed-loop control. The preliminary optimization is to screen out the optimal operation combination of the operator from the collected historical data with the goal of minimizing NOx, calibrate six groups of reference curves for coal feed rate-secondary air, load-ember air, and load-oxygen constant value, and use them as the optimization basis for closed-loop parameter self-optimization. Then, based on the preliminary optimization, the optimization problem solution scope and multi-objective optimization function are constructed according to the requirements of amplitude limit and speed limit, and the mill stop problem is taken into account, so that the system only optimizes the coal feed configuration between the operating mills without changing the total coal capacity of the boiler; The main steps of the preliminary optimization in step S6 are: A1: Remove invalid data points caused by measurement and transmission anomalies; A2: Select the steady-state operating section from the operating and test data; A3: Select appropriate upper and lower limits and step size. A4: Divide and classify the steady-state operating section according to the coal feeder coal quantity instruction or the total coal quantity instruction in a certain step length; A5: Select several control quantity combinations with the highest boiler efficiency in each coal quantity instruction range; A6: Use the average values ​​of several combined control variables as data points to connect and obtain the optimized guidance curve to obtain the preliminary optimization results. The main steps of the online closed-loop control in step S6 are: T1: Objective function: s.t.MV min <MV<MV max (8) ΔMV min <ΔMV<ΔMV max (9) NO x (k+j|k)≤NO xmax ,j=1,2,…,N (10) T rhmin ≤T rh (k+j|k)≤T rhmax ,j=1,2,…,N (11) Where, η(k+j|k) is the predicted value of boiler efficiency at N moments in the future, obtained by the established nonlinear model of boiler efficiency. N is the number of prediction steps, which should be large enough to cover the entire period from the change of the control variable to the end of the transition process of the output variable. MV represents the control variables such as oxygen constant, burnout air door opening, secondary air door opening, and coal demand bias. Equations (8) and (9) are the upper and lower limit constraints and change rate constraints for each control variable. Equation (10) is the constraint on NOx emissions, NO x (k+j|k) is the predicted value of NOx at N moments in the future, obtained by the established NOx nonlinear model. xmax is the upper limit of NOx emissions, which can be set and modified by the operator according to the unit load; Equation (11) is the constraint on the reheat steam temperature, T rh (k+j|k) is the predicted value of the reheat steam temperature at the next N moments, which is obtained by the established nonlinear model of reheat steam temperature. rhmax is the upper limit of reheat steam temperature, T rhmin It is the lower limit of reheat steam temperature, which can be set and modified by the operator according to the unit load; T2: Instructions for setting constraints: For the O2 constant and each secondary air damper, the original DCS control logic already had an open-loop control curve configured for random group load changes, provided by the boiler plant or commissioning unit. Considering its rationality and safety, when setting the O2 constant and the upper and lower limits of each secondary air damper, this curve was used as a benchmark, with fluctuations of 1% and 10% as the optimization interval. The size of this interval was adjusted based on actual conditions. Furthermore, the impact of mill start-up and shutdown on secondary air damper control must be considered. When a mill is shut down, the corresponding secondary air damper must be closed to 5% open. T3: Solving optimization problems based on the non-dominated sorting Tyrannosaurus Rex optimization algorithm NSTROA; T4: Handling grinding stoppage situations based on multiple intelligent optimization algorithms and equality constraints; The control strategy for combustion optimization is based on a nonlinear dynamic model that transforms each manipulated variable into each optimized variable. It uses a multi-objective predictive control method to provide the best combination of manipulated variables while satisfying production constraints. The optimization bias is calculated based on the optimized combustion parameters and sent to the DCS for closed-loop optimization control, ultimately achieving dynamic optimization of boiler combustion. The problem can be expressed as: In formula (12), inequality constraint ① is obtained by transforming the control quantity increment constraint, and constraint ② is the upper and lower limit constraints of the control quantity.

2. The method for self-optimization of power plant boiler combustion parameters based on extreme random tree modeling according to claim 1, characterized in that: In the step S1, the system software and hardware are deployed, including the deployment of hardware such as servers, monitors, communication card components, power supplies, communication cables, CO online detection devices, etc., and the construction of the software operating environment. The construction of the communication environment mainly includes two parts: the communication between the combustion parameter self-optimization system and the DCS system, and the internal data processing of the DCS system. The data interaction between the combustion parameter self-optimization system and the DCS system is carried out through MODBUS communication, and finally the DCS system transmits the output control signal to the corresponding actuator. In the step S1, the communication function is tested, including the verification of communication points, confirming that the basic functions of the bidirectional communication between the self-optimization system and the DCS are complete, the data points are accurate, the stability is good, there are no missing / wrong read / write required variable points, and at the same time, the optimization instructions can be accurately sent to the DCS, complete the heartbeat signal test, ensure the normal communication of the heartbeat signal between the combustion parameter self-optimization system and the DCS, and when a communication failure occurs, the heartbeat stops, and the self-optimization system safely exits without affecting the normal operation of the unit.

3. The method for self-optimization of power plant boiler combustion parameters based on extreme random tree modeling according to claim 2, characterized in that: In the step S2, after eliminating abnormal fluctuations and redundant data from the test data, 4700 groups of operation data are selected, 4200 groups are taken as the training set, and 500 groups are taken as the test set.

4. The method for self-optimization of power plant boiler combustion parameters based on extreme random tree modeling according to claim 3 is characterized by: In the step S3, the data related to the modeling parameters are mainly the selected modeling input parameters and model output parameters. The modeling input parameters include: a total of 12-dimensional parameters of six coal feeders A / B / C / D / E / F related to the coal feeder. The secondary air volume has a total of 12-dimensional parameters, which are used to describe the influence of the burner air distribution mode on the NOx emission characteristics; the over-fire air on both sides A / B has a total of 4-dimensional parameters; the rest include the flue gas and feed water temperature at the economizer outlet, load, main steam pressure and temperature, furnace outlet flue gas temperature, total air volume, total coal volume parameters. The model input parameters total 60 dimensions, and the model output parameters are the NOx emissions, CO emissions, flue gas temperature, reheater steam temperature, and oxygen content deviation on both sides A / B under the corresponding working conditions.

5. The method for self-optimization of power plant boiler combustion parameters based on extreme random tree modeling according to claim 4 is characterized in that: The five prediction models are consistent in the modeling method. Taking the NOx emission prediction model as an example, the step S3 is as follows: S31. Establish the initial nodes, prepare the sample set Cx (x = 1...P), where P is the number of base classifiers. For each sample Cx, randomly and without replacement, m (m << M) feature attributes are extracted from M feature attributes to construct P initial decision tree models. S32. Obtain the best splitting value under random features. For a single decision tree model, use the data set Cx for independent training. When splitting the nodes, each feature attribute in the decision tree model randomly generates a splitting threshold, and then calculates the bifurcation value. For the numerical characteristics of the model, the minimum mean square error is used as the bifurcation value. After comparing all the feature bifurcation values, the optimal value is selected, and the node bifurcation is realized with the feature corresponding to this optimal value. S33. Repeat the execution of S22 to further divide the nodes until they cannot be divided any more. At this time, an EP model is obtained. Repeat P times to generate all EPs. S34. Integrate all EP models and use the mean method to obtain the prediction results of the regression model Where, f x is the prediction result of the x-th decision tree.

6. The method for self-optimization of power plant boiler combustion parameters based on extreme random tree modeling according to claim 5, characterized in that: In step S4, the furnace temperature data is obtained based on the developed acoustic temperature measurement system, using the relationship between the propagation speed of the acoustic wave signal in the gas medium and the temperature, and is calculated using the following method: Where c is the acoustic wave velocity, m / s; γ is the gas specific heat index; R is the gas constant, J / mol·K; M is the gas molar weight, kg / mol; T is the medium temperature, K; L is the acoustic wave propagation distance, m; t is the acoustic wave propagation time, s; and Z is the flue gas correlation coefficient. L can be determined after the acoustic wave transceiver is deployed. Then, the flight time t required for the acoustic wave to traverse the path can be measured in real time. The average linear temperature T on the acoustic wave propagation channel is calculated according to the above formula. Several temperature measurement lines are intersected to construct a linear temperature network. The furnace cross-section temperature can then be restored using the temperature field reconstruction algorithm.

7. The method for self-optimization of power plant boiler combustion parameters based on extreme random tree modeling according to claim 6, characterized in that: In step S5, the model accuracy is calculated using the root mean square error and the coefficient of determination as evaluation indicators using the following method: Where: y i is the actual output; is the predicted output; n is the number of samples.

8. A power plant boiler combustion parameter self-optimization system based on extreme random tree modeling, characterized by: The method comprises hardware and software facilities, wherein the hardware comprises a server, a display, a communication card, a power supply, a communication cable, and a CO online detection device. The hardware is equipped with software for executing a method for self-optimization of combustion parameters of a power plant boiler based on extreme random tree modeling as described in any one of claims 1 to 7.

Images