A coal quality full-characteristics-based multi-target optimization and dynamic coal feeding control method

Abstract

The application discloses a coal quality full characteristic-based multi-target optimization and dynamic coal feeding control method, which comprises the following steps: S1. coal quality full characteristic representation and model construction: based on a coal quality database, a coking tendency prediction model and a burnout and pollutant generation prediction model are established; S2. quantitative definition of coal feeding operation parameters: quantitative indexes used for describing dynamic behaviors of a coal feeding system are defined; S3. multi-physical field mapping relationship establishment: quantitative mapping relationships among coal quality full characteristic parameters, quantitative indexes, boiler combustion efficiency, pollutant emission concentration and coking risk are established; S4. multi-target optimization and strategy solving: taking economy, environmental protection and safety as optimization targets, a multi-target optimization function is constructed, a multi-target optimization algorithm is adopted to solve a Pareto optimal solution set, a final solution is selected from the Pareto optimal solution set, and a dynamic control strategy suitable for quantitative indexes of the current coal quality is generated. The application can realize the collaborative optimization of three targets of safety, environmental protection and economy of a coal-fired boiler.

Description

Technical Field

[0001] This invention relates to the field of coal-fired boiler operation optimization technology, specifically to a multi-objective optimization and dynamic pulverized coal feeding control method based on the full characteristics of coal quality. Background Technology

[0002] In the coal-fired power generation industry, the stable, efficient, and clean operation of boilers is crucial. However, the multi-dimensional and dynamic fluctuations in the quality of coal fed into the boiler (such as changes in parameters like industrial analysis, elemental analysis, and ash properties) significantly impact boiler combustion efficiency and pollutant emissions (such as NOx). x SO x Emission control and equipment safety (such as coking and overheating) pose ongoing challenges.

[0003] Currently, the mainstream coal feeding control strategies in the industry are mostly based on PID regulation based on feedback of a single operating parameter (such as unit load), which has the following obvious defects: (1) Lack of correlation: The control logic fails to systematically establish a quantitative correlation model between coal quality characteristics and combustion response. When coal quality changes abruptly, the control command is seriously delayed, leading to combustion deterioration. (2) Single dimension: The optimization target is often limited to single economic indicators such as "combustion efficiency" or "power supply coal consumption", failing to coordinate the optimization of multiple objectives such as environmental protection (pollutant emissions) and safety (coking / overheating risk), often resulting in a situation of "paying attention to one thing but losing another". (3) Insufficient dynamic adaptability: There is a lack of self-updating mechanism for coal quality database and prediction model. In long-term operation, the "coal quality drift" and "model mismatch" problems caused by changes in coal type and equipment aging will gradually reduce the control accuracy.

[0004] Furthermore, the final operating performance of a boiler is highly dependent on the matching degree between the characteristics of the coal fed into the furnace and the pulverized coal feeding strategy. If the fuel quality is not clearly understood, any subsequent optimization and adjustments will lack a reliable basis, seriously affecting the overall effectiveness of coking prevention, pollutant control, and economic improvement. Therefore, there is an urgent need for an intelligent pulverized coal feeding control method that can deeply integrate information on the full range of coal characteristics and achieve coordinated dynamic optimization of multiple objectives, including safety, environmental protection, and economy. Summary of the Invention

[0005] The technical problem to be solved by this invention is to provide a multi-objective optimization and dynamic pulverized coal feeding control method based on the full characteristics of coal quality, so as to solve the problems of boiler coking, overheating, excessive pollutant emissions and reduced operating economy caused by coal quality fluctuations and improper pulverized coal feeding operation, and achieve the synergistic optimization of the three major objectives of safety, environmental protection and economy of coal-fired boilers.

[0006] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows.

[0007] A multi-objective optimization and dynamic coal feeding control method based on the full characteristics of coal quality includes the following steps: S1. Characterization and Model Building of Coal Quality: Construct a multi-dimensional coal quality database containing all coal quality characteristic parameters; based on the coal quality database, establish a coking tendency prediction model and a burnout and pollutant generation prediction model respectively; S2. Quantitative Definition of Powder Feeding Operation Parameters: Define quantitative indicators used to characterize the dynamic behavior of the powder feeding system, including the powder feeding variation amplitude Δ. M Frequency of powder feeding variation f and powder stability index s ; S3. Establishment of multiphysics mapping relationship: Based on the model constructed in step S1, historical operation data mining, numerical simulation and process simulation, establish a quantitative mapping relationship between the full characteristic parameters of coal quality, the quantitative indicators and boiler combustion efficiency, pollutant emission concentration and coking risk; S4. Multi-objective optimization and strategy solution: With economy, environmental protection, and safety as optimization objectives, a multi-objective optimization function is constructed. Under equipment operation constraints, a multi-objective optimization algorithm is used to solve for the Pareto optimal solution set. The final solution is selected from the Pareto optimal solution set to generate the optimal feed rate variation Δ adapted to the current coal quality. M Frequency of powder feeding variation f and powder stability index s Dynamic control strategy.

[0008] Preferably, in step S1, the coal quality full characteristic parameters include industrial analysis parameters, elemental analysis parameters, ash characteristic parameters, and combustion characteristic parameters of the coal type; the industrial analysis parameters include as-received moisture content, as-received ash content, dry ash-free volatile matter, fixed carbon, and as-received lower heating value; the elemental analysis parameters include carbon, hydrogen, oxygen, nitrogen, and total sulfur; the ash characteristic parameters include silicon dioxide, aluminum oxide, iron oxide, calcium oxide, and magnesium / potassium / sodium oxide; and the combustion characteristic parameters include ignition temperature and burnout characteristic index.

[0009] Preferably, in step S1, the specific method for establishing the coking tendency prediction model includes: S11. Feature Engineering Processing: Based on the computer-derived characteristics of coal ash composition, including the silicon-to-aluminum ratio. SR Alkali-acid ratio B / A Slagging index R s and pollution index R f At least one of them; S12. Dataset Construction: Using coal quality full characteristic parameters, mechanism-derived features, and real-time operating parameters as input vectors, and coking risk level as output label, construct a training dataset; S13. Random Forest Model Training and Probability Output: Multiple decision trees are generated using the Bootstrap sampling strategy. Random feature subspace selection is used when splitting decision tree nodes. Feature selection is performed based on the Gini impurity minimization criterion to train a random forest model. The coal quality full characteristic parameters under the current working condition, the mechanism-derived features, and real-time operating parameters are used as input vectors and input into the random forest model to output the predicted probability of the random forest model. S14. Mechanism Constraint Fusion: The probability predicted by the random forest model is weighted and fused with the probability of mechanism constraints based on physical rules through the Bayesian posterior correction formula to output the final coking risk probability.

[0010] Preferably, in step S11, the calculation formula for the mechanism-derived features is as follows: Silicon-to-aluminum ratio SR : Alkali-acid ratio B / A : Slagging Index R s : in, Total sulfur content on a dry basis; Pollution Index R f : In step S12, the coking risk level label is obtained using a semi-supervised labeling method, which combines indirect indicators such as sootblower operation frequency, abnormal increase in desuperheating water flow, and furnace heat absorption attenuation rate in historical operating data with manual records to label the level. In step S13, the Bootstrap strategy used in random forest training generates the training set for a single decision tree through random sampling with replacement, uses out-of-bag data for unbiased model validation, and sets the feature subspace size to the square root of the total number of features. The feature selection based on the Gini impurity minimization criterion specifically involves: In a randomly selected feature subspace, all possible thresholds for each feature are traversed, and the weighted Gini impurity corresponding to each candidate split is calculated. The feature and threshold combination that minimizes the weighted Gini impurity is selected as the optimal splitting scheme; where nodes... The formula for calculating the impurity of Gini is: in, For the node belonging to the first j The probability of coking-like risk; This represents the total number of coking risk categories; In step S14, the Bayesian posterior correction formula is: in, This represents the probability of final coking risk. Predict probabilities for the random forest model; Mechanism-constrained probability; The confidence weighting factor is based on the similarity between the current working conditions and the distribution of historical data. The confidence weight factor Dynamically adjust based on the Mahalanobis distance between the current operating point and the training set center: When the operating condition is in a region with high density of historical data Approaching 1 The results are primarily based on random forest predictions. When the operating condition is in a region with sparse historical data or an extrapolated region Approaching 0 The prediction results are primarily determined by mechanistic constraints.

[0011] Preferably, in step S1, the specific method for establishing the burnout and pollutant generation prediction model includes: Construct a hybrid prediction model architecture that targets NO x Generation amount, fly ash carbon content and SO x The generated data all adopt the "mechanism skeleton + data correction" model. The mechanism skeleton is constructed based on the principles of combustion dynamics, and the data correction part uses machine learning algorithms to perform nonlinear compensation on the residuals of the mechanism model. The NO x The mathematical form of the hybrid prediction model for the generated amount is: in, NO x Generate concentration prediction values; It is a thermal NO x Generation rate coefficient; For macroscopic activation energy; It is the universal gas constant; This refers to the flame temperature in the combustion core region. Oxygen concentration in the main combustion zone; To obtain the basic nitrogen content; For fuel nitrogen to NO x Conversion rate coefficient; This is a nonlinear residual correction term; It is a thermal NO x Generation mechanism term; NO is a fuel type x Generation mechanism term; This is a nonlinear residual correction term based on support vector regression, whose input variables include boiler load. , Burnout wind opening Variation in powder supply and powder stability index ; The mathematical form of the prediction model for the carbon content of fly ash is: in, The proportion of unburned carbon remaining in fly ash; For coal powder fineness; It is a dry, ash-free volatile matter; Oxygen content; The model coefficients are updated in real time using data from an online fly ash carbon content monitor and a recursive least squares algorithm to achieve model adaptation. The SO x The mathematical form of the hybrid prediction model for the generated amount is: in, For SO x Generate concentration prediction values; The total sulfur content of the coal received for use in the furnace; The sulfur conversion coefficient; For in-furnace desulfurization efficiency; This is the correction term function based on Gaussian process regression; For SO x The dominant mechanism of generation; This is a correction term based on Gaussian process regression, and its input features include bed temperature / furnace temperature. Excess air coefficient and effective calcium oxide content This correction term also outputs the confidence interval of the predicted value; The in-furnace desulfurization efficiency ,in, It is an empirical fitting index; and the effective calcium oxide content The effective content of alkaline oxides in coal ash is obtained by weighted calculation of ash component analysis data.

[0012] Preferably, in step S2, the powder feeding variation Δ M Frequency of powder feeding variation f and powder stability index s Defined quantitatively as follows: The variation range of powder feeding Δ MThe calculation formula is: in, For sampling time window; The baseline powder feeding amount; M ( t () represents the amount of powder supplied at the current moment; for Historical powder distribution before the time; Δ M ( t () represents the change in powder supply at the current moment; The frequency of powder feeding variation f The calculation formula is: in, The number of times the direction of the toner supply command changes within a unit time window; For statistical time windows; The powder feeding stability index s The calculation formula is: in, n This represents the number of sampling points; M i For the first i The powder feeding value at each sampling point; This is the arithmetic mean of the amount of powder given during the statistical period.

[0013] Preferably, in step S3, establishing a quantitative mapping relationship between the full characteristic parameters of coal, the quantitative indicators, and boiler combustion efficiency, pollutant emission concentration, and coking risk specifically includes: S31. In-depth mining of historical operational data: Extract historical operational data and establish the variation range Δ of powder distribution through association rule mining and partial least squares regression. M Frequency of powder feeding variation f and powder stability index s Preliminary correlation with coking risk, pollutant emission concentration, and burnout rate; S32. Multiphysics numerical simulation: Computational fluid dynamics (CFD) numerical simulation is used to obtain the temperature field, component field and flow field distribution in the furnace. Combined with ASPEN PLUS process simulation, the combustion efficiency, pollutant generation and coking tendency are calculated under different coal quality characteristics and pulverized coal feeding operation parameters. S33. Construction of a Quantitative Mapping Relationship Library: This involves fusing historical operational data mining results with CFD and ASPEN PLUS simulation data. Through regression analysis and interpolation fitting, a quantitative mapping relationship library is established, using coal quality characteristic parameters and pulverized feed quantitative indicators as inputs, and boiler thermal efficiency and NO2 as inputs. x Emission concentration, SOx A standardized quantitative mapping relation library is formed by using a multivariate function relation library with emission concentration, fly ash carbon content, and coking growth rate as outputs; S34. Mapping Relationship Update: Based on the running data and simulation results, the quantitative mapping relationship library is corrected online at a set period. In step S33, the quantitative mapping relationship library includes environmental mapping relationships, economic mapping relationships, safety mapping relationships, and burnout mapping relationships: The environmental mapping relationship includes NO. x Emission concentration mapping relationship and SO x Emission concentration mapping relationship; The NO x Emission concentration mapping relationship based on the change in feed Δ M Dry, ash-free volatile matter And the burnout air opening degree are used as input variables, with NO x Emission concentration is the output variable; the NO x The mathematical form of the emission concentration mapping relationship is: in, Y 1 is NO x Emission concentration forecasts; For volatile matter of NO x The generated impact coefficient; b and k The coefficient representing the nonlinear influence of powder feeding fluctuations; f ( X 3 () is a function of the burnout air opening degree; The SO x Emission concentration mapping relationship based on total sulfur content of coal received into the furnace and the variation in powder supply Δ M For input variables, in SO x Emission concentration is the output variable; the SO x The mathematical form of the emission concentration mapping relationship is: in, Y 2 is SO x Emission concentration forecasts; This is the comprehensive emission factor; To adjust the SO based on the fluctuation of powder supply x The generated sensitivity coefficient; This is an index term representing the impact of powder disturbance. The economic mapping relationship is based on flue gas oxygen content. O 2. Smoke exhaust temperature and powder stability index s The input variable is boiler thermal efficiency, and the output variable is boiler thermal efficiency; the mathematical form of the economic mapping relationship is: in, Y 3 represents the predicted value of boiler thermal efficiency; The baseline thermal efficiency; C 1 represents the penalty coefficient for oxygen deviation; O opt For optimal operating oxygen levels; C 2 represents the penalty coefficient for unstable powder distribution; C 3 represents the penalty coefficient for deviation of exhaust gas temperature; To achieve the optimal smoke exhaust temperature; The safety mapping relationship is based on the furnace outlet temperature. T ch Ash composition and alkali-acid ratio B / A and boiler load Load The input variable is the coke growth rate, and the output variable is the coke growth rate; the mathematical form of the safety mapping relationship is: in, Y 4 represents the predicted growth rate of coking. T soft This refers to the ash softening temperature. This is the temperature sensitivity coefficient; For reference load; The burnout mapping relationship is based on pulverized coal fineness. R 90 Fixed carbon content in coal The mathematical form of the burnout mapping relationship, with fuel residence time in the furnace as the input variable and fly ash carbon content as the output variable, is as follows: in, Y 5 represents the predicted carbon content of fly ash; kz This is the comprehensive impact coefficient; h This refers to the residence time of fuel inside the furnace. r The index is influenced by the length of stay; In step S34, the environmental protection mapping relationship is updated in real time, the economic mapping relationship is corrected daily, the safety mapping relationship is updated by shift, and the burnout mapping relationship is updated by hour.

[0014] Preferably, step S4 includes: S41. Construction of a Multi-Objective Optimization Model: Taking economy, environmental protection, and safety as optimization objectives, a multi-objective optimization function vector is constructed, whose mathematical form is defined as: in, Optimize function vectors for multiple objectives; The objective function is the economic performance. The objective function is environmental protection. The objective function is security. T Indicates transpose; The economic objective function Coal consumption for power supply Its value is determined by the boiler thermal efficiency predicted in step S33 and the mechanical incomplete combustion loss rate. The calculation process is as follows: First, the carbon content of fly ash is predicted using the burnout mapping relationship in step S33, and the heat loss rate due to incomplete mechanical combustion is calculated. : in, is a physical constant representing the calorific value of carbon; The ash content of the coal received before it is fed into the furnace; This refers to the fly ash proportion coefficient. The received basis lower calorific value of the coal fed into the furnace; Then, the baseline thermal efficiency is predicted using the economic mapping relationship in step S33, and the final boiler thermal efficiency is calculated. : Ultimately, coal consumption for power generation for: Where 123 is the unit conversion factor, which is a constant that converts thermal efficiency into standard coal consumption; For pipeline efficiency; This refers to the internal efficiency of the steam turbine. For mechanical efficiency; For generator efficiency; Plant power consumption rate; and ; The environmental protection objective function The comprehensive pollutant emissions are calculated, and their value is the NO value in step S33. x Emission concentration predictions and SO x Weighted normalized form of predicted emission concentrations: in, NO x Emission weighting coefficient; For SO x Emission weighting coefficient; NO x Environmental regulations emission limits; For SO x Environmental regulations emission limits; The security objective function This is the coking risk index, whose value incorporates the final coking risk probability output from the coking tendency prediction model in step S1. The current predicted coking growth rate is mapped to the safety relationship in step S33. Y 4, that is: in, This refers to the static risk weighting coefficient. This is a dynamically increasing risk weighting coefficient; S42. Constraint Settings: Set boiler wall temperature constraints and quantitative index range constraints. The boiler wall temperature constraint is used to limit the furnace wall temperature from exceeding the safety limit. The quantitative index range constraint is used to limit the pulverized coal variation range Δ. M Frequency of powder feeding variation f and powder stability index s Within the limits allowed by the equipment; S43. Multi-objective optimization solution: The NSGA-II multi-objective genetic algorithm is used to solve the Pareto optimal solution set of the multi-objective optimization function vector under the constraints. S44. Optimal Strategy Selection: Based on the power plant's operational strategy preferences, a weighted method is used to select the final solution from the Pareto optimal solution set, and the optimal feed rate variation Δ that adapts to the current coal quality is output. M Frequency of powder feeding variation f and powder stability index s The dynamic control strategy is generated and sent to the system for execution.

[0015] Preferably, in step S43, the NSGA-II algorithm includes the following steps: S431. Population initialization and real number encoding: The initial population is generated using real number encoding. Each individual represents a set of decision parameter vectors, namely, a combination of pollen feeding variation amplitude, pollen feeding variation frequency, and pollen feeding stability index. S432. Fitness Assessment: Substitute the decision parameter vector of each individual into the quantitative mapping relation library established in step S33, call the economic mapping relation and burnout mapping relation, and calculate the economic objective function value. And by invoking the environmental mapping relationship, calculate the environmental objective function value. The coking tendency prediction model established in step S1 is substituted into the model, and the coking growth rate predicted by the safety mapping relationship in step S33 is combined to calculate the safety objective function value. ; S433. Fast Non-Dominated Sort: Calculate the dominated count and dominant set of each individual, assign non-dominated individuals to the first front surface, iterate through the dominant set of each non-dominated individual and assign them to subsequent front surfaces in turn, until all individuals are stratified. S434. Crowding distance calculation: For individuals on the same frontal surface, calculate their crowding distance in the target space, and prioritize individuals with larger crowding distances to maintain population diversity; S435. Genetic operations: Select, crossover and mutation operations are performed using binary tournament selection, simulated binary crossover and polynomial mutation to generate a new generation of population; S436. Elite Preservation and Generation of New Generation: Merge the parent and offspring populations, re-perform fast non-dominated sorting and crowding distance calculation, and select the top N best individuals to generate a new generation population; S437. Iteration Termination and Solution Output: Repeat steps S432 to S436 until the preset number of iterations is reached, and output the Pareto optimal solution set.

[0016] Preferably, after step S4, the method further includes: Step S5. Dynamic self-evolution update: Establish a three-level update mechanism of data layer - model layer - strategy layer; the data layer automatically updates the coal quality database and the operation dataset when it detects coal quality fluctuations or operational efficiency deviations exceeding the threshold; the model layer uses online machine learning algorithms to iteratively optimize the coking tendency prediction model and the burnout and pollutant generation prediction model; the strategy layer synchronously adjusts the parameters of the multi-objective optimization function vector according to the model update results.

[0017] Due to the adoption of the above technical solutions, the technical progress achieved by this invention is as follows.

[0018] This invention achieves multi-objective synergistic optimization of economy, environmental protection and safety: By constructing a multi-objective function vector with economic efficiency, environmental protection, and safety as optimization goals, and employing the NSGA-II algorithm for optimization, this invention can simultaneously consider the economic benefits, environmental compliance, and operational safety of boiler operation. Compared to traditional single-objective optimization methods, this invention, through the multi-objective optimization model in step S4, effectively reduces pollutant emission concentrations while ensuring improved boiler thermal efficiency and controlling coking risks within a safe range, achieving a balance and coordination among multiple objectives.

[0019] This invention significantly improves boiler thermal efficiency and reduces coal consumption for power generation: By establishing an economic mapping relationship based on all coal quality characteristics, the boiler thermal efficiency can be accurately predicted. Furthermore, the mechanical incomplete combustion loss characterized by the fly ash carbon content predicted by the burnout mapping relationship in step S33 is innovatively incorporated into the economic objective function, thereby achieving a significant improvement in boiler combustion efficiency.

[0020] This invention effectively reduces pollutant emissions and achieves ultra-low emission operation: By constructing NO x and SO x The emission concentration mapping relationship, combined with a multi-objective optimization algorithm, spontaneously incorporates NO while pursuing economic efficiency. x and SO x Emission concentrations are kept low, meeting the requirements of national ultra-low emission standards.

[0021] This invention can accurately predict the risk of coking, ensuring the safe and stable operation of boilers. Based on the coking tendency prediction model, combined with parameters such as coal ash composition and furnace temperature field distribution, the probability of coking risk is accurately predicted, providing early warning information for operators. By optimizing the pulverized coal feeding strategy through the safety objective function that integrates static risk and dynamic growth rate in step S4, furnace coking is effectively prevented, significantly improving the safety and reliability of boiler operation, reducing unplanned shutdowns, and extending equipment service life.

[0022] This invention can improve the system's self-adaptability and adapt to complex and changing operating conditions: By establishing a quantitative mapping database of coal quality parameters and combustion performance, the system achieves rapid response and adaptive adjustment to changes in coal quality entering the furnace. The system can automatically adjust the pulverized coal feeding strategy and air distribution parameters based on coal characteristics (such as volatile matter, ash content, and sulfur content), adapting to the operational needs of different coal types and load conditions, and improving the unit's operational stability and flexibility under complex conditions such as deep peak shaving and rapid load changes.

[0023] This invention achieves significant breakthroughs in multiple dimensions through a closed-loop technology loop of "multi-dimensional coal quality database - multi-physics field mapping - adaptive multi-objective optimization - dynamic self-evolution," as detailed below: Technical aspects: The coal quality characterization dimensions are expanded to more than 15 items, the coal type identification accuracy is ≥98%, the burnout rate and pollutant generation prediction error is ≤3%, the combustion mapping deviation is reduced from the traditional 10-15% to ≤5%, the optimization strategy generation time is ≤10 seconds, the control response delay is ≤2 seconds, the model accuracy decay rate is ≤1% / month, and it can adapt to the co-firing of multiple coal types and the fluctuation of 30-100% rated load, with the adaptability range expanded by more than 3 times compared with the traditional system; Economically speaking: A 300MW unit can save 8,000-12,000 tons of standard coal and 200,000-300,000 kWh of electricity per year. The maintenance cycle of the heating surface is extended by 50%, the replacement cycle of spare parts for the pulverizer is extended by 50%, the comprehensive annual benefit is 8 million-12 million yuan, and the investment payback period is ≤2 years (1000MW units can be reduced to 1.5 years). Environmental aspects: NO x Emissions reduction of 12-18%, SO2 xWith an emission reduction of 8-12%, the ultra-low emission standards can be stably met. The 300MW unit will reduce ash and slag emissions by 5,000-8,000 tons per year, indirectly reduce carbon emissions by 20,000-30,000 tons, and can also generate carbon trading revenue.

[0024] Safety aspects: Coking risk warning is given ≥40 minutes in advance, reducing the boiler coking accident rate by more than 70%, reducing the furnace temperature difference from ±80℃ to ±68℃, reducing the unplanned shutdown rate from 3-5% to 1-2%, reducing annual downtime losses by 5-8 million yuan, and the system automation rate of ≥95% reduces the frequency of manual intervention by 70%, thus reducing operational safety risks. From an operation and maintenance perspective: the number of personnel in the coal management position can be reduced from 3-4 to 2, the fault diagnosis time can be reduced from 4-6 hours to 1-2 hours, modular deployment is supported, the implementation cycle is ≤3 months and no large-scale modification of existing equipment is required; The overall goal is to achieve comprehensive benefits that are "technologically advanced, economically feasible, environmentally compliant, safe and stable, and convenient to operate and maintain," providing key support for the intelligent and low-carbon transformation of coal-fired power plants. Attached Figure Description

[0025] Figure 1 This is a flowchart of the present invention. Detailed Implementation

[0026] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0027] A multi-objective optimization and dynamic pulverized coal feeding control method based on the full characteristics of coal quality, as shown in the figure, includes the following steps: S1. Characterization and Model Building of Coal Quality: Construct a multi-dimensional coal quality database containing all coal quality characteristic parameters; based on the coal quality database, establish a coking tendency prediction model and a burnout and pollutant generation prediction model respectively.

[0028] This step is for building the basic model for "offline / long cycle" models, which are responsible for providing physical benchmarks and probabilistic basis.

[0029] As shown in Table 1 below, the full range of coal quality parameters includes industrial analysis parameters, elemental analysis parameters, ash characteristic parameters, and combustion characteristic parameters for the coal type. The industrial analysis parameters include moisture content on an as-received basis, ash content on an as-received basis, volatile matter on a dry ash-free basis, fixed carbon, and lower heating value on an as-received basis. The elemental analysis parameters include carbon, hydrogen, oxygen, nitrogen, and total sulfur. The ash characteristic parameters include silica, alumina, iron oxide, calcium oxide, and magnesium oxide / potassium oxide / sodium oxide. The combustion characteristic parameters include ignition temperature and burnout characteristic index.

[0030] Table 1. Complete coal quality parameters

[0031] (1) The coking tendency prediction model takes the ash composition (SiO2, Al2O3, Fe2O3, etc.) and furnace temperature field distribution as inputs, and adopts the random forest + mechanism constraint algorithm to establish a quantitative correlation model between coking risk and coal quality-operating conditions. The prediction accuracy is over 90%.

[0032] To enable machine learning models (random forests) to capture patterns more effectively, it's not enough to simply input raw data; derived features must be constructed based on combustion chemistry mechanisms. These features are often the culmination of domain experts' experience and can significantly improve the model's interpretability and generalization ability.

[0033] Specifically, methods for establishing coking tendency prediction models include: S11. Feature Engineering Processing: Based on the computer-derived characteristics of coal ash composition, including the silicon-to-aluminum ratio. SR Alkali-acid ratio B / A Slagging index R s and pollution index R f At least one of them.

[0034] Specifically, the calculation formula for mechanism-derived characteristics is as follows: ①Silicon-to-aluminum ratio SR : This indicator reflects the relative proportion of acidic components in the ash and is closely related to the viscosity of the slag.

[0035] ② Alkali-acid ratio B / A : This is the most classic mechanistic indicator for judging the tendency of coking. B / A The higher the value, the lower the ash melting point usually is, and the higher the risk of coking. Especially in... B / A≈0.4~0.7 Within this range, the probability of coking increases exponentially.

[0036] ③ Slagging Index R s Taking into account the influence of sulfur content, it is defined as: in, This represents the total sulfur content on a dry basis.

[0037] R s < 0.6 Low risk of slagging; 0.6≤ R s < 2.0 Medium risk; Rs ≥ 2.0 High risk of slagging.

[0038] ④ Pollution Index R f For convective heating surfaces, the influence of sodium is considered in particular and is defined as follows: This indicator is used to provide early warning of ash accumulation in the superheater and reheater, and directly guides the operation strategy of the sootblower.

[0039] S12. Dataset Construction: Using coal quality full characteristic parameters, mechanism-derived features, and real-time operating parameters as input vectors, and coking risk level as output label, construct a training dataset.

[0040] Specifically: Input vector : Including all coal quality characteristics, including moisture content. Received base ash Dry, ash-free, and free of volatile matter. Mechanism indicators SR , B / A and real-time operating parameters (furnace outlet temperature FEGT, boiler load) Oxygen Primary air ratio Secondary air ratio ).

[0041] Output Labels Coking Risk Level (None, Mild, Moderate, Severe).

[0042] Tag Acquisition: The coking risk level is acquired using a semi-supervised labeling method, which combines indirect indicators from historical operating data, such as sootblower operation frequency, abnormal increase in desuperheating water flow, and furnace heat absorption attenuation rate, with manual records to label the level.

[0043] S13. Random Forest Model Training and Probability Output: Multiple decision trees are generated using the Bootstrap sampling strategy. Random feature subspace selection is used when splitting decision tree nodes. Feature selection is performed based on the Gini impurity minimization criterion to train the random forest model. The coal quality full characteristic parameters, mechanism-derived features, and real-time operating parameters under the current working conditions are used as input vectors and input into the random forest model to output the predicted probability of the random forest model.

[0044] Specifically, in the Bootstrap strategy used in random forest training, the training set for each decision tree is generated through random sampling with replacement, out-of-bag data is used for unbiased model validation, and the size of the feature subspace is set to the order of the square root of the total number of features. Technical effect: By randomly perturbing the training set, the sensitivity of individual trees to specific noisy samples is smoothed, significantly improving the model's robustness under sensor drift conditions.

[0045] The coal combustion process involves dozens of dimensions of characteristics, including coal quality, flow field, and temperature field. If all characteristics are considered, the process is easily dominated by strongly correlated variables (such as load), masking the influence of microscopic coal quality. Feature selection is based on the Gini impurity minimization criterion, specifically as follows: In a randomly selected feature subspace, all possible thresholds for each feature are traversed, and the weighted Gini impurity corresponding to each candidate split is calculated. The feature and threshold combination that minimizes the weighted Gini impurity is selected as the optimal splitting scheme; where nodes... The formula for calculating the impurity of Gini is: in, For the node belonging to the first j The probability of coking-like risk; This represents the total number of coking risk categories.

[0046] Technical effect: By forcing some decision trees to focus on "internal factors of coal quality" (such as Fe2O3) and some to focus on "external factors of operation" (such as damper opening), the decoupling and full utilization of multi-source characteristics are achieved.

[0047] S14. Mechanism Constraint Fusion: The probability predicted by the random forest model is weighted and fused with the probability of mechanism constraints based on physical rules through the Bayesian posterior correction formula to output the final coking risk probability.

[0048] The fusion of mechanistic constraints is intended to prevent misjudgments by pure data models, which may produce predictions that contradict common sense in physics.

[0049] Specifically, the Bayesian posterior correction formula is as follows: in, The final coking risk probability is denoted as [0,1]. The closer the value is to 1, the higher the coking risk. This value is the final criterion after the weighted fusion of the data-driven model and the physical mechanism model.

[0050] The probability prediction for the random forest model, which is the voting ratio (Probability Estimate) of the "high-risk coagulation" category output by the random forest classifier, represents a statistical pattern judgment based on historical operating data.

[0051] Mechanism-constrained probabilities are hard-constrained probabilities generated based on physical rules. Example of rule setting: If the flue gas temperature at the furnace outlet (FEGT) > the ash deformation temperature (DT), then let Otherwise, according to B / A The ratio is linearly mapped over the interval [0, 0.5].

[0052] The confidence weighting factor, based on the similarity between the current operating conditions and historical data distribution, ranges from [0,1] and is used to dynamically adjust the reference weights of the data model and the mechanism model. The calculation logic is as follows: Confidence weighting factor Dynamically adjust based on the Mahalanobis distance between the current operating point and the training set center: When the operating condition is in a region with high density of historical data Approaching 1 (Trust data model), primarily based on random forest prediction results; When the operating condition is in a region with sparse historical data or an extrapolated region Approaching 0 (Forced regression to the physical baseline to ensure safety), with the prediction results primarily constrained by the mechanism.

[0053] (2) Methods for establishing prediction models of burnout and pollutant formation include: Construct a hybrid prediction model architecture that targets NO x Generation amount, fly ash carbon content and SO x The generated data all adopt a "mechanism framework + data correction" approach. The mechanism framework is constructed based on combustion dynamics principles, while the data correction uses machine learning algorithms to nonlinearly compensate for the residuals of the mechanism model. Details are as follows: ①NO x The mathematical form of the hybrid prediction model for the generated amount is: in, NO x Generate concentration prediction values; It is a thermal NO x The generation rate coefficient is a constant obtained by regression fitting of historical data, reflecting the sensitivity of a specific furnace type to temperature; The macroscopic activation energy describes the energy barrier for the occurrence of nitrogen oxidation reaction, and its value is usually taken within the equivalent range of the Zeldovich mechanism. It is the universal gas constant; The flame temperature in the core combustion zone can be obtained from data from a dual-color thermometer or through heat balance calculations. The oxygen concentration in the main combustion zone characterizes the concentration of the combustion-supporting agent. The nitrogen content received is from real-time coal quality data; it is a fuel-type NO. x The material basis; For fuel nitrogen to NO x The conversion rate coefficient reflects the proportion of nitrogen in coal oxidized to NO, and is usually related to the volatile matter release characteristics; This is a nonlinear residual correction term, a scalar value output by the support vector regression model, used to compensate for the effects of complex physical processes such as flow field turbulence intensity and burnout wind mixing delay on NO that cannot be described by simple mechanistic formulas. x The impact; It is a thermal NO x Generation mechanism term; NO is a fuel type x Generation mechanism term; This is a nonlinear residual correction term based on support vector regression, whose input variables include boiler load. , Burnout wind opening Variation in powder supply and powder stability index This structure ensures the correctness of physical trends (such as temperature rise and NO increase). x (Upgrade), and also used real-time data to capture the characteristics of specific devices.

[0054] ② The burnout rate is mainly affected by the fineness of pulverized coal, volatile matter content, and residence time.

[0055] The mathematical form of the prediction model for fly ash carbon content (burnout rate) is as follows: in, The proportion of unburned carbon remaining in fly ash; The fineness of the pulverized coal (screen residue) is the greatest; the larger the residue, the more difficult it is to burn completely. It is a dry, ash-free volatile matter; the lower the value, the more difficult it is to burn completely (the index is negative). The oxygen content is higher, which facilitates complete combustion (the index is negative). The model coefficients are updated in real time using data from an online fly ash carbon content monitor and a recursive least squares algorithm to achieve model adaptation.

[0056] ③With NO x Different, SO x The formation of sulfur mainly depends on the competition between the oxidation of fuel sulfur and the fixation of sulfur in the furnace (ash absorption).

[0057] SO x The mathematical form of the hybrid prediction model for the generated amount is: in, For SO x Generate concentration prediction values; This refers to the total sulfur content of coal received into the furnace, which is SO₂. x The source data generated; The sulfur conversion coefficient, theoretically However, this needs to be adjusted based on the actual combustion efficiency. The efficiency of in-furnace desulfurization is mainly affected by alkaline oxides in coal ash (especially...). The effect of content is calculated as follows: ,in, It is an empirical fitting index; and the effective calcium oxide content The effective content of alkaline oxides in coal ash is obtained by weighted calculation of ash component analysis data.

[0058] For SO x The dominant mechanism of generation; This is the correction term function based on Gaussian Process Regression (GPR); This is a correction term based on Gaussian process regression, and its input features include bed temperature / furnace temperature. (Affecting the equilibrium of sulfur fixation reaction), excess air coefficient (Affecting the oxidizing atmosphere) and effective calcium oxide content This correction term also outputs the confidence interval of the predicted value, which is crucial for environmental compliance control (avoiding the probabilistic risk of exceeding emission standards).

[0059] S2. Quantitative Definition of Powder Feeding Operation Parameters: Define quantitative indicators used to characterize the dynamic behavior of the powder feeding system, including the powder feeding variation amplitude Δ. M Frequency of powder feeding variation fand powder stability index s .

[0060] Powder variation Δ M Frequency of powder feeding variation f and powder stability index s Defined quantitatively as follows: (1) Variation in powder feeding Δ M The ratio of the maximum fluctuation in feed rate per unit time to the baseline feed rate reflects the adjustment intensity. Its calculation formula is: in, For sampling time window; The baseline powder feeding amount; M ( t ) represents the current amount of toner supplied; represents the historical amount of toner supplied before that time; Δ M ( t () represents the change in powder supply at the current moment.

[0061] This indicator limits the "aggressiveness" of pulverized coal feed adjustment to prevent sudden changes in pulverized coal concentration from causing flameout or deflagration.

[0062] (2) Frequency of powder feeding variation f The number of times the powder feed rate is adjusted per unit time reflects the adjustment frequency. The calculation formula is: in, The number of times the direction of the toner supply command changes within a unit time window; This represents the statistical time window.

[0063] limit f It can reduce wear on the powder feeder and smooth the combustion process.

[0064] (3) Powder feeding stability index s The coefficient of variation (standard deviation / mean) of the feed rate time series reflects the stability of the feed system. Its calculation formula is: in, n This represents the number of sampling points; M i For the first i The powder feeding value at each sampling point; This is the arithmetic mean of the amount of powder given during the statistical period.

[0065] This indicator reflects the volatility of the powder flow. If s An abnormal increase may indicate that the coal feeder has stopped feeding coal, is blocked by coal, or that pulverized coal is flowing out by gravity.

[0066] S3. Establishment of Multiphysics Mapping Relationship: Based on the model constructed in step S1, historical operation data mining, numerical simulation and process simulation, establish a quantitative mapping relationship between coal quality full characteristic parameters, quantitative indicators and boiler combustion efficiency, pollutant emission concentration and coking risk.

[0067] This step involves establishing a fast mapping function library for "online / short cycle" data to map the baseline from step S1 to the current control variable's change magnitude Δ. M Frequency of powder feeding variation f and powder stability index s Establish functional relationships as the computational kernel for S4 optimization.

[0068] Establish a quantitative mapping relationship between coal quality parameters, quantitative indicators, and boiler combustion efficiency, pollutant emission concentration, and coking risk, specifically including: S31. In-depth mining of historical operation data: Extract historical operation data from DCS and SIS systems, and establish the powder feeding variation range Δ through association rule mining and partial least squares regression. M Frequency of powder feeding variation f and powder stability index s Preliminary correlation with coking risk, pollutant emission concentration, and burnout rate.

[0069] S32. Multiphysics numerical simulation: Computational fluid dynamics (CFD) numerical simulation is used to obtain the temperature field, component field and flow field distribution in the furnace. Combined with ASPEN PLUS process simulation, the combustion efficiency, pollutant generation and coking tendency are calculated under different coal quality characteristics and pulverized coal feeding operation parameters.

[0070] S33. Construction of a Quantitative Mapping Relationship Library: This involves fusing historical operational data mining results with CFD and ASPEN PLUS simulation data. Through regression analysis and interpolation fitting, a quantitative mapping relationship library is established, using coal quality characteristic parameters and pulverized feed quantitative indicators as inputs, and boiler thermal efficiency and NO2 as inputs. x Emission concentration, SO x A multivariate function relation library with emission concentration, fly ash carbon content, and coking growth rate as outputs is formed to create a standardized quantitative mapping relation library, which serves as the "lookup basis" or "computation kernel" for subsequent multi-objective optimization.

[0071] As shown in Table 2, the quantitative mapping relationship library includes environmental mapping relationships, economic mapping relationships, safety mapping relationships, and burnout mapping relationships: (1) Environmental mapping relationships are no longer merely static predictions, but reflect the marginal impact of control variables on emissions, including NO. x Emission concentration mapping relationship and SO x Emission concentration mapping relationship; ①NOx Emission concentration mapping relationship based on the change in feed Δ M Dry, ash-free volatile matter And the burnout air opening degree are used as input variables, with NO x Emission concentration is the output variable; NO x The mathematical form of the emission concentration mapping relationship is: in, Y 1 is NO x Emission concentration forecasts; For volatile matter of NO x The generated impact coefficient; b and k The coefficient representing the nonlinear influence of powder feeding fluctuations; f ( X 3 ) is a function of the burnout air opening.

[0072] ②SO x Emission concentration mapping relationship based on total sulfur content of coal received into the furnace and the variation in powder supply Δ M For input variables, in SO x Emission concentration is the output variable; SO x The mathematical form of the emission concentration mapping relationship is: in, Y 2 is SO x Emission concentration forecasts; This is the comprehensive emission factor; To adjust the SO based on the fluctuation of powder supply x The generated sensitivity coefficient; This is an index term representing the impact of powder disturbance.

[0073] (2) Economic mapping relationship based on flue gas oxygen content O 2. Smoke exhaust temperature and powder stability index s The input variable is boiler thermal efficiency, and the output variable is boiler thermal efficiency; the mathematical form of the economic mapping relationship is: in, Y 3 represents the predicted value of boiler thermal efficiency; The baseline thermal efficiency; C 1 represents the penalty coefficient for oxygen deviation; O opt For optimal operating oxygen levels; C 2 represents the penalty coefficient for unstable powder distribution; C 3 represents the penalty coefficient for deviation of exhaust gas temperature; The optimal smoke exhaust temperature.

[0074] The function exhibits a parabolic distribution, which not only penalizes deviations in oxygen levels but also captures combustion pulsation losses caused by powder instability.

[0075] (3) Safety mapping relationship based on furnace outlet temperature T ch Ash composition, alkali-acid ratio (B / A), and boiler load. L The input variable is coke growth rate, and the output variable is coke growth rate; the mathematical form of the safety mapping relationship is: in, Y 4 represents the predicted growth rate of coking. T soft This refers to the ash softening temperature. This is the temperature sensitivity coefficient; For reference load, the rated load of the boiler is usually taken; This characterizes the amplification effect of increased furnace volumetric heat load on slagging rate under high load. This dynamic rate term will directly participate in the construction of the safety objective function in S4.

[0076] Unlike the "final coking risk probability" output in step S1 (which represents the overall safety in the current state), this is different. Y 4. The dynamic rate of coking with increasing temperature and load was characterized.

[0077] (4) Combustion mapping relationship based on coal powder fineness R 90 Fixed carbon content in coal The mathematical form of the burnout mapping relationship, with fuel residence time in the furnace as the input variable and fly ash carbon content as the output variable, is as follows: in, Y 5 represents the predicted carbon content of fly ash; kz This is the comprehensive impact coefficient; h This refers to the residence time of fuel inside the furnace. r The index is influenced by the length of stay.

[0078] This formula improves the logic of burnout rate prediction from two dimensions: internal factors (the ratio of fixed carbon to volatile matter) and external factors (fineness and residence time).

[0079] Table 2 Functional Architecture of Quantitative Mapping Relationship Library

[0080] S34. Mapping Relationship Update: Based on the running data and simulation results, the quantitative mapping relationship library is corrected online at a set period.

[0081] Specifically, the environmental mapping relationship is updated in real time (every 5 minutes), the economic mapping relationship is corrected daily, the safety mapping relationship is updated by shift (8 hours), and the burnout mapping relationship is updated hourly (every hour).

[0082] S4. Multi-objective optimization and strategy solution: With economy, environmental protection, and safety as optimization objectives, a multi-objective optimization function is constructed. Under equipment operation constraints, a multi-objective optimization algorithm is used to solve for the Pareto optimal solution set. The final solution is selected from the Pareto optimal solution set to generate the optimal feed rate variation Δ adapted to the current coal quality. M Frequency of powder feeding variation f and powder stability index s The dynamic control strategy is as follows: S41. Construction of a Multi-Objective Optimization Model: Taking economy, environmental protection, and safety as optimization objectives, a multi-objective optimization function vector is constructed, whose mathematical form is defined as: in, Optimize function vectors for multiple objectives; The objective function is the economic performance. The objective function is environmental protection. The objective function is security. T This indicates transpose.

[0083] Multi-objective optimization function vector The goal is to find the parameter vector. X Make All indicators in the target function are minimized simultaneously, that is, the economic objective function, the environmental objective function, and the safety objective function are minimized simultaneously: ① Economic objective function Coal consumption for power supply Its value is determined by the boiler thermal efficiency predicted in step S33 and the mechanical incomplete combustion loss rate. The calculation process is as follows: First, the carbon content of fly ash is predicted using the burnout mapping relationship in step S33, and the heat loss rate due to incomplete mechanical combustion is calculated. : in, is a physical constant representing the calorific value of carbon; The ash content of the coal received before it is fed into the furnace; This refers to the fly ash proportion coefficient. The received basis lower calorific value of the coal fed into the furnace; Then, the baseline thermal efficiency is predicted using the economic mapping relationship in step S33, and the final boiler thermal efficiency is calculated. : Ultimately, coal consumption for power generation for: Where 123 is the unit conversion factor, which is a constant that converts thermal efficiency into standard coal consumption; For pipeline efficiency; This refers to the internal efficiency of the steam turbine. For mechanical efficiency; For generator efficiency; Plant power consumption rate; and .

[0084] ② Environmental protection objective function The comprehensive pollutant emissions are calculated, and their value is the NO value in step S33. x Emission concentration predictions and SO x Weighted normalized form of predicted emission concentrations: in, NO x Emission weighting coefficient; For SO x Emission weighting coefficient; NO x Environmental regulations emission limits; For SO x The function determines the emission limits set by environmental regulations. It ensures that the total score spikes dramatically when any indicator approaches the red line.

[0085] ③ Security objective function This is the coking risk index, whose value incorporates the final coking risk probability output from the coking tendency prediction model in step S1. (Static Risk) and Step S33 Safety Mapping Relationship: Current Predicted Coking Growth Rate Y 4 (Dynamic growth), that is: in, This refers to the static risk weighting coefficient. This is a dynamic growth risk weighting coefficient. This formula ensures that even if the baseline risk is low, if the fan-giving strategy leads to an excessively rapid growth rate, the safety target will still be deemed unsatisfactory.

[0086] The multi-objective optimization function vector successfully constructed a trade-off Pareto front: Pursuing extremely low environmental performance usually requires lowering combustion temperatures or delaying mixing, which may lead to increased fuel economy.

[0087] The pursuit of extremely high economic efficiency usually requires low-oxygen combustion and high temperatures, which in turn leads to a reduction in safety.

[0088] The task of multi-objective optimization of function vectors is to find the optimal balance point among these contradictions.

[0089] S42. Constraint Settings: Set boiler wall temperature constraints and quantitative index range constraints. The boiler wall temperature constraint limits the furnace wall temperature to within safe limits, while the quantitative index range constraint limits the pulverized coal variation range Δ. M Frequency of powder feeding variation f and powder stability index s Within the limits allowed by the equipment. Specifically: ① Boiler wall temperature constraint: in, This is the measured value of the furnace wall temperature; This refers to the safe limit for furnace wall temperature. ② Quantitative indicator range constraints: , ,

[0090] in, These are the lower and upper limits of the variation range for powder application; These are the lower and upper limits of the frequency of powder feeding variation; These are the lower and upper limits of the powder stability index.

[0091] S43. Multi-objective optimization solution: The NSGA-II multi-objective genetic algorithm is used to solve the Pareto optimal solution set of the multi-objective optimization function vector under constraints.

[0092] Specifically, the NSGA-II algorithm includes the following steps: S431. Population Initialization and Real Number Encoding: The initial population is generated using real number encoding. Each individual represents a set of decision parameter vectors, namely, a combination of pollen feeding variation amplitude, pollen feeding variation frequency, and pollen feeding stability index.

[0093] S432. Fitness Assessment: Substitute the decision parameter vector of each individual into the quantitative mapping relation library established in step S33, call the economic mapping relation and burnout mapping relation, and calculate the economic objective function value. And by invoking the environmental mapping relationship, calculate the environmental objective function value. The coking tendency prediction model established in step S1 is substituted into the model, and the coking growth rate predicted by the safety mapping relationship in step S33 is combined to calculate the safety objective function value. ; S433. Fast Non-Dominated Sort: Calculate the dominated count and dominant set of each individual, assign non-dominated individuals to the first front surface, iterate through the dominant set of each non-dominated individual and assign them to subsequent front surfaces in turn, until all individuals are stratified. S434. Crowding Distance Calculation: For individuals on the same frontier, calculate their crowding distance in the target space, and prioritize individuals with large crowding distances (i.e., solutions located in sparse regions) to maintain population diversity. Specifically, crowding distance The calculation formula is as follows: in, For the first i The crowding distance of each individual solution; the larger the value, the sparser the region where the solution is located and the better the diversity. m The index of the objective function; M The total number of objective functions; ), To group individuals in the same layer according to their order of birth m After sorting the target values, the first... i The objective function values ​​of two adjacent individuals of each individual; For this frontier in the 1st m Maximum and minimum values ​​in each target direction.

[0094] S435. Genetic operations: Select, crossover and mutation operations are performed using binary tournament selection, simulated binary crossover and polynomial mutation to generate a new generation of population; Specifically, binary tournament selection is based on non-dominated ranking rank (the smaller the better) and crowding distance (the larger the better); simulated binary crossover (SBX) is used to simulate continuous changes in real variables; and polynomial mutation is used to introduce small perturbations.

[0095] S436. Elite Preservation and Generation of New Generation: Merge the parent and offspring populations, re-perform fast non-dominated sorting and crowding distance calculation, and select the top N best individuals to generate a new generation population; S437. Iteration Termination and Solution Output: Repeat steps S432 to S436 until the preset number of iterations is reached, and output the Pareto optimal solution set.

[0096] S44. Optimal Strategy Selection: Based on the power plant's operational strategy preferences, a weighted method is used to select the final solution from the Pareto optimal solution set, and the optimal feed rate variation Δ that is suitable for the current coal quality is output. M Frequency of powder feeding variation f and powder stability index sThe generated dynamic control strategy is then sent to the DCS system for execution.

[0097] Step S5. Dynamic self-evolutionary update: Establish a three-level update mechanism of data layer - model layer - strategy layer; the data layer automatically updates the coal quality database and the operation dataset when it detects coal quality fluctuations or operational efficiency deviations exceeding the threshold; the model layer uses online machine learning algorithms to iteratively optimize the coking tendency prediction model and the burnout and pollutant generation prediction model; the strategy layer synchronously adjusts the parameters of the multi-objective optimization function vector according to the model update results.

Claims

1. A multi-objective optimization and dynamic pulverized coal feeding control method based on the full characteristics of coal quality, characterized in that: Includes the following steps: S1. Characterization and Model Building of Coal Quality: Construct a multi-dimensional coal quality database containing all coal quality characteristic parameters; based on the coal quality database, establish a coking tendency prediction model and a burnout and pollutant generation prediction model respectively; S2. Quantitative Definition of Powder Feeding Operation Parameters: Define quantitative indicators used to characterize the dynamic behavior of the powder feeding system, including the powder feeding variation amplitude Δ. M Frequency of powder feeding variation f and powder stability index σ ; S3. Establishment of multiphysics mapping relationship: Based on the model constructed in step S1, historical operation data mining, numerical simulation and process simulation, establish a quantitative mapping relationship between the full characteristic parameters of coal quality, the quantitative indicators and boiler combustion efficiency, pollutant emission concentration and coking risk; S4. Multi-objective optimization and strategy solution: With economy, environmental protection, and safety as optimization objectives, a multi-objective optimization function is constructed. Under equipment operation constraints, a multi-objective optimization algorithm is used to solve for the Pareto optimal solution set. The final solution is selected from the Pareto optimal solution set to generate the optimal feed rate variation Δ adapted to the current coal quality. M Frequency of powder feeding variation f and powder stability index σ Dynamic control strategy.

2. The method for multi-objective optimization and dynamic pulverized coal feeding control based on the full characteristics of coal quality according to claim 1, characterized in that: In step S1, the coal quality parameters include industrial analysis parameters, elemental analysis parameters, ash characteristic parameters, and combustion characteristic parameters of the coal type; the industrial analysis parameters include as-received moisture content, as-received ash content, dry ash-free volatile matter, fixed carbon, and as-received lower heating value; the elemental analysis parameters include carbon, hydrogen, oxygen, nitrogen, and total sulfur; the ash characteristic parameters include silicon dioxide, aluminum oxide, iron oxide, calcium oxide, and magnesium / potassium / sodium oxide; and the combustion characteristic parameters include ignition temperature and burnout characteristic index.

3. The method for multi-objective optimization and dynamic pulverized coal feeding control based on the full characteristics of coal quality according to claim 1, characterized in that: In step S1, the specific method for establishing the coking tendency prediction model includes: S11. Feature Engineering Processing: Based on the computer-derived characteristics of coal ash composition, including the silicon-to-aluminum ratio. SR Alkali-acid ratio B / A Slagging index R s and pollution index R f At least one of them; S12. Dataset Construction: Using coal quality full characteristic parameters, mechanism-derived features, and real-time operating parameters as input vectors, and coking risk level as output label, construct a training dataset; S13. Random Forest Model Training and Probability Output: Multiple decision trees are generated using the Bootstrap sampling strategy. Random feature subspace selection is used when splitting decision tree nodes. Feature selection is performed based on the Gini impurity minimization criterion to train a random forest model. The coal quality full characteristic parameters under the current working condition, the mechanism-derived features, and real-time operating parameters are used as input vectors and input into the random forest model to output the predicted probability of the random forest model. S14. Mechanism Constraint Fusion: The probability predicted by the random forest model is weighted and fused with the probability of mechanism constraints based on physical rules through the Bayesian posterior correction formula to output the final coking risk probability.

4. The method for multi-objective optimization and dynamic pulverized coal feeding control based on the full characteristics of coal quality according to claim 3, characterized in that: In step S11, the calculation formula for mechanism-derived features is as follows: silicon-aluminum ratio SR : Alkali-acid ratio B / A : Slagging Index R s : in, Total sulfur content on a dry basis; Pollution Index R f : In step S12, the coking risk level label is obtained using a semi-supervised labeling method, which combines indirect indicators such as sootblower operation frequency, abnormal increase in desuperheating water flow, and furnace heat absorption attenuation rate in historical operating data with manual records to label the level. In step S13, the Bootstrap strategy used in random forest training generates the training set for a single decision tree through random sampling with replacement, uses out-of-bag data for unbiased model validation, and sets the feature subspace size to the square root of the total number of features. The feature selection based on the Gini impurity minimization criterion specifically involves: In a randomly selected feature subspace, all possible thresholds for each feature are traversed, and the weighted Gini impurity corresponding to each candidate split is calculated. The feature and threshold combination that minimizes the weighted Gini impurity is selected as the optimal splitting scheme; where nodes... The formula for calculating the impurity of Gini is: in, For the node belonging to the first j The probability of coking-like risk; This represents the total number of coking risk categories; In step S14, the Bayesian posterior correction formula is: in, This represents the probability of final coking risk. Predict probabilities for the random forest model; Mechanism-constrained probability; The confidence weighting factor is based on the similarity between the current operating conditions and the distribution of historical data. The confidence weight factor Dynamically adjust based on the Mahalanobis distance between the current operating point and the training set center: When the operating condition is in a region with high density of historical data Approaching 1 The results are primarily based on random forest predictions. When the operating condition is in a region with sparse historical data or an extrapolated region Approaching 0 The prediction results are primarily determined by mechanistic constraints.

5. The method for multi-objective optimization and dynamic pulverized coal feeding control based on the full characteristics of coal quality according to claim 4, characterized in that: In step S1, the specific method for establishing the burnout and pollutant generation prediction model includes: Construct a hybrid prediction model architecture that targets NO x Generation amount, fly ash carbon content and SO x The generated data all adopt the "mechanism skeleton + data correction" model. The mechanism skeleton is constructed based on the principles of combustion dynamics, and the data correction part uses machine learning algorithms to perform nonlinear compensation on the residuals of the mechanism model. The NO x The mathematical form of the hybrid prediction model for the generated amount is: in, NO x Generate concentration prediction values; It is a thermal NO x Generation rate coefficient; For macroscopic activation energy; It is the universal gas constant; This refers to the flame temperature in the combustion core region. Oxygen concentration in the main combustion zone; To obtain the basic nitrogen content; For fuel nitrogen to NO x Conversion rate coefficient; This is a nonlinear residual correction term; It is a thermal NO x Generation mechanism term; NO is a fuel type x Generation mechanism term; This is a nonlinear residual correction term based on support vector regression, whose input variables include boiler load. , Burnout wind opening Variation in powder supply and powder stability index ; The mathematical form of the prediction model for the carbon content of fly ash is: in, The proportion of unburned carbon remaining in fly ash; For coal powder fineness; It is a dry, ash-free volatile matter; Oxygen content; The model coefficients are updated in real time using data from an online fly ash carbon content monitor and a recursive least squares algorithm to achieve model adaptation. The SO x The mathematical form of the hybrid prediction model for the generated amount is: in, For SO x Generate concentration prediction values; The total sulfur content of the coal received into the furnace; The sulfur conversion coefficient; For in-furnace desulfurization efficiency; This is the correction term function based on Gaussian process regression; For SO x The dominant mechanism of generation; This is a correction term based on Gaussian process regression, and its input features include bed temperature / furnace temperature. Excess air coefficient and effective calcium oxide content This correction term also outputs the confidence interval of the predicted value; The in-furnace desulfurization efficiency ,in, It is an empirical fitting index; and the effective calcium oxide content The effective content of alkaline oxides in coal ash is obtained by weighted calculation of ash component analysis data.

6. The method for multi-objective optimization and dynamic pulverized coal feeding control based on the full characteristics of coal quality according to claim 5, characterized in that: In step S2, the change in powder feeding amplitude Δ M Frequency of powder feeding variation f and powder stability index σ Defined quantitatively as follows: The variation range of powder feeding Δ M The calculation formula is: in, For sampling time window; The baseline powder feeding amount; M ( t () represents the amount of powder supplied at the current moment; for Historical powder distribution before the time; Δ M ( t () represents the change in powder supply at the current moment; The frequency of powder feeding variation f The calculation formula is: in, The number of times the direction of the powder feeding command changes within a unit time window; For statistical time windows; The powder feeding stability index σ The calculation formula is: in, n This represents the number of sampling points; M i For the first i The powder feeding value at each sampling point; This is the arithmetic mean of the amount of powder given during the statistical period.

7. The method for multi-objective optimization and dynamic pulverized coal feeding control based on the full characteristics of coal quality according to claim 6, characterized in that: In step S3, establishing a quantitative mapping relationship between the full characteristic parameters of coal, the quantitative indicators, and boiler combustion efficiency, pollutant emission concentration, and coking risk specifically includes: S31. In-depth mining of historical operational data: Extract historical operational data and establish the variation range Δ of powder distribution through association rule mining and partial least squares regression. M Frequency of powder feeding variation f and powder stability index σ Preliminary correlation with coking risk, pollutant emission concentration, and burnout rate; S32. Multiphysics numerical simulation: Computational fluid dynamics (CFD) numerical simulation is used to obtain the temperature field, component field and flow field distribution in the furnace. Combined with ASPEN PLUS process simulation, the combustion efficiency, pollutant generation and coking tendency are calculated under different coal quality characteristics and pulverized coal feeding operation parameters. S33. Construction of a Quantitative Mapping Relationship Library: This involves fusing historical operational data mining results with CFD and ASPEN PLUS simulation data. Through regression analysis and interpolation fitting, a quantitative mapping relationship library is established, using coal quality characteristic parameters and pulverized feed quantitative indicators as inputs, and boiler thermal efficiency and NO2 as inputs. x Emission concentration, SO x A standardized quantitative mapping relation library is formed by using a multivariate function relation library with emission concentration, fly ash carbon content, and coking growth rate as outputs; S34. Mapping Relationship Update: Based on the running data and simulation results, the quantitative mapping relationship library is corrected online at a set period. In step S33, the quantitative mapping relationship library includes environmental mapping relationships, economic mapping relationships, safety mapping relationships, and burnout mapping relationships: The environmental mapping relationship includes NO. x Emission concentration mapping relationship and SO x Emission concentration mapping relationship; The NO x Emission concentration mapping relationship based on the change in feed Δ M Dry, ash-free volatile matter And the burnout air opening degree are used as input variables, with NO x Emission concentration is the output variable; the NO x The mathematical form of the emission concentration mapping relationship is: in, Y 1 is NO x Emission concentration forecasts; For volatile matter of NO x The generated impact coefficient; b and k The coefficient representing the nonlinear influence of powder feeding fluctuations; f ( X 3 () is a function of the burnout air opening degree; The SO x Emission concentration mapping relationship based on total sulfur content of coal received into the furnace and the variation in powder supply Δ M For input variables, in SO x Emission concentration is the output variable; the SO x The mathematical form of the emission concentration mapping relationship is: in, Y 2 is SO x Emission concentration forecasts; This is the comprehensive emission factor; To adjust the SO based on the fluctuation of powder supply x The generated sensitivity coefficient; This is an index term representing the impact of powder disturbance. The economic mapping relationship is based on flue gas oxygen content. O 2. Smoke exhaust temperature and powder stability index σ The input variable is boiler thermal efficiency, and the output variable is boiler thermal efficiency; the mathematical form of the economic mapping relationship is: in, Y 3 represents the predicted value of boiler thermal efficiency; The baseline thermal efficiency; C 1 represents the penalty coefficient for oxygen deviation; O opt For optimal operating oxygen levels; C 2 represents the penalty coefficient for unstable powder distribution; C 3 represents the penalty coefficient for deviation of exhaust gas temperature; To achieve the optimal smoke exhaust temperature; The safety mapping relationship is based on the furnace outlet temperature. T ch Ash composition and alkali-acid ratio B / A and boiler load Load The input variable is the coke growth rate, and the output variable is the coke growth rate; the mathematical form of the safety mapping relationship is: in, Y 4 represents the predicted growth rate of coking. T soft This refers to the ash softening temperature. This is the temperature sensitivity coefficient; For reference load; The burnout mapping relationship is based on the fineness of pulverized coal. R 90 Fixed carbon content in coal The mathematical form of the burnout mapping relationship, with fuel residence time in the furnace as the input variable and fly ash carbon content as the output variable, is as follows: in, Y 5 represents the predicted carbon content of fly ash; kz This is the comprehensive impact coefficient; h This refers to the residence time of fuel inside the furnace. r The index is influenced by the length of stay; In step S34, the environmental protection mapping relationship is updated in real time, the economic mapping relationship is corrected daily, the safety mapping relationship is updated by shift, and the burnout mapping relationship is updated by hour.

8. The method for multi-objective optimization and dynamic pulverized coal feeding control based on the full characteristics of coal quality according to claim 7, characterized in that: Step S4 includes: S41. Construction of a Multi-Objective Optimization Model: Taking economy, environmental protection, and safety as optimization objectives, a multi-objective optimization function vector is constructed, whose mathematical form is defined as: in, For multi-objective optimization function vectors; The objective function is the economic performance. The objective function is environmental protection. The objective function is security. T Indicates transpose; The economic objective function Coal consumption for power supply Its value is jointly determined by the boiler thermal efficiency predicted in step S33 and the mechanical incomplete combustion loss rate calculated from the carbon content of fly ash. The calculation process is as follows: First, the carbon content of fly ash is predicted using the burnout mapping relationship in step S33, and the heat loss rate due to incomplete mechanical combustion is calculated. : in, is a physical constant representing the calorific value of carbon; The received ash content of the coal fed into the furnace; This refers to the fly ash proportion coefficient. The received basis lower calorific value of the coal fed into the furnace; Then, the baseline thermal efficiency is predicted using the economic mapping relationship in step S33, and the final boiler thermal efficiency is calculated. : Ultimately, coal consumption for power generation for: Where 123 is the unit conversion factor, a constant that converts thermal efficiency into standard coal consumption; For pipeline efficiency; This refers to the internal efficiency of the steam turbine. For mechanical efficiency; For generator efficiency; Plant power consumption rate; and ; The environmental protection objective function The comprehensive pollutant emissions are represented by the NO value in step S33. x Emission concentration predictions and SO x Weighted normalized form of predicted emission concentrations: in, NO x Emission weighting coefficient; For SO x Emission weighting coefficient; NO x Environmental regulations emission limits; For SO x Environmental regulations emission limits; The security objective function This is the coking risk index, whose value incorporates the final coking risk probability output from the coking tendency prediction model in step S1. The current predicted coking growth rate is mapped to the safety relationship in step S33. Y 4, that is: in, This refers to the static risk weighting coefficient. This is a dynamically increasing risk weighting coefficient; S42. Constraint Settings: Set boiler wall temperature constraints and quantitative index range constraints. The boiler wall temperature constraint is used to limit the furnace wall temperature from exceeding the safety limit. The quantitative index range constraint is used to limit the pulverized coal variation range Δ. M Frequency of powder feeding variation f and powder stability index σ Within the limits allowed by the equipment; S43. Multi-objective optimization solution: The NSGA-II multi-objective genetic algorithm is used to solve the Pareto optimal solution set of the multi-objective optimization function vector under the constraints. S44. Optimal Strategy Selection: Based on the power plant's operational strategy preferences, a weighted method is used to select the final solution from the Pareto optimal solution set, and the optimal feed rate variation Δ that adapts to the current coal quality is output. M Frequency of powder feeding variation f and powder stability index σ The dynamic control strategy is generated and sent to the system for execution.

9. The method for multi-objective optimization and dynamic pulverized coal feeding control based on the full characteristics of coal quality according to claim 8, characterized in that: In step S43, the NSGA-II algorithm includes the following steps: S431. Population initialization and real number encoding: The initial population is generated using real number encoding. Each individual represents a set of decision parameter vectors, namely, a combination of pollen feeding variation amplitude, pollen feeding variation frequency, and pollen feeding stability index. S432. Fitness Assessment: Substitute the decision parameter vector of each individual into the quantitative mapping relation library established in step S33, call the economic mapping relation and burnout mapping relation, and calculate the economic objective function value. And by invoking the environmental mapping relationship, calculate the environmental objective function value. The coking tendency prediction model established in step S1 is substituted into the model, and the coking growth rate predicted by the safety mapping relationship in step S33 is combined to calculate the safety objective function value. ; S433. Fast Non-Dominated Sort: Calculate the dominated count and dominant set of each individual, assign non-dominated individuals to the first front surface, iterate through the dominant set of each non-dominated individual and assign them to subsequent front surfaces in turn, until all individuals are stratified. S434. Crowding distance calculation: For individuals on the same frontal surface, calculate their crowding distance in the target space, and prioritize individuals with larger crowding distances to maintain population diversity; S435. Genetic operations: Select, crossover and mutation operations are performed using binary tournament selection, simulated binary crossover and polynomial mutation to generate a new generation of population; S436. Elite Preservation and Generation of New Generation: Merge the parent and offspring populations, re-perform fast non-dominated sorting and crowding distance calculation, and select the top N best individuals to generate a new generation population; S437. Iteration Termination and Solution Output: Repeat steps S432 to S436 until the preset number of iterations is reached, and output the Pareto optimal solution set.

10. The method for multi-objective optimization and dynamic pulverized coal feeding control based on the full characteristics of coal quality according to claim 9, characterized in that: The process following step S4 also includes: Step S5. Dynamic self-evolution update: Establish a three-level update mechanism of data layer - model layer - strategy layer; the data layer automatically updates the coal quality database and the operation dataset when it detects coal quality fluctuations or operational efficiency deviations exceeding the threshold; the model layer uses online machine learning algorithms to iteratively optimize the coking tendency prediction model and the burnout and pollutant generation prediction model; the strategy layer synchronously adjusts the parameters of the multi-objective optimization function vector according to the model update results.

Images