AI-driven coal blending method for stability and economic benefit of thermal power plant combustion

Abstract

The application discloses an AI-driven double-target intelligent coal blending method for combustion stability and economic benefit of a thermal power plant, and relates to the technical field of coal blending optimization of a thermal power plant. The method comprises the following steps: constructing a coal type combustion fingerprint feature database; inputting combustion fingerprint feature data of a candidate coal blending scheme into a boiler combustion risk prediction model based on a weighted support vector machine to obtain corresponding combustion risk probability values; taking a coal blending ratio as an optimization variable, constructing a double-target optimization function, and relaxing a preset constraint condition by using a penalty function, wherein the double-target optimization function comprises an economic function with the minimum fuel procurement cost as a target and a safety function with the minimum combustion risk probability value as a target; performing Pareto optimization on the economic function and the safety function based on a multi-objective evolutionary algorithm to output a Pareto optimal coal blending scheme set; and executing the coal blending scheme. In this way, the safety of boiler combustion and the economy of fuel procurement are synergistically optimized.

Description

Technical Field

[0001] This invention relates to the field of coal blending optimization technology for thermal power plants, and in particular to an AI-driven intelligent coal blending method with dual objectives of combustion stability and economic benefits for thermal power plants. Background Technology

[0002] Coal blending in thermal power plants is a core element for the safe and stable operation of boilers and the reduction of fuel costs; its optimization directly impacts the plant's safety and economic benefits. Currently, coal blending technology in thermal power plants still suffers from several fundamental shortcomings, making it difficult to meet the demands of high-quality development in the power industry. These shortcomings include:

[0003] (1) Insufficient accuracy of combustion risk prediction: Existing technologies mostly use standard SVM for boiler combustion risk prediction, but there is a serious class imbalance problem in the operating data of thermal power plants. Normal combustion samples account for more than 95%, while high-risk abnormal samples such as slagging and stable combustion failure are a minority. Standard SVM uses the same misclassification penalty for all samples, which easily leads to the problem of classification bias towards the majority class, resulting in low accuracy and high false negative rate in identifying high-risk abnormal samples, which cannot meet the safety management requirements of power plants.

[0004] (2) There is a timeliness bottleneck in coal quality data acquisition: Existing coal blending optimization relies on the intrinsic combustion characteristic parameters of coal obtained by thermogravimetric analysis (TGA), but TGA detection takes a long time and cannot track the characteristic changes of newly entered coal in real time, resulting in poor real-time model training and coal blending optimization, making it difficult to achieve dynamic coal blending.

[0005] (3) Coal blending prediction does not consider nonlinear effects: Traditional coal blending methods derive the combustion characteristics of blended coal based on the assumption of linear weighted average, without considering the nonlinear combustion effects after different coal types are blended, resulting in large deviations in the prediction of combustion characteristics, and a large gap between the actual implementation effect of the coal blending scheme and the theoretical prediction.

[0006] (4) Coal blending optimization is difficult to balance safety and economy: Existing coal blending optimization is mostly single-objective optimization, or although dual objectives are set, hard constraints are used. The non-dominated sorting genetic algorithm (NSGA-II) is prone to having no feasible solution due to the narrow feasible domain when searching for optimization. At the same time, there are extreme decision problems of emphasizing economy over safety or safety over economy, which cannot achieve synergistic optimization of the two.

[0007] (5) The model lacks a dynamic adaptation mechanism: Most of the existing coal blending models are static models and are not dynamically updated according to the actual operating data of the boiler, changes in coal type, and fluctuations in market prices. As the unit ages and the coal type is changed, the prediction accuracy and optimization effect of the model will continue to decline, resulting in poor engineering implementation.

[0008] While some existing research attempts to apply weighted SVM to unbalanced data classification and NSGA-II to coal blending optimization, none of these studies have made targeted improvements based on the actual operating conditions of coal blending in thermal power plants, thus failing to fundamentally solve the aforementioned problems. Therefore, developing an intelligent coal blending method for thermal power plants that balances combustion risk prediction accuracy, data real-time performance, synergistic optimization of safety and economy, and dynamic adaptation to engineering conditions has become an urgent need in this field. Summary of the Invention

[0009] To address the core deficiencies in existing coal blending technologies for thermal power plants, this invention provides an AI-driven intelligent coal blending method with dual objectives of combustion stability and economic benefits. It fundamentally solves the problem of low accuracy and high false negative rate in identifying a few high-risk anomalies such as slagging and combustion instability when the standard support vector machine (SVM) model faces imbalanced operating data categories. Simultaneously, it overcomes the limitations of the linear assumptions in traditional coal blending, achieving synergistic optimization of boiler combustion safety and fuel procurement economy.

[0010] In a first aspect, embodiments of the present invention provide an AI-driven intelligent coal blending method for thermal power plants with dual objectives of combustion stability and economic benefits, including:

[0011] Construct a coal type combustion fingerprint feature database that includes basic coal quality parameters, intrinsic combustion characteristic parameters, and safety boundary parameters;

[0012] The combustion fingerprint feature data corresponding to the candidate coal blending scheme is input into a pre-constructed boiler combustion risk prediction model based on weighted support vector machine to obtain the combustion risk probability value of the candidate coal blending scheme.

[0013] Using the coal blending ratio as the optimization variable, a dual-objective optimization function is constructed and a penalty function is used to relax the preset constraints. The dual-objective optimization function includes an economic function aimed at minimizing fuel procurement costs and a safety function aimed at minimizing the combustion risk probability value.

[0014] Pareto optimization is performed on the economic function and safety function based on a multi-objective evolutionary algorithm to output the Pareto optimal coal blending scheme set.

[0015] Based on the actual operating conditions of the thermal power unit, a coal blending scheme is selected and executed from the Pareto optimal coal blending scheme set.

[0016] As a preferred embodiment, a coal type combustion fingerprint feature database is constructed, including basic coal quality parameters, intrinsic combustion characteristic parameters, and safety boundary parameters, comprising:

[0017] Basic coal quality parameters, safety boundary parameters, and intrinsic combustion characteristic parameters obtained through thermogravimetric analysis were collected for single coal and mixed coal. The collected parameters were preprocessed to obtain a standardized set of characteristic parameters.

[0018] If there are coal types for which thermogravimetric analysis data has not been obtained, the combustion intrinsic characteristic parameters are predicted using a pre-trained regression-assisted model based on the basic coal quality parameters of that coal type, and the predicted values ​​are used as the combustion intrinsic characteristic parameters of that coal type.

[0019] The standardized feature parameter set and the predicted intrinsic combustion characteristic parameters are integrated to form a combustion fingerprint feature database.

[0020] As a preferred embodiment, the boiler combustion risk prediction model based on weighted support vector machine is constructed through the following steps:

[0021] Obtain a training sample set, in which high-risk abnormal combustion conditions of boilers are used as positive samples and normal stable combustion conditions are used as negative samples.

[0022] A corresponding weighting factor is set for the positive class samples and the negative class samples respectively, wherein the weighting factor is determined by the ratio of the number of negative class samples to the number of positive class samples;

[0023] Using the combustion fingerprint feature data of the coal blending scheme and the boiler operating boundary parameters as input features, and the combustion condition category as the output label, a boiler combustion risk prediction model based on weighted support vector machine is trained.

[0024] In a preferred embodiment, the weighting factor is determined by the ratio of the number of negative class samples to the number of positive class samples, including:

[0025] The ratio of the weighting factor for positive samples to the weighting factor for negative samples is determined based on the ratio of the number of negative samples to the number of positive samples.

[0026] The specific value of the weighting factor is determined by cross-validation within the range of 0 to 1, based on the ratio of the weighting factor of the positive class samples to the weighting factor of the negative class samples.

[0027] As a preferred embodiment, the economic objective function for:

[0028] ;

[0029] Security objective function for:

[0030] ;

[0031] in, : Coal blending scheme optimization variable vector : No. The blending ratio of different types of coal satisfies the non-negativity constraint. and normalization constraints ; The total quantity of coal types used in blending. For the first The unit quality purchase price of coal; Coal blending scheme The probability of high-risk anomalies, where the superscript T indicates that the row vector... Transpose to a column vector.

[0032] As a preferred embodiment, a penalty function is used to relax the preset constraints, including:

[0033] For coal blending schemes that violate preset constraints, calculate the amount of excess for each constraint and compare the amount of excess with a preset threshold.

[0034] If the excess exceeds a preset threshold, the coal blending scheme will be directly eliminated during the population evolution process;

[0035] If the overscalar quantity does not exceed the preset threshold, a penalty term is determined based on the overscalar quantity and the preset penalty coefficient, and the penalty term is included in the economic function, with the corrected economic function serving as the optimization objective.

[0036] As a preferred implementation, the economic function and safety function are optimized using a multi-objective evolutionary algorithm via Pareto optimization, outputting a set of Pareto optimal coal blending schemes, including:

[0037] Based on a multi-objective evolutionary algorithm, a population containing multiple coal blending schemes is generated during initialization. Each coal blending scheme consists of the blending ratio of each type of coal.

[0038] For each individual in the population, its fuel purchase cost and combustion risk probability are calculated according to the economic function and safety function, respectively. Individuals that violate the constraints are corrected based on the penalty function to obtain the fitness value of the individual.

[0039] The population individuals are non-dominated and ordered, Pareto ranks are assigned, and crowding is calculated for individuals in the same rank.

[0040] The offspring population is generated through selection, crossover, and mutation operations. The parent population and the offspring population are merged and then subjected to non-dominated sorting and crowding screening to generate a new generation population.

[0041] The process is iterated until the preset maximum number of iterations is reached or the Pareto front of the population tends to converge, and the Pareto optimal coal blending scheme set is output.

[0042] As a preferred embodiment, after selecting and executing a coal blending scheme from the Pareto optimal coal blending scheme set according to the actual operating conditions of the thermal power unit, the method further includes:

[0043] Boiler operation data is collected in real time and compared with combustion risk probability values ​​to generate incremental labeled samples;

[0044] When the cumulative number of incrementally labeled samples reaches a preset threshold, the incrementally labeled samples are merged with the training sample set to obtain the merged training sample set.

[0045] The boiler combustion risk prediction model based on weighted support vector machine is retrained using the merged training sample set to update and optimize the model parameters.

[0046] In a preferred embodiment, the method further includes:

[0047] When new coal types are added to the coal yard, coal market prices fluctuate beyond the preset range, or the operating conditions of thermal power units undergo significant changes, the optimization process based on a multi-objective evolutionary algorithm is restarted to generate an updated Pareto optimal coal blending scheme set.

[0048] As a preferred embodiment, the boiler combustion risk prediction model based on weighted support vector machine is a weighted C-support vector machine or a weighted C-support vector machine. - Support Vector Machine.

[0049] Secondly, embodiments of the present invention also provide an AI-driven intelligent coal blending system for thermal power plants with dual objectives of combustion stability and economic benefits, comprising:

[0050] The combustion fingerprint feature database construction module is used to construct a coal type combustion fingerprint feature database that includes basic coal quality parameters, intrinsic combustion characteristic parameters, and safety boundary parameters.

[0051] The combustion risk prediction module is used to input the combustion fingerprint feature data corresponding to the candidate coal blending scheme into a pre-constructed boiler combustion risk prediction model based on weighted support vector machine to obtain the combustion risk probability value of the candidate coal blending scheme.

[0052] A dual-objective optimization function construction module is used to construct a dual-objective optimization function with the coal blending ratio as the optimization variable and to relax the preset constraints using a penalty function. The dual-objective optimization function includes an economic function aimed at minimizing fuel procurement costs and a safety function aimed at minimizing the combustion risk probability value.

[0053] The optimal coal blending scheme set solution module is used to perform Pareto optimization on the economic function and safety function based on a multi-objective evolutionary algorithm, and output the Pareto optimal coal blending scheme set.

[0054] The coal blending scheme execution module is used to select and execute a coal blending scheme from the Pareto optimal coal blending scheme set according to the actual operating conditions of the thermal power unit.

[0055] Thirdly, embodiments of the present invention also provide an electronic device, the electronic device comprising:

[0056] One or more processors;

[0057] Storage device for storing one or more programs;

[0058] When the one or more programs are executed by the one or more processors, the one or more processors implement the AI-driven intelligent coal blending method for dual objectives of combustion stability and economic benefits in thermal power plants as described in any embodiment of the present invention.

[0059] Fourthly, embodiments of the present invention also provide a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the AI-driven intelligent coal blending method for combustion stability and economic benefits in thermal power plants as described in any embodiment of the present invention.

[0060] Compared with existing technologies, the present invention achieves the following beneficial effects:

[0061] (1) This invention improves the standard support vector machine by introducing a category-differentiated weighting factor, which effectively eliminates the classification bias problem caused by the imbalance of the categories of thermal power plant operating data and significantly reduces the false negative rate of high-risk abnormal samples such as slagging and stable combustion failure. Combined with Platt scaling, it realizes the probabilistic quantitative characterization of combustion risk, providing accurate safety quantitative basis for coal blending optimization.

[0062] (2) The present invention designs a combustion fingerprint soft measurement supplementary scheme, which uses easily obtainable basic coal quality parameters to predict combustion intrinsic characteristic parameters through regression auxiliary model, solves the timeliness bottleneck caused by the long acquisition cycle of thermogravimetric analysis data, takes into account the integrity of data and the real-time performance of coal blending optimization, and realizes rapid adaptation of new coal types entering the plant.

[0063] (3) This invention maps low-dimensional coal quality characteristics to high-dimensional feature space through kernel function, effectively fitting the nonlinear combustion effect after different coal types are blended, breaking through the limitations of the traditional coal blending method based on the linear weighted average assumption, and making the prediction results of blended coal combustion characteristics more consistent with the actual combustion conditions of the boiler.

[0064] (4) This invention constructs a dual-objective optimization system with a quantitative penalty function, which incorporates combustion safety and fuel economy into a unified optimization framework. It combines Pareto optimization with non-dominated sorting genetic algorithm-II, which avoids the optimization failure problem caused by the narrow feasible region and overcomes the extreme decision-making tendency of single-objective optimization. It minimizes fuel procurement costs while ensuring boiler combustion safety.

[0065] (5) This invention provides two optional models: weighted C-support vector machine and weighted ν-support vector machine, which take into account both prediction accuracy and engineering debugging convenience. At the same time, it designs a model dynamic update mechanism based on sliding time window, which can adapt to actual working condition changes such as unit aging, coal type change, and market fluctuations, and has good engineering implementation and cross-unit universality. Attached Figure Description

[0066] Other features, objects, and advantages of the invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings. The drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings:

[0067] Figure 1 This is a flowchart of the AI-driven intelligent coal blending method for combustion stability and economic benefits in thermal power plants provided in this embodiment of the invention.

[0068] Figure 2 This is a flowchart of the weighted support vector machine model construction provided in an embodiment of the present invention;

[0069] Figure 3 This is a flowchart of the dual-objective optimization function and penalty function processing provided in the embodiments of the present invention;

[0070] Figure 4 This is a Pareto optimization flowchart based on NSGA-II provided in an embodiment of the present invention;

[0071] Figure 5 This is a schematic diagram of the structure of the AI-driven intelligent coal blending system for thermal power plants with dual objectives of combustion stability and economic benefits, provided in an embodiment of the present invention.

[0072] Figure 6 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0073] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and not intended to limit it. Furthermore, it should be noted that, for ease of description, the accompanying drawings show only the parts relevant to the present invention, and not all of the structures.

[0074] Before discussing the exemplary embodiments in more detail, it should be mentioned that some exemplary embodiments are described as processes or methods depicted as flowcharts. Although the flowcharts describe operations (or steps) as being processed sequentially, many of these operations (or steps) may be performed in parallel, concurrently, or simultaneously. Furthermore, the order of the operations may be rearranged. The process may be terminated when its operation is completed, but may also have additional steps not included in the figures. The process may correspond to a method, function, procedure, subroutine, subroutine, etc.

[0075] Example 1

[0076] like Figure 1 The diagram shows a flowchart of an AI-driven intelligent coal blending method 100 for dual objectives of combustion stability and economic benefits in thermal power plants, as provided in Embodiment 1 of the present invention. The method 100 specifically includes the following steps:

[0077] S110: Construct a coal type combustion fingerprint feature database that includes basic coal quality parameters, intrinsic combustion characteristic parameters, and safety boundary parameters.

[0078] As a preferred embodiment, this step provides basic data support for model training and coal blending optimization. Data sources include incoming coal quality testing reports from thermal power plants, laboratory thermogravimetric analysis (TGA) reports, historical operating data from distributed control systems (DCS) of boilers, boiler performance test reports, and basic ledgers of coal types in coal yard inventory. The implementation includes three parts:

[0079] (1) Feature parameter acquisition and definition

[0080] Collect and input full-dimensional characteristic parameters of single coal and mixed coal, defining them as coal-specific "combustion fingerprints," which specifically include three types of parameters:

[0081] Basic coal quality parameters: industrial analysis (moisture, ash, volatile matter, fixed carbon), elemental analysis (carbon, hydrogen, oxygen, nitrogen, total sulfur), calorific value;

[0082] Intrinsic combustion characteristic parameters: Ignition temperature obtained through TGA Burnout temperature Maximum combustion rate and corresponding temperature, and comprehensive combustion characteristic index S;

[0083] Safety boundary parameters: ash melting characteristic temperature (deformation temperature DT, softening temperature ST, flow temperature FT), ash viscosity-temperature characteristic curve, and grindability index.

[0084] (2) Data preprocessing: Perform data preprocessing operations on the collected parameters, including removing outliers, filling in missing values, and standardizing all feature parameters.

[0085] (3) Supplementary solution for combustion fingerprint soft measurement

[0086] To address the issue of missing intrinsic combustion characteristic parameters for some coal types due to the lack of thermogravimetric analysis (TGA), this method proposes a soft measurement supplementary scheme: for newly arrived coal types for which TGA cannot be completed in a timely manner, the intrinsic combustion characteristic parameters of the coal type are predicted using readily available basic industrial analysis and elemental analysis data through a pre-trained regression-assisted model. The predicted values ​​are then added to the combustion fingerprint database to ensure data integrity and real-time coal blending optimization.

[0087] S120: Input the combustion fingerprint feature data corresponding to the candidate coal blending scheme into the pre-constructed boiler combustion risk prediction model based on weighted support vector machine to obtain the combustion risk probability value of the candidate coal blending scheme.

[0088] In a preferred embodiment, this step constructs a mapping relationship between coal blending schemes and combustion risks using an improved weighted support vector machine (SVM). This fundamentally addresses the model classification bias problem caused by severe imbalance in the categories of thermal power plant operating data, achieving a probabilistic and quantitative characterization of combustion risks. The specific implementation includes the following steps S121 to S124:

[0089] Step S121: Model task definition and sample set construction

[0090] The weighted SVM proposed in this invention is a binary classification model, primarily used to predict the probability of high-risk combustion anomalies in boilers corresponding to coal blending schemes. The training samples are divided into positive and negative classes, specifically defined as follows:

[0091] Positive class samples: labels This represents high-risk abnormal operating conditions in boiler combustion, including scenarios such as slagging, significant fluctuations in furnace negative pressure, delayed ignition, and fly ash carbon content exceeding the threshold. The sample size is denoted as [sample number]. ;

[0092] Negative class samples: labels This represents the normal and stable combustion conditions of the boiler, and the sample size is denoted as [sample number]. .

[0093] The total training sample set is denoted as The total number of samples Input feature vector (N is the feature dimension) It is composed of the coal blending combustion fingerprint parameters corresponding to the coal blending scheme and the boiler operating boundary parameters (load, air distribution method, furnace temperature).

[0094] Step S122: Model Construction and Derivation of Weighted C-SVM in this Invention

[0095] This step constructs a mapping relationship between coal blending schemes and combustion risks using an improved weighted support vector machine (SVM). The core solution addresses the classification bias problem caused by severe imbalance in the data categories of thermal power plant operating conditions, achieving a probabilistic and quantitative representation of combustion risks. Specifically, it includes the following steps S1221 to S1224:

[0096] Step S1221: The original optimization problem of the weighted C-SVM of this invention is defined as follows:

[0097] ;

[0098] Constraints:

[0099] ;

[0100] Where w is the classification hyperplane normal vector, used to define the classification decision boundary direction of the weighted C-SVM in the high-dimensional feature space; b is the classification hyperplane bias term, used to adjust the position of the weighted C-SVM classification decision boundary, which together with the normal vector w forms the complete classification hyperplane. : Nonlinear mapping function, used to map the feature vector of low-dimensional coal blending scheme to a high-dimensional feature space, so as to fit the nonlinear relationship in the prediction of coal combustion risk; : The slack variable of the i-th training sample, representing the degree of misclassification of the sample, used to measure the degree to which the sample deviates from the classification hyperplane constraint; C: Global penalty parameter, one of the core hyperparameters of weighted C-SVM, used to balance model complexity (by... (Characteristics) and the overall severity of penalties for misclassification; : The weighting factor of the i-th training sample, used to assign differentiated misclassification penalty weights to samples of different categories; : The category label of the i-th training sample, with a value of +1 (representing a high-risk abnormal boiler combustion condition) or -1 (representing a normal and stable boiler combustion condition). : The input feature vector of the i-th training sample.

[0101] Step S1222: Set weighting factors for positive and negative samples respectively, and eliminate the impact of class imbalance through differentiated weights. Specific rules:

[0102] (1) The weighting factor of the positive class samples (a few outliers) is denoted as The weighting factor for the negative class samples (most normal classes) is denoted as... ;

[0103] (2) The core ratio relationship of the weighting factors satisfies Even though the weighting factor for minority class samples is inversely proportional to the number of samples, it assigns them a higher penalty weight for misclassification.

[0104] (3) Under the premise of maintaining this ratio, and The specific values ​​were obtained through 5-fold cross-validation. Optimize within the range to adapt to the distribution characteristics of operating data of different thermal power units.

[0105] Step S1223: Transformation and solution of the dual problem: By introducing Lagrange multipliers , The original optimization problem is transformed into a dual problem, simplifying the solution process:

[0106] (1) Construction of the Lagrange function: Lagrange function Defined as:

[0107] ;

[0108] in, The Lagrange multiplier corresponding to the i-th training sample is non-negative and is used to handle the classification hyperplane constraint. Its optimal solution determines the selection of support vectors. The Lagrange multiplier corresponding to the i-th training sample is non-negative and is specifically used to handle slack variables. Constraints;

[0109] (2) Derivation of optimality conditions: Based on the optimality conditions of saddle points, the following are derived: , , Taking the partial derivative and setting it to 0, we get:

[0110] ;

[0111] in, Lagrange function Setting the first-order partial derivative of the normal vector w of the classification hyperplane to 0 is one of the core conditions for solving the saddle point in the optimization problem.

[0112] Lagrange function Setting the first-order partial derivative of the bias term b on the classification hyperplane to 0 allows us to derive the optimality constraints associated with the bias term.

[0113] Lagrange function Relaxing variables for the i-th sample Setting the first-order partial derivative of the factor to 0 allows us to establish the quantitative relationship between the penalty parameter, the weighting factor, and the Lagrange multiplier.

[0114] (3) Dual optimization problem and solution: Substitute the above optimality conditions into the Lagrangian function and eliminate , , This leads to the dual optimization problem:

[0115] ;

[0116] Constraints:

[0117] ;

[0118] in, Dual optimization objective function, elimination , , The Lagrange multipliers obtained later The objective function for optimizing variables; α: a vector consisting of the Lagrange multipliers corresponding to all training samples, i.e. , is the core optimization variable of the dual optimization problem; j: the index identifier of the training sample, used to traverse all training samples, and is related to the index Together they form a double summation, representing the pairwise correlation between samples; The kernel function is a core component of the weighted C-SVM of this invention. It calculates the kernel function in the low-dimensional feature space. and The kernel value indirectly enables inner product operations in high-dimensional feature spaces. This avoids computational complexity in high-dimensional spaces. The input feature vector of the j-th training sample consists of the coal combustion fingerprint parameters and the boiler operating boundary parameters. : The class label of the j-th training sample.

[0119] Step S1224: Decision Function and Risk Quantification

[0120] Based on the above optimal solution, the decision function of the weighted C-SVM of this invention is constructed, and the quantitative characterization of combustion risk is realized:

[0121] (1) The classification decision function is defined as follows:

[0122] ;

[0123] in, For the support vector set (satisfying) (samples) For the optimal bias term, pass through normal support vectors (satisfying...) The calculation is as follows:

[0124] ;

[0125] in, The number of normal support vectors, when When, it is determined to be a high-risk abnormal working condition; when When this condition is met, it is considered a normal operating condition.

[0126] in, The classification decision function output of the weighted C-SVM of this invention is used to determine the boiler operating condition type corresponding to the coal blending scheme to be predicted, with a value of +1 (high-risk abnormal operating condition) or -1 (normal stable operating condition). The sign function is the core activation function of the weighted C-SVM decision function in this invention. It is used to output the classification result based on the positive or negative value of the expression in parentheses. When the expression is greater than 0, it outputs +1, and when it is less than 0, it outputs -1. : No. The optimal Lagrange multiplier for each training sample is obtained by solving the dual problem using the Sequence Minimum Optimization (SMO) algorithm. Its non-zero value corresponds to the support vector, which determines the contribution weight of the sample to the classification hyperplane. Kernel function operation, used to calculate the kernel vector in the support vector set. One sample and the sample to be predicted The kernel value indirectly enables inner product operations in high-dimensional feature spaces, thus avoiding the curse of dimensionality. : Optimal bias term, calculated from the normal support vector set, used to adjust the position of the classification decision boundary to ensure classification accuracy; x: Input feature vector of the coal blending scheme to be predicted; The number of normal support vectors, i.e., those that satisfy... The total number of samples, used for the optimal bias term. The average calculation; Normal support vector set refers to the set of support vectors in the weighted C-SVM of this invention that satisfies... The sample set is used to eliminate the interference of boundary support vectors on the calculation of bias terms; The optimal Lagrange multiplier for the j-th sample in the support vector set SV, and... Homogeneity is used to calculate the weighted sum of kernel values ​​corresponding to normal support vectors; The kernel function operation is used to calculate the kernel value between the j-th sample and the i-th normal support vector in the support vector set SV, and it is the optimal bias term. The core term of the calculation.

[0127] (2) Risk probability quantification

[0128] A classification decision function is constructed based on the optimal solution. The decision function outputs the signed distance from the sample to the classification hyperplane. Using Platt scaling method to Mapped to High-risk anomaly occurrence probability in interval The slope and intercept parameters of the Platt scaling are obtained by maximizing the log-likelihood estimate on the validation set, ensuring that the mapped values ​​are consistent with the actual values. The risk distribution characteristics are consistent with actual working conditions.

[0129] Step S123: Weighted method of the present invention - Optional model building for SVM

[0130] To address the issue of the unintuitive selection of parameter C in weighted C-SVM, this invention simultaneously proposes a weighted... -SVM, as an optional implementation, only optimizes the parameter form. The model derivation, dual solution, and risk quantification methods are exactly the same as weighted C-SVM, making it more suitable for on-site engineering debugging. The specific implementation includes the following:

[0131] (1) Primitive optimization problem

[0132] Weighted The original optimization problem of SVM is defined as follows:

[0133] ;

[0134] Constraints:

[0135] , ;

[0136] in, ρ is an interpretability parameter used to limit the maximum proportion of boundary support vectors and the minimum proportion of total support vectors; ρ: weighting parameter of this invention. The margin parameter in SVM defines the size of the margin between the classification hyperplane and the samples. - One of the core parameters of the SVM model.

[0137] (2) Dual optimization problem: Through Lagrange dual transformation, the weighted optimization problem is obtained. - The dual optimization problem of SVM:

[0138] ;

[0139] Constraints:

[0140] ;

[0141] Among them, constraints Derived from the KKT optimality condition, and consistent with the standard The constraint derivation logic of -SVM is completely consistent.

[0142] (3) Decision-making and quantification: weighted average The decision function construction, symbolic distance calculation, and Platt scaling of risk probabilities in the -SVM are exactly the same as those in the weighted C-SVM described above, and will not be repeated here. This model uses parameters... Replacing C is more in line with the debugging habits on the engineering site.

[0143] Step S124: Model Training and Validation

[0144] (1) Dataset partitioning: The labeled sample set is divided into... The dataset is divided into training, validation and test sets to ensure that the ratio of positive and negative samples in each dataset is consistent with the original dataset.

[0145] (2) Hyperparameter optimization: The C, σ, and σ of the weighted C-SVM are optimized by combining 5-fold cross-validation with grid search algorithm. or weighted v-SVM , The core optimization metric is the F1 score of the minority class (high-risk abnormal samples).

[0146] (3) Model validation: The model performance is validated using a test set. The core evaluation indicators include minority class F1 score, AUC value, and false negative rate, to ensure that the false negative rate of high-risk combustion events is minimized while maintaining the overall classification accuracy.

[0147] S130: Using the coal blending ratio as the optimization variable, a dual-objective optimization function is constructed and a penalty function is used to relax the preset constraints. The dual-objective optimization function includes an economic function aimed at minimizing fuel procurement costs and a safety function aimed at minimizing the combustion risk probability value.

[0148] In a preferred embodiment, this step uses the coal blending ratio as the optimization variable to construct a dual-objective optimization system that balances safety and economy, and introduces a penalty function method to achieve constraint relaxation, avoiding an excessively narrow feasible region for optimization. Specifically, it includes steps S131 to S133:

[0149] Step S131: Optimize variable definitions

[0150] The optimization variable is the mass percentage of each coal type involved in blending, denoted as... Where n is the quantity of coal types involved in blending. The blending ratio of the i-th type of coal satisfies and The superscript T indicates that the row vector Transpose to a column vector.

[0151] Step S132: Construction of the dual-objective optimization function

[0152] (1) Economic objective function: The objective is to minimize the procurement cost per unit mass of mixed coal. The function expression is as follows:

[0153] ;

[0154] In the formula, Let be the unit quality purchase price of the i-th type of coal;

[0155] (2) Safety objective function: The objective is to minimize the combustion risk probability output by the weighted SVM of this invention. The function expression is as follows:

[0156] ;

[0157] In the formula, Coal blending scheme The probability of high-risk anomalies is obtained by weighted SVM combined with Platt scaling.

[0158] Step S133: Design of Constraints and Penalty Functions

[0159] (1) Boundary definition: Five types of constraints that the coal blending scheme must meet:

[0160] ① Coal quality constraints: The received basis lower heating value, volatile matter, total sulfur content and ash content of the blended coal must meet the upper and lower limits of the coal quality requirements for boiler design;

[0161] ② Environmental constraints: Sulfur dioxide emissions after coal-fired combustion nitrogen oxides The concentration of dust emissions must comply with national emission standards;

[0162] ③ Safety constraints: softening temperature of blended coal ignition temperature Burnout temperature The boiler must meet the safety requirements for stable combustion and prevention of slagging.

[0163] ④ Inventory constraints: The blending ratio of each type of coal shall not exceed the available quantity of the existing coal yard inventory;

[0164] ⑤ Operational constraints: The grindability index of the blended coal must meet the safe operation requirements of the pulverizing system.

[0165] (2) Specific form of penalty function: The penalty function method is used to integrate the constraints into the objective function. To achieve constraint relaxation, the specific rules are as follows:

[0166] ① Handling minor violations: For coal blending schemes that slightly violate constraints, a penalty value is calculated based on the severity of the violation and incorporated into the objective function; taking excessive total sulfur content as an example, the compliance upper limit is set as follows: The scheme is divided into total sulfur. ,like Then the overscalar The penalty item is ( The penalty coefficient is adjusted according to the consequences of violations (coal quality indicators are taken as [10, 100], environmental protection indicators are taken as [100, 500]). The corrected objective function is: ;

[0167] ② Handling of serious violations: Coal blending schemes that seriously violate the constraints will be directly eliminated in the subsequent optimization process to ensure the basic feasibility of the coal blending schemes.

[0168] S140: Based on the multi-objective evolutionary algorithm, Pareto optimization is performed on the economic function and safety function to output the Pareto optimal coal blending scheme set.

[0169] This step encapsulates the combustion fingerprint feature database, weighted support vector machine model, bi-objective optimization function, and penalty function into a non-dominated sorting genetic algorithm-II. It iterative optimization outputs the Pareto optimal coal blending scheme set. The specific process is as follows: Figure 4 As shown:

[0170] (1) Step S141: Population initialization

[0171] The population size is set to 100, the number of iterations to 200, the crossover probability to 0.9, and the mutation probability to 0.1. An initial population is randomly generated, with each individual representing a set of conditions. and The coal blending ratio scheme.

[0172] (2) Step S142: Fitness calculation

[0173] The coal blending scheme corresponding to each individual is input into the weighted SVM of this invention to calculate the combustion risk value. Combined with the penalty function rule in step S33, calculate the corrected fuel cost. The fitness values ​​of the two objectives are obtained. , ).

[0174] (3) Step S143: Non-dominated sorting and crowding calculation

[0175] Individuals in the population are non-dominated and ranked according to Pareto order. Crowding is calculated for individuals of the same order to measure the dispersion of individuals in the solution space. Elite individuals with high order and high crowding are preferentially retained.

[0176] (4) Step S144: Genetic operation

[0177] The offspring population is generated through selection, crossover, and mutation operations. The parent population is then merged with the offspring population, and non-dominated sorting and crowding screening are performed again to generate a new generation population.

[0178] (5) Step S145: Iteration terminated

[0179] The iteration terminates when the preset maximum number of iterations is reached, or when the Pareto front of the population converges, and the final Pareto optimal solution set is output. Each individual in this solution set represents an optimal coal blending scheme that satisfies all constraints and is mutually exclusive in terms of safety and economy.

[0180] S150: Select and execute a coal blending scheme from the Pareto optimal coal blending scheme set according to the actual operating conditions of the thermal power unit.

[0181] As a preferred embodiment, this step enables the implementation of the coal blending scheme and continuous optimization of the model, ensuring the dynamic adaptability of the method, specifically including:

[0182] The power plant operation and fuel management personnel, based on the current unit load plan, equipment health status, coal yard inventory, and coal market price fluctuations, select the best coal blending scheme suitable for the current operating conditions from the Pareto optimal solution set and issue it to the fuel operation department for execution.

[0183] During the dynamic update of the coal blending scheme, the actual operating data of the boiler is collected in real time through the DCS. The predicted combustion risk is compared with the actual results to generate incremental labeled samples. The incremental labeled samples are merged with the historical training sample set to obtain the merged training sample set. The weighted SVM model of this invention is updated by using a sliding time window or periodic full training, and outdated historical operating condition data is removed to ensure that the model always reflects the recent real combustion characteristics of the boiler.

[0184] When new coal types are added to the coal yard, coal prices fluctuate significantly, or the operating conditions of the generating units undergo major changes, the system automatically restarts the NSGA-II optimization process. By combining the updated combustion fingerprint database and the weighted SVM model, a new Pareto optimal solution set is generated, enabling dynamic adaptive adjustment of the coal blending strategy.

[0185] As a preferred embodiment, to verify the effectiveness of the method of the present invention, an experiment was conducted based on the actual operating data of a 300MW thermal power unit.

[0186] (1) Experimental scenario: Based on actual operating data of thermal power plants, the performance of different support vector machine (SVM) variants in the prediction task of high-risk anomalies (slagging, stable combustion failure, etc.) of boiler combustion was verified. The focus was on comparing the ability of each model to identify minority anomaly samples in the class imbalance scenario (normal samples accounted for 95.2% and high-risk anomaly samples accounted for 4.8%).

[0187] (2) Data set: 12,000 samples of operating data of a 300MW thermal power unit for 6 months were selected. The feature dimension is 28 (covering coal combustion fingerprint and boiler operating parameters). Stratified sampling was used to divide the data into training set (8,400 samples) and test set (3,600 samples) in a 7:3 ratio.

[0188] (3) Experimental setup:

[0189] Kernel function: All models use a Gaussian radial basis function (RBF).

[0190] Basic SVM: using default hyperparameters (penalty parameter C=1, kernel bandwidth) No cross-validation optimization is performed;

[0191] C-SVM, -SVM, Weighted C-SVM, Weighted -SVM: Hyperparameters are optimized using 5-fold cross-validation combined with grid search (C-SVM) , -SVM Kernel function bandwidth ;

[0192] Weighting factor settings: Weighted C-SVM / Weighted - The weighting factor of SVM is first determined according to the inverse ratio of the minority class sample size to the majority class sample size. Then, a small-scale grid search is performed around this ratio (18 to 22 times) to finally determine the optimal weighting factor;

[0193] Key performance indicators: Minority class F1 score (balancing precision and recall), AUC (overall classification performance), and false negative rate (FNR, the proportion of high-risk outliers that were not identified).

[0194] Table 1

[0195]

[0196] The experimental data clearly show that the weighting mechanism is the core of improving the ability to identify high-risk anomalies in imbalanced data:

[0197] The basic SVM uses default hyperparameters and has no class weighting design, resulting in an F1 score of only 0.62 for minority class outliers and a false negative rate as high as 28.5%, which cannot meet the safety management requirements of thermal power plants.

[0198] C-SVM and - After optimizing the hyperparameters through 5-fold cross-validation, SVM's performance is improved compared to the basic SVM (F1 score increases by 0.09 and 0.11 respectively, and false negative rate decreases by 9.3 and 10.9 percentage points respectively), but the false negative rate is still over 17%, making it difficult to meet the core engineering requirement of "low false negative rate".

[0199] In summary, the weighted C-SVM and weighted... -SVM, through a weighting factor setting of "inverse proportion of sample size + small-range optimization", significantly optimizes the recognition performance of minority class samples: compared with the optimized C-SVM, the weighted C-SVM improves the minority class F1 score by 18 percentage points and reduces the false negative rate by 13.4 percentage points; compared with the basic SVM with default parameters, the weighted C-SVM improves the minority class F1 score by 27 percentage points and reduces the false negative rate by 22.7 percentage points.

[0200] Weighted -SVM performs slightly better than weighted C-SVM (F1 score is 0.02 higher, and false negative rate is 1.5 percentage points lower), which is attributed to the more intuitive physical meaning of the ν parameter (directly related to the proportion of support vectors), higher efficiency in hyperparameter tuning, and better fit for the practical needs of engineering sites.

[0201] The above results verify the effectiveness of the weighted SVM model proposed in this invention in the task of predicting combustion risks in thermal power plants. It can ensure the overall classification accuracy (AUC exceeding 0.94) and minimize the risk of missed reporting of high-risk abnormal events, which is fully compatible with the "safety first" management principle of thermal power plants.

[0202] Based on the above embodiments, the core beneficial effects of the present invention are as follows:

[0203] (1) This invention improves the standard support vector machine by introducing a category-differentiated weighting factor, which effectively eliminates the classification bias problem caused by the imbalance of the categories of thermal power plant operating data and significantly reduces the false negative rate of high-risk abnormal samples such as slagging and stable combustion failure. Combined with Platt scaling, it realizes the probabilistic quantitative characterization of combustion risk, providing accurate safety quantitative basis for coal blending optimization.

[0204] (2) The present invention designs a combustion fingerprint soft measurement supplementary scheme, which uses easily obtainable basic coal quality parameters to predict combustion intrinsic characteristic parameters through regression auxiliary model, solves the timeliness bottleneck caused by the long acquisition cycle of thermogravimetric analysis data, takes into account the integrity of data and the real-time performance of coal blending optimization, and realizes rapid adaptation of new coal types entering the plant.

[0205] (3) This invention maps low-dimensional coal quality characteristics to high-dimensional feature space through kernel function, effectively fitting the nonlinear combustion effect after different coal types are blended, breaking through the limitations of the traditional coal blending method based on the linear weighted average assumption, and making the prediction results of blended coal combustion characteristics more consistent with the actual combustion conditions of the boiler.

[0206] (4) This invention constructs a dual-objective optimization system with a quantitative penalty function, which incorporates combustion safety and fuel economy into a unified optimization framework. It combines Pareto optimization with non-dominated sorting genetic algorithm-II, which avoids the optimization failure problem caused by the narrow feasible region and overcomes the extreme decision-making tendency of single-objective optimization. It minimizes fuel procurement costs while ensuring boiler combustion safety.

[0207] (5) This invention provides two optional models: weighted C-support vector machine and weighted ν-support vector machine, which take into account both prediction accuracy and engineering debugging convenience. At the same time, it designs a model dynamic update mechanism based on sliding time window, which can adapt to actual working condition changes such as unit aging, coal type change, and market fluctuations, and has good engineering implementation and cross-unit universality.

[0208] Example 2

[0209] Figure 5 This is a schematic diagram of the structure of the AI-driven intelligent coal blending system 500 for thermal power plants, which aims to achieve both combustion stability and economic benefits, provided in Embodiment 5 of the present invention. Figure 5 As shown, the system includes:

[0210] The combustion fingerprint feature database construction module 510 is used to construct a coal type combustion fingerprint feature database containing basic coal quality parameters, intrinsic combustion characteristic parameters, and safety boundary parameters.

[0211] Combustion risk prediction module 520 is used to input the combustion fingerprint feature data corresponding to the candidate coal blending scheme into a pre-constructed boiler combustion risk prediction model based on weighted support vector machine to obtain the combustion risk probability value of the candidate coal blending scheme.

[0212] The dual-objective optimization function construction module 530 is used to construct a dual-objective optimization function with the coal blending ratio as the optimization variable and to relax the preset constraints using a penalty function. The dual-objective optimization function includes an economic function aimed at minimizing fuel procurement costs and a safety function aimed at minimizing the combustion risk probability value.

[0213] The optimal coal blending scheme set solution module 540 is used to perform Pareto optimization on the economic function and safety function based on a multi-objective evolutionary algorithm, and output the Pareto optimal coal blending scheme set.

[0214] The coal blending scheme execution module 550 is used to select and execute a coal blending scheme from the Pareto optimal coal blending scheme set according to the actual operating conditions of the thermal power unit.

[0215] The AI-driven intelligent coal blending method for thermal power plants with dual objectives of combustion stability and economic benefits provided in the embodiments of the present invention can execute the AI-driven intelligent coal blending method for thermal power plants with dual objectives of combustion stability and economic benefits provided in any of the embodiments of the present invention. It has the corresponding functions and beneficial effects of executing the AI-driven intelligent coal blending method for thermal power plants with dual objectives of combustion stability and economic benefits. For detailed process, please refer to the relevant operations of the AI-driven intelligent coal blending method for thermal power plants with dual objectives of combustion stability and economic benefits in the foregoing embodiments.

[0216] Example 3

[0217] Figure 6 This is a schematic diagram of the structure of an electronic device provided in Embodiment Six of the present invention. The electronic device 10 is intended to represent various forms of digital computers, and may also represent various forms of mobile devices. The components shown herein, their connections and relationships, and their functions are merely examples and are not intended to limit the implementation of the invention described and / or claimed herein.

[0218] like Figure 6 As shown, the electronic device 10 includes at least one processor 11 and a memory, such as a read-only memory (ROM) 12 or a random access memory (RAM) 13, communicatively connected to the at least one processor 11. The memory stores computer programs executable by the at least one processor. The processor 11 can perform various appropriate actions and processes based on the computer program stored in the ROM 12 or loaded into the RAM 13 from storage unit 18. The RAM 13 can also store various programs and data required for the operation of the electronic device 10. The processor 11, ROM 12, and RAM 13 are interconnected via a bus 14. An input / output (I / O) interface 15 is also connected to the bus 14.

[0219] Multiple components in electronic device 10 are connected to I / O interface 15, including: input unit 16, such as keyboard, mouse, etc.; output unit 17, such as various types of displays, speakers, etc.; storage unit 18, such as disk, optical disk, etc.; and communication unit 19, such as network card, modem, wireless transceiver, etc. Communication unit 19 allows electronic device 10 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0220] Processor 11 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of processor 11 include, but are not limited to, central processing unit (CPU), graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various processors running machine learning model algorithms, digital signal processors (DSPs), and any suitable processor, controller, microcontroller, etc. Processor 11 implements the AI-driven dual-objective intelligent coal blending method for combustion stability and economic benefits in thermal power plants described above.

[0221] It should be understood that the various forms of processes shown above can be used, with steps reordered, added, or deleted. For example, the steps described in this invention can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution of this invention can be achieved, and this is not limited herein.

[0222] The above embodiments are merely illustrative examples and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. An AI-driven intelligent coal blending method for thermal power plants with dual objectives of combustion stability and economic benefits, characterized in that... include: Construct a coal type combustion fingerprint feature database that includes basic coal quality parameters, intrinsic combustion characteristic parameters, and safety boundary parameters; The combustion fingerprint feature data corresponding to the candidate coal blending scheme is input into a pre-constructed boiler combustion risk prediction model based on weighted support vector machine to obtain the combustion risk probability value of the candidate coal blending scheme. Using the coal blending ratio as the optimization variable, a dual-objective optimization function is constructed and a penalty function is used to relax the preset constraints. The dual-objective optimization function includes an economic function aimed at minimizing fuel procurement costs and a safety function aimed at minimizing the combustion risk probability value. Pareto optimization is performed on the economic function and safety function based on a multi-objective evolutionary algorithm to output the Pareto optimal coal blending scheme set. Based on the actual operating conditions of the thermal power unit, a coal blending scheme is selected and executed from the Pareto optimal coal blending scheme set.

2. The method according to claim 1, characterized in that, A coal combustion fingerprint feature database was constructed, comprising basic coal quality parameters, intrinsic combustion characteristic parameters, and safety boundary parameters, including: Basic coal quality parameters, safety boundary parameters, and intrinsic combustion characteristic parameters obtained through thermogravimetric analysis were collected for single coal and mixed coal. The collected parameters were preprocessed to obtain a standardized set of characteristic parameters. If there are coal types for which thermogravimetric analysis data has not been obtained, the combustion intrinsic characteristic parameters are predicted using a pre-trained regression-assisted model based on the basic coal quality parameters of that coal type, and the predicted values ​​are used as the combustion intrinsic characteristic parameters of that coal type. The standardized feature parameter set and the predicted intrinsic combustion characteristic parameters are integrated to form a combustion fingerprint feature database.

3. The method according to claim 1, characterized in that, The boiler combustion risk prediction model based on weighted support vector machine is constructed through the following steps: Obtain a training sample set, in which high-risk abnormal combustion conditions of boilers are used as positive samples and normal stable combustion conditions are used as negative samples. A corresponding weighting factor is set for the positive class samples and the negative class samples respectively, wherein the weighting factor is determined by the ratio of the number of negative class samples to the number of positive class samples; Using the combustion fingerprint feature data of the coal blending scheme and the boiler operating boundary parameters as input features, and the combustion condition category as the output label, a boiler combustion risk prediction model based on weighted support vector machine is trained.

4. The method according to claim 3, characterized in that, The weighting factor is determined by the ratio of the number of negative class samples to the number of positive class samples, including: The ratio of the weighting factor for positive samples to the weighting factor for negative samples is determined based on the ratio of the number of negative samples to the number of positive samples. The specific value of the weighting factor is determined by cross-validation within the range of 0 to 1, based on the ratio of the weighting factor of the positive class samples to the weighting factor of the negative class samples.

5. The method according to claim 1, characterized in that, The economic function for: ； Security functions for: ； in, : Coal blending scheme optimization variable vector : No. The blending ratio of different types of coal satisfies the non-negativity constraint. and normalization constraints ; The total quantity of coal types used in blending. For the first The unit quality purchase price of coal; Coal blending scheme The probability of high-risk anomalies; the superscript T indicates that the row vector Transpose to a column vector.

6. The method according to claim 5, characterized in that, The pre-defined constraints are relaxed using a penalty function, including: For coal blending schemes that violate preset constraints, calculate the amount of excess for each constraint and compare the amount of excess with a preset threshold. If the excess exceeds a preset threshold, the coal blending scheme will be directly eliminated during the population evolution process; If the overscalar quantity does not exceed the preset threshold, a penalty term is determined based on the overscalar quantity and the preset penalty coefficient, and the penalty term is included in the economic function, with the corrected economic function serving as the optimization objective.

7. The method according to claim 1, characterized in that, The Pareto optimization of the economic and safety functions is performed based on a multi-objective evolutionary algorithm, outputting a set of Pareto optimal coal blending schemes, including: Based on a multi-objective evolutionary algorithm, a population containing multiple coal blending schemes is generated during initialization. Each coal blending scheme consists of the blending ratio of each type of coal. For each individual in the population, its fuel purchase cost and combustion risk probability are calculated according to the economic function and safety function, respectively. Individuals that violate the constraints are corrected based on the penalty function to obtain the fitness value of the individual. The population individuals are non-dominated and ordered, Pareto ranks are assigned, and crowding is calculated for individuals in the same rank. The offspring population is generated through selection, crossover, and mutation operations. The parent population and the offspring population are merged and then subjected to non-dominated sorting and crowding screening to generate a new generation population. The process is iterated until the preset maximum number of iterations is reached or the Pareto front of the population tends to converge, and the Pareto optimal coal blending scheme set is output.

8. The method according to claim 3, characterized in that, After selecting and executing a coal blending scheme from the Pareto optimal coal blending scheme set based on the actual operating conditions of the thermal power unit, the process also includes: Boiler operation data is collected in real time and compared with combustion risk probability values ​​to generate incremental labeled samples; When the cumulative number of incrementally labeled samples reaches a preset threshold, the incrementally labeled samples are merged with the training sample set to obtain the merged training sample set. The boiler combustion risk prediction model based on weighted support vector machine is retrained using the merged training sample set to update and optimize the model parameters.

9. The method according to claim 1, characterized in that, The method further includes: When new coal types are added to the coal yard, coal market prices fluctuate beyond the preset range, or the operating conditions of thermal power units undergo significant changes, the optimization process based on a multi-objective evolutionary algorithm is restarted to generate an updated Pareto optimal coal blending scheme set.

10. The method according to claim 1, characterized in that, The boiler combustion risk prediction model based on weighted support vector machine is a weighted C-support vector machine or a weighted C-support vector machine. - Support Vector Machine.

Images