Sintering batching optimization scheme

Abstract

The invention discloses a sintering batching optimization scheme, and relates to the technical field of sintering ore batching control, and the technical key points are as follows: a random forest proxy model is used to predict the performance of finished ore, and a nonlinear mapping relationship between an independent variable and a dependent variable is used as a target function; carrying out Pareto optimal solution set search through non-dominated sorting and crowding distance calculation by adopting an NSGA-II (Non-dominated Sorting Genetic Algorithm-II) algorithm; calculating the weight of each target index by using an entropy weight method based on a Pareto solution set; weight is determined from a Pareto solution set generated by an NSGA-II algorithm, an optimal batching scheme closest to an ideal solution and far away from a negative ideal solution is screened out through a TOPSIS method, and multi-target comprehensive weighing is achieved; according to the method, sintering batching optimization based on finished ore performance feedback is realized, dynamic optimization of a batching plan can be effectively guided, the sinter quality control level is improved, the production cost is reduced, and technical support is provided for efficient and stable operation of a blast furnace ironmaking process.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of sintering ore batching control, and particularly relates to a sintering ore batching optimization scheme. BACKGROUND

[0002] As a basic pillar industry of the national economy, the sintering ore batching link of the steel industry is a key hub connecting the pretreatment of mineral resources and the blast furnace ironmaking, and has a decisive influence on the quality, cost and efficiency of steel production.

[0003] The iron ore resources in China have the significant characteristics of abundant total amount but high proportion of poor ores, and more than 70% of the iron ores need to be pretreated through crushing, screening and beneficiation processes to improve the grade, and then the sintering ore with suitable particle size, strength and chemical composition is formed. In this process, the quality indexes of the sintering ore (such as iron content, basicity, magnesium oxide and aluminum oxide content) directly determine the technical and economic indexes of the blast furnace ironmaking, for example, the blast furnace coke ratio can be reduced by 2%-3% and the blast furnace output can be increased by 3%-5% when the iron content is improved by 1%; when the basicity fluctuates by more than ±0.05, the stability of the blast furnace condition will be significantly reduced, resulting in the fluctuation of the molten iron quality and the rise of energy consumption. Therefore, accurate control of the sintering ore composition is the premise of guaranteeing the efficient operation of the blast furnace.

[0004] However, the sintering ore batching faces the challenge of complex multivariate constraints and target coordination restrictions. The raw material system usually contains dozens of minerals (such as hematite, magnetite, limonite, fluxing ore and returned ore), and the chemical composition (TFe, CaO, SiO2, etc.), physical properties (particle size, hydrophilicity) and cost of each raw material are significantly different, which needs to optimize the raw material combination cost while meeting the nonlinear composition constraints such as basicity and iron content. Under the traditional manual experience decision mode, the batching plan is formulated by relying on the technical personnel to calculate according to the historical data and operation experience, and there are three core problems: 1) insufficient composition control accuracy; 2) low cost optimization efficiency; 3) information processing lag: the integration of raw material attributes (composition, price, inventory) and test data relies on manual input, and the plan adjustment cycle is as long as several hours, which easily affects the production plan and wastes production resources.

[0005] Therefore, the present application aims to provide a sintering ore batching optimization scheme to solve the above problems. SUMMARY

[0006] The present application aims to provide a sintering ore batching optimization scheme to solve the above problems. The present application adopts a three-level architecture of "agent model dimension reduction -> multi-objective optimization search -> objective decision screening", integrates random forest (RF), NSGA-II, entropy weight method and TOPSIS method, forms a closed loop from data modeling to engineering decision, can effectively improve the sintering ore quality control level, reduce the production cost, and provide technical support for the efficient and stable operation of the blast furnace ironmaking process.

[0007] To achieve the above object, the technical scheme of the present application is as follows:

[0008] The present application provides a sintering burden optimization scheme, which comprises the following steps:

[0009] S1, obtaining sintering whole-process data;

[0010] S2, preprocessing the sintering process data to obtain the preprocessed whole-process data;

[0011] S3, constructing a random forest agent model to fit the nonlinear mapping relationship between the burden and the finished product performance;

[0012] S4, using NSGA-II multi-objective genetic algorithm to perform a Pareto optimal solution set search;

[0013] S5, using entropy weight method to calculate the weight;

[0014] S6, searching in the feasible region of the decision variable by using the TOPSIS method combined with the calculated weight to obtain a multi-objective optimization model;

[0015] S7, obtaining the optimal solution of the sintering burden scheme based on the multi-objective optimization model.

[0016] In S2, the preprocessing of the whole-process data includes: detecting and processing the abnormal values of the whole-process data by using the Relyada criterion.

[0017] In S3, for the nonlinear mapping relationship between the independent variables (raw material ratio, chemical composition) and the dependent variables (performance index, cost) in the sintering burden optimization, a data-driven prediction agent model is constructed by using the random forest algorithm.

[0018] Random forest forms a prediction model with strong generalization ability by integrating T decision trees (T tree ). Each tree is based on Bootstrap sampling to extract 63.2% samples from the training set to construct a sub-data set, and the remaining 36.8% unselected "out-of-bag data" (OOB Data) is used for unbiased error estimation.

[0019] The model accuracy is evaluated by using quantitative indicators, including the determination coefficient (R 2 ), root mean square error (RMSE), to ensure that the agent model can accurately map the nonlinear relationship between the independent variables and the dependent variables.

[0020] In S4, NSGA-II algorithm is used to search for the Pareto optimal solution set. The nonlinear mapping relationship between the independent variable value predicted by the random forest and the dependent variable and the lowest cost are taken as the objective function, the population is divided into different levels through non-dominated sorting, and the retention of the Pareto frontier solution is ensured; the diversity of solutions is maintained by combining the crowding distance calculation to avoid local convergence, and the population is iteratively updated through genetic operations such as selection, crossover and mutation, and a non-inferior solution set containing multi-objective trade-offs is generated in the solution space that meets the independent variable constraints (raw material ratio range).

[0021] Through genetic operations such as selection, crossover and mutation, the non-dominated sorting method is introduced to reduce the computational complexity of the algorithm, the method of crowding degree is used to ensure the diversity of individuals in the population, and the tournament selection is used to ensure that excellent individuals have a greater probability of being retained. A non-inferior solution set containing multi-objective trade-offs is generated in the solution space that meets the independent variable constraints (raw material ratio range).

[0022] In S5, entropy weight method is introduced for weight calculation. Entropy weight method is an attribute weight determination method based on information theory. It determines the weight of each attribute by analyzing the information entropy in the original data matrix. Specifically, entropy weight method determines the weight of each attribute according to the size of its information entropy. The larger the information entropy of an attribute, the smaller its weight. Based on information entropy theory, the information amount is reflected by the dispersion degree of target data.

[0023] In S6, TOPSIS method is used to calculate the weight of entropy weight method in the solution space that meets the independent variable constraints to find the optimal solution. The positive ideal solution and the negative ideal solution are defined. The positive ideal solution is the scheme in which all objectives reach the optimal value, and the negative ideal solution is the scheme in which all objectives reach the worst value. By determining the positive ideal solution and the negative ideal solution, a reference standard can be provided for each scheme, which is convenient for comparison and evaluation. The multi-objective problem is converted into the distance measurement of each scheme and ideal solution, and the closeness index is used to comprehensively evaluate the pros and cons of the scheme.

[0024] In S7, the weighted normalized matrix is constructed by combining the entropy weight method, and the Euclidean distance between each solution and the ideal solution is calculated, and the scheme with the largest closeness is finally selected as the optimal solution. This process converts the multi-objective trade-off into a quantifiable single-index ranking.

[0025] Compared with the prior art, the beneficial effects of the present scheme are:

[0026] The application constructs an intelligent optimization scheme including "complex relationship modeling-multi-objective search-objective decision", reconstructs the sinter burdening optimization process through data intelligent technology, realizes the leap from "experience decision" to "intelligent decision", and realizes the leap from "experience decision" to "intelligent decision". Through the adoption of "proxy model dimension reduction-multi-objective optimization search-objective decision screening" three-level architecture, the random forest (RF), NSGA-II, entropy weight method and TOPSIS method are combined to form a closed loop from data modeling to engineering decision, which can effectively improve the sinter quality control level, reduce the production cost, and provide technical support for the efficient and stable operation of the blast furnace ironmaking process. BRIEF DESCRIPTION OF DRAWINGS

[0027] Figure 1 is a technical framework schematic diagram in the embodiment of the application;

[0028] Figure 2 is a random forest algorithm prediction model flow schematic diagram in the embodiment of the application;

[0029] Figure 3 is a NSGA-II algorithm flow schematic diagram in the embodiment of the application;

[0030] Figure 4 is a part of the prediction result visualization diagram of the random forest proxy model in the training process in the embodiment of the application;

[0031] Figure 5 is a part of the prediction decision tree error visualization diagram of the random forest proxy model in the training process in the embodiment of the application. DETAILED DESCRIPTION

[0032] In order to enable the personnel in the art to better understand the present application scheme, the technical scheme of the present application will be further described in detail below in conjunction with the embodiments of the present application and the drawings. Obviously, the described embodiments are only a part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by the person skilled in the art without creative labor should belong to the scope of protection of the present application.

[0033] It should be noted that the embodiments in the present application and the features in the embodiments can be combined with each other without conflict. The present application will be described in detail below in conjunction with the embodiments.

[0034] Embodiment:

[0035] Through the sinter burdening optimization scheme of the present application, a sinter burdening optimization scheme based on finished product performance feedback is constructed, and the specific scheme includes:

[0036] S1, obtaining historical data

[0037] The sintering cup experimental data of a steel plant are collected, as shown in Table 1, including four types of raw material types, sintering process parameters, sinter chemical composition and sinter performance parameters.

[0038] Table 1 sintering process data summary

[0039]

[0040]

[0041] S2, data preprocessing

[0042] Including outlier detection and processing.

[0043] S3, establishing a random forest prediction model

[0044] A random forest model is constructed for each dependent variable. Each random forest integrates T decision trees (T tree ) to form a prediction model with strong generalization ability. Each tree is based on bootstrap sampling (Bootstrap) to extract 63.2% samples from the training set to construct a sub-data set, and the remaining 36.8% unselected “out-of-bag data” (OOB Data) is used for unbiased error estimation.

[0045] S4, model evaluation and reverse normalization

[0046] The fitting accuracy and generalization ability of the random forest regression prediction model are evaluated by the coefficient of determination (R 2 ), and the root mean square error (RMSE). The prediction results are inversely restored by reverse normalization method.

[0047] S5, constructing NSGA-II multi-objective genetic algorithm model

[0048] Including the dependent variable value predicted by random forest and the minimum cost as the objective function, supporting the addition of custom objective functions (such as cost calculation, energy consumption evaluation), dividing the population into different levels through non-dominated sorting, ensuring the preservation of Pareto frontier solutions; combined with the calculation of crowding distance to maintain the diversity of solutions, avoid local convergence, through genetic operations such as selection, crossover, mutation to update the population, the non-dominated sorting method is introduced to reduce the computational complexity of the algorithm, the method of crowding degree is used to ensure the diversity of individuals in the population, and the tournament selection is used to ensure that excellent individuals have a greater probability of being retained. In the solution space that meets the independent variable constraints (raw material ratio range), a non-inferior solution set containing multi-objective trade-offs is generated.

[0049] S6, constructing an entropy weight method module

[0050] It includes standardizing the decision matrix; introducing the entropy weight method to calculate the weight and find the ratio of each indicator under each weight; finding the information entropy of each indicator based on the information entropy theory; determining the weight of each indicator and reflecting its information content through the discrete degree of the target data.

[0051] S7, TOPSIS method module

[0052] The TOPSIS model, also known as the superiority and inferiority distance method, is a method for ranking a finite number of evaluation objects based on their proximity to an idealized goal. It can evaluate the relative superiority of existing objects. The specific method steps include using a standardized matrix and the weights of each indicator to calculate a weighted standardized matrix; determining the ideal solution and the negative ideal solution. The positive ideal solution is a solution in which all objectives reach the optimal value, and the negative ideal solution is a solution in which all objectives reach the worst value. By determining the positive ideal solution and the negative ideal solution, a reference standard can be provided for each solution to facilitate comparison and evaluation; calculating the Euclidean distance from each evaluation indicator to the positive and negative ideal distances; calculating the relative proximity of each evaluation object, sorting the ingredient solutions according to the proximity, and selecting the solution with the highest proximity as the optimal solution.

[0053] Furthermore, in step S2, the process is as follows:

[0054] S2-1, such as Figure 2 As shown, for outliers, the Laida criterion (3σ criterion) is used to judge outliers. For any data in the feature column, if it satisfies If the data is considered an outlier, the row of data is invalid and deleted. The standard deviation of each feature column is calculated as follows:

[0055]

[0056] Where μ is the data mean.

[0057] Furthermore, in step S3, the process is as follows:

[0058] S3-1. Build a random forest model for each dependent variable.

[0059] Specifically, Bootstrap sampling extracts T subsample sets {D1, D2, ... D T}, the sample size of each subset is N. The probability of each sample being selected in a single sampling is:

[0060]

[0061] The samples that were not selected (36.8%) constitute the out-of-bag dataset (OOB) and are used for error estimation and feature importance analysis.

[0062] Specifically, the feature is randomly selected, and the total number of features is M. For each subset Dt, a regression decision tree ht(x) is constructed, and m features are randomly selected when the node is split (m≤M).

[0063] The goal of node splitting in the regression tree is to minimize the mean square error (MSE). For feature j and split threshold and R R (j,s), the optimal split pair (j * ,s * ) is selected:

[0064]

[0065] where R L and R R are the sample sets of the left and right child nodes after splitting, c L and c R are the predicted values (i.e., the mean of the subset sample target) of the left and right child nodes:

[0066]

[0067] Specifically, the prediction result is integrated, and the mean of the prediction of all decision trees is taken as the final output, and the uncertainty of the prediction of a single tree is reduced through ensemble learning. The formula of the final output is:

[0068]

[0069] where is the average result; h t (x) is the prediction result of a single decision tree; and T is the number of decision trees.

[0070] S3-2, unbiased error estimation using out-of-bag data

[0071] The importance of the feature is measured by the change in out-of-bag error after permuting the feature value. After randomly permuting the feature, the contribution of the feature is measured by the change in out-of-bag error. If the error significantly increases after permutation, it indicates that the feature has an important influence on the model prediction. This index can not only be used for feature sorting, but also can identify key raw material parameters that have nonlinear interaction with the target function.

[0072] S4-1, the determination coefficient (R 2 ) and the root mean square error (RMSE) are used to evaluate the fitting precision and generalization ability of the random forest regression prediction model. The prediction results are restored by inverse normalization.

[0073] S5-1, first initialize the population POP, then perform non-dominated sorting and crowding degree calculation to obtain the evolution generation Gen=1.

[0074] Specifically, the population initialization randomly generates an initial population within the constraint range of the independent variables, and for each individual, the corresponding dependent variable value is predicted by the random forest surrogate model as the objective function value, and the optimal objective function such as maximizing iron grade or minimizing production cost is added.

[0075] Constraint range: (i = 1, …, p, j = 1, …, m, p is the population size)

[0076] Specifically, non-dominated sorting, for any two individuals x a and x b in the population, if for all objective functions k = 1, …, n, f k (x a ) ≤ f k (x b ), and there is at least one objective function such that f1(x a ) < f1(x b ), then x a dominates x b , denoted as x a < x b .

[0077] The population is divided into different levels F1, F2, … by non-dominated sorting, where F1 is the set of all non-dominated individuals, F2 is the set of non-dominated individuals among the remaining individuals after removing F1, and so on.

[0078] Specifically, the crowding distance calculation, for the individual x i in the same level F j , the calculation steps of its crowding distance d j are as follows:

[0079] 1) For each objective function k = 1, …, n, sort the individuals in Fi by objective function value f k from small to large.

[0080] 2) The crowding distance of the boundary individuals (the first and last individuals after sorting) is set to infinity, i.e.

[0081] 3) For other individuals j, the crowding distance calculation formula is:

[0082]

[0083] Where and are the maximum and minimum values of the kth objective function in F i , respectively.

[0084] S5-2 then performs selection-crossover-mutation operations on the population to generate a child population POP_n; then the parent population POP and the child population POP_n are merged and non-dominated sorting and congestion calculation are performed.

[0085] Specifically, tournament selection is adopted: after merging the offspring population and the parent population, individuals are randomly selected from the population, and excellent individuals with higher Pareto levels and larger crowding distances are selected to enter the mating pool to avoid premature convergence.

[0086] Specifically, simulated binary crossover is used: for the selected parent individual x a , x b , generate offspring individuals x c , x d , perform crossover work according to the formula, the crossover formula is as follows:

[0087]

[0088] where δ j Controls the crossover strength for random numbers that follow a specific distribution.

[0089] Specifically, polynomial mutation is used: for each dimension j of individual x, the mutated value x′ is generated j :

[0090] x′ j =x j +(U j -L j )·δ j

[0091] Among them, δ j Follows a multinomial distribution.

[0092] S5-3. Use the elite strategy to retain excellent individuals to generate a new parent population POP; finally, determine whether the genetic generation meets the set value. If not, repeat the selection-crossover-mutation operation. If so, the evolution is complete.

[0093] S6-1. Standardization of decision matrix.

[0094] Specifically, the Pareto solution set contains n solutions and k objectives, forming a decision matrix X = [x ij ] n×k where x ij is the jth target value of the i-th solution.

[0095] Specifically, the target value is divided into positive indicators and negative indicators, and the normalization matrix R=[r ij ] n×k , the standardization formula is as follows:

[0096] For positive indicators (such as iron grade, the larger the value, the better):

[0097] For negative indicators (such as production costs, the smaller the value, the better):

[0098] S6-2. Calculate the ratio of each indicator under each plan.

[0099] Specifically, according to Calculating the proportion of the jth target value of the i-th plan to the plan's indicators is actually to calculate the size of the plan's variation.

[0100] S6-3. Calculate the information entropy of each indicator.

[0101] Specifically, calculate the entropy value e of the jth indicator j , measure the degree of disorder of indicator information:

[0102]

[0103] If p ij =0, define p ij lnp ij =0, ensuring that the entropy value is non-negative.

[0104] S6-4. Determine the weight of each indicator.

[0105] Specifically, calculate the index difference coefficient g i =1-e j , reflecting the amount of effective information provided by the indicator.

[0106] Specifically, normalized weights:

[0107] S7-1. Calculate the relative progress of each evaluation object.

[0108] Specifically, the weight vector w=[w1,w2,...,w k ], for the standardized matrix V = [v ij ] n×k ,v ij =r ij w j Perform weighted processing:

[0109] The weight reflects the importance of the indicator, and the weighted matrix reflects the difference in contribution of each goal to decision-making.

[0110] S7-2. Determine the ideal solution and negative ideal solution

[0111] Specifically, the positive ideal solution is the optimal value of each target, that is, the maximum value of the positive index and the minimum value of the negative index; the negative ideal solution is the worst value of each target, that is, the minimum value of the positive index and the maximum value of the negative index.

[0112] Positive ideal solution If the jth target is positive;

[0113] If the jth target is positive.

[0114] Positive ideal solution If the jth target is positive;

[0115] If the jth target is positive.

[0116] S7-3, calculate the Euclidean distance of each evaluation index to the positive and negative ideal distance.

[0117] Specifically, the distance of each scheme to the positive ideal solution and the negative ideal solution is calculated:

[0118]

[0119] S7-4, calculate the relative progress C of each evaluation object i According to the closeness, the schemes are ranked in order of good and bad, and the final evaluation result is obtained.

[0120] Specifically, the relative closeness calculation formula is:

[0121] C i The closer to 1, the closer the scheme is to the positive ideal solution and away from the negative ideal solution, and the better the comprehensive performance.

[0122] The specific working process and working principle of the disclosed sintering burden optimization scheme based on finished product ore performance feedback can refer to the working process and working principle of the sintering burden optimization scheme based on finished product ore performance feedback in the present embodiment.

[0123] The above specific embodiments are only an explanation of the present application, and are not a limitation of the present application, and those skilled in the art can make modifications to the present embodiments without creative contribution after reading the present specification, but as long as the modifications are within the scope of the claims of the present application, they are protected by the patent law.

Claims

1. A sintering ingredient optimization scheme, characterized by: The scheme comprises the following steps: S1. Acquire the full process data of sintering ore, wherein the full process data includes raw material data, process parameters and sintering ore performance parameters; S2. Preprocessing the sintering process data to obtain the preprocessed full process data; S3. Construct a random forest proxy model to efficiently fit the nonlinear mapping relationship between ingredients and finished ore properties; S4, using NSGA-II multi-objective genetic algorithm to perform Pareto optimal solution set search; S5. Calculate weights using entropy weight method; S6. Using the TOPSIS method and the calculated weights, we search within the feasible region of the decision variables and obtain a multi-objective optimization model. S7. Based on the multi-objective optimization model, an optimal solution for sintering ingredients is obtained.

2. The sintering ingredient optimization solution according to claim 1, characterized in that: In S2, the preprocessing method is to detect and process outliers in the entire process data using the Laida criterion.

3. The sintering ingredient optimization solution according to claim 1, characterized in that: In S3, a random forest proxy model is used to construct a data-driven prediction framework, and a nonlinear mapping relationship between independent variables and dependent variables is generated as the objective function.

4. The sintering ingredient optimization solution according to claim 3, characterized in that: The random forest forms a prediction model with strong generalization ability by integrating T decision trees; each tree extracts 63.2% of the samples from the training set based on bootstrap sampling to construct a sub-dataset, and the remaining 36.8% is used for unbiased error estimation.

5. The sintering ingredient optimization solution according to claim 4, characterized in that: In S4, the NSGA-II algorithm is used to search for the Pareto optimal solution set, the population is divided into levels through non-dominated sorting, the Pareto frontier solution is retained, and the diversity of solutions is maintained by combining the crowding distance calculation.

6. The sintering ingredient optimization solution according to claim 5, characterized in that: The dependent variable value predicted by the random forest and the minimum cost are used as the objective function, a genetic algorithm is used to iteratively update the population, and a non-inferior solution set containing multi-objective trade-offs is generated in the solution space that satisfies the independent variable constraints.

7. The sintering ingredient optimization solution according to claim 1, characterized in that: In S5, the entropy weight method determines the weight of each attribute by analyzing the information entropy in the original data matrix.

8. The sintering ingredient optimization solution according to claim 1, characterized in that: In S6, the TOPSIS method is used to make the optimal solution decision. By defining positive and negative ideal solutions, the multi-objective problem is converted into a distance measurement between the batching scheme and the ideal solution, and the pros and cons of the scheme are comprehensively evaluated using the closeness index.

9. The sintering ingredient optimization solution according to claim 8, characterized in that: The weighted normalized matrix is ​​constructed by combining the weights of the entropy weight method, and the Euclidean distance between each solution and the ideal solution is calculated. Finally, the ingredient solution with the greatest closeness is selected as the optimal solution.

Images