Converter end point temperature intelligent prediction method
By combining a large language model with a four-layer architecture model of process data preprocessing and physical constraints, the accuracy and interpretability issues of converter end-point temperature prediction are solved, achieving high-precision, real-time temperature prediction that can adapt to complex operating conditions.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-03
- Publication Date
- 2026-03-13
AI Technical Summary
Predicting the final temperature of a converter faces challenges such as low accuracy, lack of interpretability, difficulty in integrating metallurgical physical constraints, inability to effectively extract semantic information from process parameters, and difficulty in balancing real-time performance and accuracy using traditional methods.
An intelligent prediction method based on a large language model is adopted. Through a four-layer architecture model of process data preprocessing, semantic transformation, thermodynamic and kinetic constraints, combined with a lightweight temperature prediction model and knowledge distillation technology, the physical constraints and high accuracy of temperature prediction are achieved.
It achieves high-precision prediction of converter end temperature, improves the interpretability of the model and its acceptance in industrial settings, reduces computational complexity, adapts to fluctuations under different operating conditions, and has real-time performance and robustness.
Smart Images
Figure CN121659754A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of converter steelmaking technology, and in particular to a method for intelligent prediction of converter endpoint temperature. Background Technology
[0002] Predicting the final temperature of a converter is a core technical challenge in converter steelmaking, directly determining steel quality, energy consumption control, and production efficiency. Accurate final temperature enables precise process control, avoiding energy waste caused by overheating and quality defects and other quality problems caused by insufficient temperature, thus playing a crucial role in improving the economic benefits of steel enterprises.
[0003] Key technical challenges in predicting converter endpoint temperatures include: highly nonlinear temperature changes, the complexity of multi-element coupled reactions, the uncertainty of operating condition fluctuations, and low real-time prediction accuracy. Traditional temperature prediction methods have significant limitations in addressing these challenges.
[0004] Limitations of traditional temperature prediction methods: (1) Experience judgment method: relies on the experience accumulation of operators to estimate temperature, which is highly subjective, with large fluctuations in prediction accuracy and difficulty in dealing with complex working conditions; (2) Mechanism model method: establishes temperature prediction model based on thermodynamic and kinetic equations. Although the physical meaning is clear, the parameter calibration is complicated, the adaptability to non-standard working conditions is poor, and the prediction error is usually large; (3) Statistical regression method: establishes the regression relationship between temperature and process parameters through historical data, which requires a lot of feature engineering. The model has limited generalization ability and fails to predict abnormal working conditions.
[0005] The shortcomings of existing deep learning temperature prediction methods are: (1) They are black-box characteristics, lack interpretability, have low acceptance in industrial settings, and are difficult to gain the trust of operators; (2) They are difficult to effectively integrate professional knowledge and physical constraints in the metallurgical field, and the prediction results may violate basic thermodynamic laws; (3) They do not fully explore the temperature-related process mechanisms and fail to make full use of the inherent laws of temperature changes; (4) It is difficult to balance real-time performance and prediction accuracy, and high-precision models often have excessive computational complexity.
[0006] Data feature challenges in temperature prediction: (1) It is difficult to mine the temperature semantic information of numerical process parameters, lack a deep understanding of the factors affecting temperature, and it is difficult to establish the causal relationship between parameters and temperature changes; (2) Temperature changes involve the coupling effect of multiple metallurgical reactions such as decarburization, dephosphorization, and desulfurization, and the reaction heat release mode is complex; (3) The temperature control mode under different working conditions is significantly different, and the model needs to have strong adaptability; (4) The lag and discreteness of temperature measurement increase the technical difficulty of real-time prediction.
[0007] Physical constraints in temperature prediction: Existing prediction methods generally lack effective constraints on the physical laws governing temperature changes. Prediction results often violate the principle of thermal equilibrium, exceed reasonable temperature ranges, and are inconsistent with reaction kinetics, which seriously affects the reliability and practicality of the prediction results.
[0008] In recent years, large language models have demonstrated groundbreaking capabilities in complex reasoning and semantic understanding, providing new technical pathways for solving semantic modeling, knowledge fusion, and interpretability issues in temperature prediction. However, research on the application of large language models in industrial temperature prediction is still in its early stages, especially in the complex industrial scenario of converter final temperature, where its application potential needs further exploration and verification. Summary of the Invention
[0009] This application provides an intelligent prediction method for converter endpoint temperature, which addresses key technical problems in converter endpoint temperature prediction such as low accuracy, lack of interpretability, difficulty in integrating metallurgical physical constraints, inability to effectively mine semantic information of process parameters, and difficulty in balancing real-time performance and accuracy. It provides an intelligent prediction solution based on a large language model.
[0010] This application provides a method for intelligent prediction of converter endpoint temperature, the method comprising: The process data of the converter steelmaking process is collected and intelligently preprocessed to output preprocessed process data. The process data includes molten iron composition, molten iron temperature, molten iron charging amount, oxygen blowing time, total oxygen consumption, number of furnace turnings, slag splashing time, lance pressing time, scrap steel ratio, amount of various slag-forming materials added, furnace number, steel grade, work group, shift number, smelting timestamp, tapping timestamp, and final temperature. The intelligent preprocessing includes automatic feature type identification and classification, multi-dimensional feature distribution analysis, advanced correlation analysis and temperature feature relationship mining, converter process constraint definition and temperature boundary verification, key process parameter identification and temperature influencing factor analysis, intelligent missing value handling strategy, IQR outlier detection and processing, numerical coding of classification features, data standardization preprocessing, and temperature prediction-oriented data quality verification. The numerical parameters in the preprocessed process data are converted into structured natural language descriptions, and semantic process data is output. A pre-trained large language model is performed, and the semantic data is used for domain-specific fine-tuning to output the fine-tuned large language model. Thermodynamic constraints, kinetic constraints, and process feasibility constraints are embedded in the fine-tuned large language model to form a model framework with physical constraints. Based on the physical constraint model framework, a four-layer architecture model including an input layer, an encoding layer, a constraint layer, and an output layer is constructed, and the semantic process data and composite loss function are used for training to output a trained temperature prediction model. The trained temperature prediction model is subjected to knowledge distillation and dynamic quantization to output an optimized lightweight temperature prediction model. Input the test data into the optimized lightweight temperature prediction model and output the final temperature prediction result.
[0011] Furthermore, the automatic identification and classification of feature types includes: The timestamp data is identified as a time feature; wherein, the timestamp data includes smelting timestamps and steel tapping timestamps; Continuous variables are identified as numerical features; wherein, the continuous variables include molten iron composition, molten iron temperature, molten iron charging amount, oxygen blowing time, total oxygen consumption, number of furnace turnings, slag splashing time, lance pressing time, scrap steel ratio, and the amount of various slag-forming materials added; Discrete variables are identified as categorical features; wherein, the discrete variables include furnace number, steel grade, work group, and shift. The endpoint temperature is identified as the target feature.
[0012] Furthermore, the multi-dimensional feature distribution analysis includes: Statistical characteristic analysis: Calculate the mean of each feature. μ Standard deviation σ Minimum value (min), maximum value (max), quartiles; Skewness analysis: using skewness coefficient To assess the symmetry of the data distribution, the formula for calculating the skewness coefficient is: in E Represents the expected value, when Skew A value greater than 1 indicates that the data exhibits a right-skewed distribution. Skew A skewness of less than -1 indicates a left-skewed distribution, and a skewness close to 0 indicates a relatively symmetrical distribution. The skewness coefficient is used to determine the distribution characteristics of temperature data. Distribution type detection: Shapiro-Wilk normality test was used, and the test statistic was: in x (i) These are the sorted sample values. a i The Shapiro-Wilk coefficient, The sample mean. i n is the sample number, and n is the number of samples; WThe test statistic is set to a range of [0,1]. The closer the value is to 1, the more the data conforms to a normal distribution. The calculated W value is compared with the theoretical distribution to obtain the corresponding p value. The p value is based on the W statistic and the sample size n, and is obtained with reference to its specific theoretical distribution. When the p value is > 0.05, the normal distribution hypothesis is accepted, and when the p value is ≤ 0.05, the normal distribution hypothesis is rejected. Missing value proportion assessment: Calculate the missing value rate: in This represents the number of missing values. For the total number of samples, when Delete the feature if it exceeds 90%. If the percentage is less than 80%, it will be retained for further processing.
[0013] Furthermore, the advanced correlation analysis and temperature feature relationship mining include: Pearson correlation analysis: Calculating the correlation coefficient r : in and Variables X and Y The mean, X The process parameters are represented by Y, the final temperature is represented by Y, and the correlation coefficient is represented by Y. r The value range of is [-1, 1]. and Representing the i Process parameters and endpoint temperature in each sample n The number of samples; Temperature correlation analysis: For the target of predicting the final temperature, key process parameters that are highly correlated with the final temperature are identified and classified according to their correlation level: |r|>0.7 is highly correlated, |r|>0.5 is strongly correlated, |r|>0.3 is moderately correlated, and |r|≤0.3 is weakly correlated. Highly correlated and strongly correlated features are given priority for use in the temperature prediction model. Strongly correlated feature pair identification: Construct a correlation matrix to automatically identify feature pairs with an absolute correlation coefficient greater than 0.7 for multicollinearity detection.
[0014] Furthermore, the converter process constraint definition and temperature boundary verification include: The composition range of molten iron was set as follows: pig iron C content (3.5, 4.8)%, Si content (0.3, 1.2)%, Mn content (0.3, 1.0)%, P content (0.05, 0.25)%, S content (0.01, 0.06)%; the final composition was C content (0.03, 0.25)%, Si content (0.01, 0.35)%, Mn content (0.10, 0.80)%, P content (0.005, 0.025)%, S content (0.005, 0.025)%. Based on the molten iron temperature range of (1280, 1400) °C, the endpoint temperature range 1 of (1580, 1680) °C, and the endpoint temperature range 2 of (1580, 1680) °C, the predicted results are ensured to meet physical feasibility through a constraint boundary clipping operation: in, This is the endpoint temperature value after physical constraint verification and correction. This represents the lower limit constraint value for temperature. This represents the upper limit constraint value for temperature. The endpoint temperature value predicted by the model; The time constraints are set as follows: oxygen blowing time range (600, 1200) seconds, slag splashing time range (60, 300) seconds, and lance pressing time range (10, 60) seconds; boundary replacement processing is performed when the time parameters are abnormal. Within the range of blast furnace pig iron production (60, 120) tons, scrap steel ratio (0.10, 0.25), and oxygen consumption per unit volume (45, 65) m³, the following parameters are considered: 3 Based on / t steel, ensure that the material balance meets process requirements: in, This refers to the total mass of materials used in the converter smelting process. To ensure the quality of the molten iron added to the converter, To ensure the quality of the scrap steel added to the converter, The quality of other added materials.
[0015] Furthermore, the identification of key process parameters and the analysis of temperature influence factors include: Identify the raw material addition amount characteristics that begin with the first target character as the raw material parameter group. The raw material addition amount characteristics include slag-forming and alloying materials such as limestone, fluorite, and ferrosilicon, and calculate the total amount of raw materials. Identify the first component feature that begins with the second target character as the iron composition parameter group. The first component feature includes the content of C, Si, Mn, P and S elements. Identify the second component feature that begins with the third target character as the endpoint component parameter group to reflect the component control effect of the smelting process; Key operating parameters are identified as a group of process parameters, including oxygen blowing time, total oxygen consumption, number of furnace turnings, slag splashing time, and lance pressing time. Temperature-related parameters are identified as a group of thermal parameters, including pig iron temperature and final temperature.
[0016] Furthermore, the intelligent missing value handling strategy includes: Hierarchical missing value identification: Quantitative evaluation is performed using the missing value rate of each feature. The formula for calculating the missing value rate of each feature is as follows: in The feature missing rate, This represents the number of missing values. Total number of samples Missing value removal strategy: When the number of samples with missing values in the dataset is less than 10% and the missing values are randomly distributed, directly delete the rows containing missing values; if the missing values of a certain feature exceed 80%, delete that feature column to avoid features with insufficient information affecting the temperature prediction effect. Intelligent imputation of numerical features: For numerical features, if the data is normally distributed, the mean is used to impute missing values; when the data is skewed or has outliers, the median is used to impute missing values. Categorical feature imputation: The mode imputation method is used to handle missing discrete variables, and the category with the highest frequency is selected as the imputation value; Target variable handling: Directly delete sample rows with missing values for the endpoint temperature; The IQR outlier detection and processing includes: Outlier boundary calculation: The interquartile range is calculated using the following formula. : in, It is the first quartile. It is the third quartile; The criteria for outlier determination are: When process parameters A value is considered an outlier if it meets the outlier criteria. (2) Outlier Proportion Statistics: Calculate the outlier proportion for each feature using the following formula: in The number of outliers. The total number of samples; Conditional boundary replacement: percentage of outliers When the percentage is less than 5%, the boundary value replacement method is used to replace the outliers with the corresponding boundary values in order to maintain data continuity. Outlier retention strategy: Outlier percentage When the value is ≥ 5%, keep the original value to avoid information loss and prevent over-processing from causing the loss of important process information.
[0017] Furthermore, the domain knowledge-driven feature engineering includes: The total amount of raw materials is used as a comprehensive characteristic of raw materials; The temperature equilibrium characteristics are calculated using the following formula. : in For total input heat, For total heat loss, The total mass of the materials. This represents the average specific heat capacity. The process efficiency characteristic is calculated using the following formula. : in This represents the total oxygen consumption. Oxygen blowing time; The reaction intensity index is calculated using the following formula. : in The temperature at which molten iron enters the furnace. Total amount of materials; The numerical encoding of the classification features includes: Label encoding processing: Low-cardinality categorical features are numerically converted using LabelEncoder, transforming discrete variables such as furnace number, steel grade, and work group into continuous numerical values. Encoding consistency verification: Ensures the stability and reproducibility of encoding mapping, and guarantees that the same encoding rules are used during the training and deployment phases; Encoding feature extension: The processed features are added to the numerical feature set and processed in a unified manner with other continuous variables; The temperature prediction-oriented data quality verification includes Constraint violation detection: Count the number of data points that violate process constraints and calculate the violation rate. : in The number of data points that violate the constraints. Total number of data points Statistical consistency verification: Compare the statistical characteristics of the data before and after processing, and calculate the mean difference rate. : in, This represents the sample mean after data preprocessing. This represents the sample mean of the original data. Temperature distribution consistency check: The focus is on verifying the consistency of the distribution of temperature-related features to ensure that the preprocessed temperature data retains its original physical meaning and statistical regularity.
[0018] Furthermore, the thermodynamic constraints include: Temperature change pattern: Heat of reaction calculate: Thermal balance constraint: Input heat = Output heat + Accumulated heat; The dynamic constraints include: Relationship between decarbonization reaction and temperature: in The carbon content (%) in molten iron. For time, is the decarbonization rate function, representing the functional relationship of the rate of change of carbon content.
[0019] Rate of temperature rise: in For temperature, is the decarbonization rate function, representing the functional relationship of the rate of change of carbon content.
[0020] Temperature range constraint: 1550°C ≤ T ≤ 1750°C; The process feasibility constraints include the temperature distribution range based on historical data; The methods for embedding thermodynamic constraints, kinetic constraints, and process feasibility constraints in the fine-tuned large language model include soft constraints, hard constraints, and adaptive weights; wherein, the soft constraints include adding physical constraint terms to the loss function, the hard constraints include using a projection algorithm to ensure that the prediction results satisfy physical laws, and the adaptive weights include dynamically adjusting the constraint weights according to the characteristics of the operating conditions.
[0021] Furthermore, the composite loss function is expressed as: in For composite loss function, This is due to the loss of accuracy in temperature prediction; Loss due to violation of physical constraints; This represents the semantic consistency loss. These are the physical constraint weighting coefficients; This is the semantic consistency weight coefficient.
[0022] The intelligent prediction method for converter endpoint temperature provided in this application has at least the following beneficial effects: 1) This application leverages the powerful complex pattern recognition and semantic understanding capabilities of a large language model to deeply explore the highly nonlinear intrinsic relationship between process parameters and endpoint temperature, overcoming the limitations of traditional mechanistic models and statistical regression methods. Experimental results show that the prediction accuracy of this method can reach over 99%, far exceeding that of traditional methods.
[0023] 2) This application innovatively converts numerical process parameters into natural language descriptions, making the model's input and decision-making process easier for on-site operators to understand and trust, and greatly improving the acceptance of artificial intelligence models in industrial settings.
[0024] 3) This application fundamentally avoids problems such as prediction results violating thermodynamic laws and exceeding reasonable ranges by embedding physical constraints such as heat balance equations, reaction kinetics, and temperature ranges into the model in the form of soft and hard constraints, thus ensuring the reliability and practicality of the prediction results.
[0025] 4) This application adopts a lightweight large language model and combines inference optimization techniques such as LoRA efficient fine-tuning, knowledge distillation and dynamic quantization to significantly reduce the computational complexity and inference time of the model while maintaining high prediction accuracy, thus meeting the real-time requirements of the converter steelmaking process.
[0026] 5) This application integrates the entire process from data preprocessing and feature engineering to model training and optimization. It has automated missing value handling, outlier detection, feature generation and quality verification capabilities. It can adapt to fluctuations in different working conditions and shows good robustness to abnormal data and complex working conditions. Attached Figure Description
[0027] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.
[0028] Figure 1 A flowchart illustrating an intelligent prediction method for converter endpoint temperature provided in this application embodiment; Figure 2 The interface diagram of the converter endpoint temperature prediction system provided in the embodiments of this application shows the prediction results; Figure 3 The loss function convergence curve of the converter endpoint temperature prediction model based on a large language model provided in the embodiments of this application during the training process.
[0029] The accompanying drawings illustrate specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concept of this application to those skilled in the art through reference to particular embodiments. Detailed Implementation
[0030] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.
[0031] The collection, storage, use, processing, transmission, provision, and disclosure of financial data or user data involved in the technical solution of this application all comply with the provisions of relevant laws and regulations and do not violate public order and good morals.
[0032] It should be noted that in the embodiments of this application, certain software, components, models and other existing solutions in the industry may be mentioned. These should be regarded as exemplary and are only intended to illustrate the feasibility of implementing the technical solution of this application. However, it does not mean that the applicant has used or necessarily used the solution.
[0033] The technical solution of this application and how the technical solution of this application solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of this application will now be described with reference to the accompanying drawings.
[0034] This application provides a method for intelligent prediction of converter endpoint temperature. This method can be configured and executed on a computer in the form of a software program or modular components. Before implementing this method, the required environment should be configured on the computer according to actual needs. Figure 1 A flowchart of a method for intelligent prediction of converter endpoint temperature provided in this application embodiment is shown below. Figure 1 As shown, the intelligent prediction method for converter endpoint temperature includes the following steps S10 to S70.
[0035] Step S10: Collect process data during the converter steelmaking process, perform intelligent preprocessing on the process data, and output the preprocessed process data; In this embodiment, the process data includes the following 23 types of data: 1) Iron composition: percentage content of C, Si, Mn, P, and S in molten iron, % 2) Molten iron temperature, °C; 3) Amount of molten iron charged, tons / ladle; 4) Oxygen blowing time, seconds; 5) Total oxygen consumption, m 3 ; 6) Number of furnace turnings, times; 7) Slag splashing time, seconds; 8) Recoil control time, seconds; 9) Scrap steel ratio, % 10) Amount of quicklime added, kg; 11) Amount of limestone added, kg; 12) Amount of raw stone added, kg; 13) Amount of fluorite added, kg; 14) Amount of ferrosilicon added, kg; 15) Amount of magnetic separation slag added, kg; 16) Amount of other slag-forming materials added, kg; 17) Furnace number, classification characteristics; 18) Steel grades and classification characteristics; 19) Work groups, classification characteristics; 20) Class schedules and classification characteristics; 21) Smelting timestamp, time characteristics; 22) Steel tapping timestamp, time characteristics; 23) Endpoint temperature, °C.
[0036] Intelligent preprocessing of process data includes the following sub-steps S101-S111.
[0037] S101: Automatic feature type identification and classification, identifying the following four types of features: (1) Time characteristics: timestamp data such as converter smelting time and tapping time; (2) Numerical characteristics: continuous variables such as molten iron composition and operating parameters; (3) Classification characteristics: discrete variables such as furnace number, steel grade, work group, and shift; (4) Target characteristic: endpoint temperature.
[0038] S102: Multidimensional feature distribution analysis, including the following steps S1021-S1024: S1021. Statistical Characteristic Analysis: Calculate the mean μ, standard deviation σ, minimum value min, maximum value max, and quartiles Q1 and Q3 for each characteristic using the following formula: Calculate the sample mean, where X i Let be the value of the i-th sample, and n be the total number of samples. The standard deviation is expressed by the formula: Calculate, where μ represents the typical level of the process parameter and σ represents the degree of fluctuation of the process parameter; S1022, Skewness Analysis: Using skewness coefficient: Evaluate the symmetry of the data distribution, where E represents the expected value, when Skew A value greater than 1 indicates that the data exhibits a right-skewed distribution. Skew A value less than -1 indicates a left-skewed distribution, and a skewness close to 0 indicates a relatively symmetrical distribution. This index is used to determine the distribution characteristics of temperature data. S1023, Distribution type detection: Use the Shapiro-Wilk normality test, test statistic: in These are the sorted sample values. The Shapiro-Wilk coefficient, The sample mean. W To test the statistic, the value range is [0,1]. The closer the value is to 1, the more the data conforms to a normal distribution. By comparing the calculated W value with the theoretical distribution, the corresponding p value is obtained. The p value is based on the W statistic and the sample size n, referring to its specific theoretical distribution. When the p value > 0.05, the normal distribution hypothesis is accepted; when the p value ≤ 0.05, the normal distribution hypothesis is rejected. This provides a basis for the selection of subsequent temperature prediction models. S1024. Missing value ratio assessment: Calculate the missing value rate. : in This represents the number of missing values. For the total number of samples, when Delete the feature if it exceeds 90%. If the percentage is less than 80%, it will be retained for further processing.
[0039] S103: Advanced Correlation Analysis and Temperature Feature Relationship Mining, including the following steps S1031-S1033: S1031, Pearson correlation analysis: Calculate the linear correlation coefficient: in and Variables X and Y The mean, X Representative process parameters, Y The value of the correlation coefficient r represents the endpoint temperature and ranges from [-1, 1]. S1032. Temperature Correlation Analysis: For the target of predicting the final temperature, identify key process parameters that are highly correlated with the final temperature and classify them according to their correlation level: |r|>0.7 is highly correlated, |r|>0.5 is strongly correlated, |r|>0.3 is moderately correlated, and |r|≤0.3 is weakly correlated. Highly correlated and strongly correlated features are preferred for the temperature prediction model. S1033. Identification of Strongly Correlated Feature Pairs: Construct a correlation matrix and automatically identify feature pairs with an absolute correlation coefficient greater than 0.7 for multicollinearity detection, avoiding redundant features from affecting temperature prediction accuracy.
[0040] S104: Converter process constraint definition and temperature boundary verification, including the following steps S1041-S1044: S1041. Determine the compositional constraint boundary: Define the reasonable range of molten iron composition: pig iron C content (3.5, 4.8)%, Si content (0.3, 1.2)%, Mn content (0.3, 1.0)%, P content (0.05, 0.25)%, S content (0.01, 0.06)%; endpoint composition C content (0.03, 0.25)%, Si content (0.01, 0.35)%, Mn content (0.10, 0.80)%, P content (0.005, 0.025)%, S content (0.005, 0.025)%; S1042. Perform temperature constraint verification: Molten iron temperature range (1280, 1400) °C, endpoint temperature one range (1580, 1680) °C, endpoint temperature two range (1580, 1680) °C. Ensure the prediction results meet physical feasibility through constraint boundary clipping operations. in, This is the endpoint temperature value after physical constraint verification and correction. This represents the lower limit constraint value for temperature. This represents the upper limit constraint value for temperature. The endpoint temperature value predicted by the model; S1043. Perform time constraint verification: oxygen blowing time range (600, 1200) seconds, slag splashing time range (60, 300) seconds, lance pressing time range (10, 60) seconds, and perform boundary replacement processing when time parameters are abnormal. S1044. Determine material balance constraints: blast furnace pig iron quantity range (60, 120) tons, scrap steel ratio range (0.10, 0.25), oxygen consumption range (45, 65) m³. 3 / t steel, ensuring the material balance meets process requirements: in, This refers to the total mass of materials used in the converter smelting process. To ensure the quality of the molten iron added to the converter, To ensure the quality of the scrap steel added to the converter, The quality of other added materials.
[0041] S105: Identification of key process parameters and analysis of temperature influence factors, including the following steps S1051-S1055: S1051. Identify raw material parameter groups: Identify raw material addition characteristics starting with '_', including slag-forming and alloying materials such as limestone, fluorite, and ferrosilicon, and calculate the total raw material amount. : in M i This represents the weight of the i-th type of raw material, where n is the total number of types of raw materials. S1052. Identify the molten iron composition parameter group: Identify the first component characteristic starting with 'pig iron_', including the content of elements such as C, Si, Mn, P, and S. These parameters directly affect the release of reaction heat and temperature changes. S1053. Identify the endpoint component parameter group: Identify the second component feature that starts with 'endpoint_', which reflects the component control effect of the smelting process and is closely related to the endpoint temperature; S1054. Identify process parameter groups: including key operating parameters such as oxygen blowing time, total oxygen consumption, number of furnace turnings, slag splashing time, and lance pressing time, which directly affect the accuracy of temperature control; S1055. Identify thermal parameter groups: including pig iron temperature, endpoint temperature, and other temperature-related parameters, which constitute the core feature set for temperature prediction.
[0042] S106: Intelligent missing value handling strategy, including the following steps S1061-S1065: S1061, Hierarchical Missing Value Identification: Calculate the missing value rate of each feature for quantitative evaluation. The formula for calculating the missing value rate of each feature is as follows: in The feature missing rate, This represents the number of missing values. The total number of samples.
[0043] S1062 Missing value deletion strategy: When the number of samples with missing values in the dataset is less than 10% and the missing values are randomly distributed, directly delete the rows containing missing values; if the missing values of a certain feature exceed 80%, delete that feature column to avoid the influence of insufficient information on the temperature prediction effect. S1063. Intelligent imputation of numerical features: For numerical features, if the data is normally distributed (determined by Shapiro-Wilk test p>0.05), the mean is used to impute missing values; when the data has a skewed distribution or outliers, the median is used to impute missing values to ensure the robustness of the imputed values. S1064. Categorical feature imputation: The mode imputation method is used to handle missing discrete variables, and the category with the highest frequency is selected as the imputation value. S1065. Target variable processing: Directly delete sample rows with missing values for the endpoint temperature to ensure the integrity and reliability of the training data.
[0044] S107: IQR outlier detection and handling, including the following steps S1071-S1074: S1071, Outlier Boundary Calculation: The outlier range is defined as follows: Interquartile range The result is obtained through the following formula: in It is the first quartile (25th percentile). It is the third quartile (75th percentile); The criteria for outlier determination are: When process parameters If the above conditions are met, it is determined to be an outlier; S1072. Outlier Ratio Statistics: Calculate the number and percentage of outliers for each feature, and the outlier ratio. Calculate using the following formula: in The number of outliers. The proportion of outliers in the total sample size. Used to guide subsequent processing strategies; S1073, Conditional Boundary Replacement: Outlier Ratio When the percentage is less than 5%, the boundary value replacement method is used to replace the outliers with the corresponding boundary values in order to maintain data continuity. S1074. Outlier Retention Strategy: When the proportion of outliers is too high (greater than or equal to 5%), retain the original values to avoid information loss and prevent over-processing from causing the loss of important process information. S108: Domain knowledge-driven feature engineering, including the following steps S1081-S1084: S1081. Calculation of the comprehensive characteristics of raw materials: The total amount of raw materials is used as the comprehensive characteristics of raw materials. The calculation formula is as follows: This formula sums up the weights of various additives, where Indicates the first i Weight of each raw material n The total number of raw material types covers all slag-forming and alloying materials, including limestone, fluorite, and ferrosilicon. S1082. Calculation of Temperature Balance Characteristics: Based on the temperature-related characteristics calculated according to the heat balance equation, the following formula is used to predict temperature changes: in For total input heat, For total heat loss, The total mass of the materials. The average specific heat capacity reflects the heat balance during the process. S1083. Calculate process efficiency characteristics, including oxygen utilization efficiency. Calculate using the following formula: Total oxygen consumption Divide by oxygen blowing time Adding 1 to the result is used as oxygen utilization efficiency. Adding 1 to the denominator is used to avoid division by zero errors. This indicator reflects the oxygen consumption rate per unit time and the level of process efficiency. S1084. Calculation of reaction intensity index: The intensity of the composite process based on temperature-time-flow rate is calculated using the following formula: The reaction intensity index Taking into account the temperature of molten iron entering the furnace Total oxygen consumption Oxygen blowing time and total amount of materials The combined effect of the reaction process is used to quantify the intensity level of the reaction process.
[0045] S109: Numerical encoding of classification features, including the following steps S1091-S1093: S1091, Label Encoding Processing: Use LabelEncoder to convert low cardinality classification features into numerical values, converting discrete variables such as furnace number, steel grade, and shift group into continuous numerical values. S1092, Encoding Consistency Verification: Ensure the stability and reproducibility of encoding mapping, and guarantee that the same encoding rules are used during the training and deployment phases; S1093, Encoding Feature Extension: The processed features are added to the numerical feature set and processed in a unified manner with other continuous variables.
[0046] S110: Data standardization preprocessing, including the following steps S1101-S1103: S1101, Z-score standardization: To achieve a data distribution with a mean of 0 and a standard deviation of 1, the standardization formula is: For raw data X Perform the conversion, where μ The sample mean. σ This is the sample standard deviation. The sample mean is calculated using the following formula: The sample standard deviation is calculated using the following formula: The standardized data satisfy the following normal distribution characteristics: S1102, Standardized parameter storage: Store the mean μ and variance σ parameters for use in the deployment phase to ensure that training and inference use a consistent standardized benchmark; S1103. Distribution Characteristics Verification: Check the statistical characteristics of the standardized data to verify whether the data conforms to the expected normal distribution, where μ represents the typical level of the process parameter, σ represents the degree of fluctuation of the process parameter, and Z value represents the degree of deviation of the current parameter from the typical level.
[0047] S111, Data quality verification guided by temperature prediction, including S1111-S1113: S1111, Constraint Violation Detection: Count the number of data points that violate process constraints, and calculate the constraint violation rate. in The number of data points that violate the constraints. Total number of data points S1112. Statistical Consistency Verification: Compare the statistical characteristics of the data before and after processing, calculate the mean difference rate, and ensure that data preprocessing does not change the basic statistical characteristics of the original data. The formula for calculating the mean difference rate is: in, This represents the sample mean after data preprocessing. This represents the sample mean of the original data. S1113. Temperature distribution consistency check: Focus on verifying the consistency of temperature-related feature distribution to ensure that the preprocessed temperature data retains its original physical meaning and statistical regularity.
[0048] Step S20: Convert the numerical parameters in the preprocessed process data into structured natural language descriptions and output semantic process data.
[0049] In this embodiment, step S20 can be implemented through the following steps S201-S203.
[0050] S3.1: Construct a semantic mapping rule base for process parameters.
[0051] First, based on metallurgical expertise, historical operational experience, and process specifications, a structured semantic mapping rule base is constructed within the computer system. This rule base is preferably stored in JSON or XML format, defining the mapping relationship between the numerical ranges of key process parameters and their corresponding natural language descriptions. Core mapping rules include, but are not limited to: 1) Composition parameter mapping rules: For the composition parameters of molten iron and the endpoint, semantic division is performed based on their content levels.
[0052] Taking "carbon (C) content" of molten iron as an example: When the content is in the range (4.0, 5.0]%, it is mapped to the description: "High carbon molten iron, heavy decarburization task, large heat generation"; When the content is in the range (3.8, 4.0]%, it is mapped to the description: "Medium carbon molten iron, moderate decarburization reaction"; When the content is in the range [3.5, 3.8]%, it is mapped to the description: "Low carbon molten iron, light decarburization burden".
[0053] 2) Temperature parameter mapping rules: For the molten iron temperature and the final temperature parameters, the state is described according to their temperature levels.
[0054] Taking "molten iron temperature" as an example: When the temperature is in the range [0, 1300)℃, it is mapped to the description: "Low-temperature molten iron requires additional heating, resulting in increased energy consumption"; When the temperature is in the range [1300, 1400]℃, it is mapped to the description: "moderate temperature, standard operating range, good thermal balance"; When the temperature is in the range (1400, 1500)℃, it is mapped to the description: "High temperature molten iron, violent reaction, process needs to be controlled".
[0055] 3) Operation parameter mapping rules: For operation parameters such as oxygen blowing and feeding, the operation intensity is described according to their magnitude.
[0056] Taking "total oxygen consumption" as an example: When the consumption is in the range (55000,∞) Nm 3 At that time, the mapping was described as: "High-flow oxygen supply enhances the decarbonization reaction, and the temperature rises rapidly"; When the consumption is in the range [45000, 55000] Nm 3 At that time, the mapping is described as: "The oxygen supply is moderate, and the reaction is stable"; When the consumption is in the range [0, 45000) Nm 3 At that time, the mapping was described as: "Low-flow oxygen supply, mild reaction".
[0057] S201: Create and execute a semantic converter: Based on the mapping rule base, create a semantic converter to automatically convert numerical data into text.
[0058] In a preferred embodiment, the converter's execution flow is as follows: the converter loads a predefined semantic mapping rule base; for an input process data sample (i.e., a row of data), the converter traverses its key process parameters; for each parameter, it matches it in the rule base according to its name and value to find the corresponding semantic description text; and it combines the semantic description text of all parameters with a preset instruction template to generate a structured dictionary object containing "instruction-input" pairs.
[0059] S203: Perform batch data transformation and output: Call the semantic converter to perform batch transformation on the entire dataset after preprocessing in step S10. Save the generated structured text data in a standard format, such as JSONL, as the training dataset for subsequent fine-tuning of the large language model. Through the above processing, the original numerical data is transformed into text rich in domain knowledge.
[0060] Step S30: Pre-train the large language model and perform domain-specific fine-tuning using semantic data to output the fine-tuned large language model.
[0061] In this embodiment, step S30 can be implemented through the following steps S301-S303.
[0062] S301: Select a lightweight large language model.
[0063] To achieve a balance between accuracy and speed in industrial scenarios, this embodiment selects a pre-trained large language model with a parameter scale of 7 to 8 billion (7B-8B) as the base model. This large-scale model maintains powerful language understanding and reasoning capabilities while possessing faster reasoning speed and lower computational resource consumption, meeting the real-time requirements of industrial deployment.
[0064] Specific criteria for model selection, as exemplified, include: 1) Fast inference speed: The model structure is efficient and can achieve sub-second or even sub-second response on conventional industrial-grade GPUs.
[0065] 2) Strong Chinese comprehension ability: The pre-training corpus contains high-quality, large-scale Chinese text, ensuring that the model has excellent semantic parsing ability for Chinese process descriptions.
[0066] 3) Industrial deployment friendly: The model should have a mature deployment toolchain (such as vLLM, TensorRT-LLM) and broad community support, so as to facilitate integration into existing production systems.
[0067] In a preferred embodiment, the Llama-3.1-8B-Instruct model is selected as the base model. This model performs well in multiple Chinese and English benchmark tests, and its instruction-following version is particularly suitable for prediction tasks based on semantic prompts.
[0068] S302: LoRA-based domain-specific fine-tuning.
[0069] To adapt the selected general-purpose language model to the highly specialized task of converter temperature prediction, while avoiding excessive fine-tuning costs and overfitting risks, this embodiment employs LoRA (Low-Rank Adaptive) technology for efficient parameter fine-tuning. The core of LoRA fine-tuning lies in freezing the original weights of the pre-trained model and injecting trainable low-rank adapters only next to key layers such as the model's attention mechanism. This significantly reduces the number of trainable parameters (typically by more than 90%), substantially improving training efficiency and reducing hardware requirements.
[0070] S303: Build a prompt template library and an adaptive prompt framework.
[0071] To stabilize the model input format and enhance its expert reasoning capabilities, this embodiment constructs a dedicated prompt template library for converter temperature prediction. This library contains various structured prompt templates used to encapsulate semantic process data during fine-tuning and inference.
[0072] In addition, an adaptive prompting engineering framework was designed. This framework can dynamically select the most suitable prompt template from the template library or make minor adjustments to the wording of the instructions in the template based on the complexity of the input conditions, such as the presence of abnormal parameters or whether it is a special steel type, in order to guide the model to make more accurate predictions.
[0073] Step S40: Embed thermodynamic constraints, kinetic constraints, and process feasibility constraints into the fine-tuned large language model to form a model framework with physical constraints.
[0074] In this embodiment, the thermodynamic constraint includes: Temperature change pattern: Heat of reaction calculate: Thermal balance constraint: Input heat = Output heat + Accumulated heat; The dynamic constraints include: Relationship between decarbonization reaction and temperature: in The carbon content (%) in molten iron. For time, is the decarbonization rate function, representing the functional relationship of the rate of change of carbon content.
[0075] Rate of temperature rise: in For temperature, is the decarbonization rate function, representing the functional relationship of the rate of change of carbon content.
[0076] Minimum reaction time constraint; Temperature range constraint: 1550°C ≤ T ≤ 1750°C; The process feasibility constraints include the temperature distribution range based on historical data; The methods for embedding thermodynamic constraints, kinetic constraints, and process feasibility constraints in the fine-tuned large language model include soft constraints, hard constraints, and adaptive weights; wherein, the soft constraints include adding physical constraint terms to the loss function, the hard constraints include using a projection algorithm to ensure that the prediction results satisfy physical laws, and the adaptive weights include dynamically adjusting the constraint weights according to the characteristics of the operating conditions.
[0077] Step S50: Based on the model framework with physical constraints, construct a four-layer architecture model including an input layer, an encoding layer, a constraint layer, and an output layer, and train it using semantic process data and a composite loss function to output the trained temperature prediction model.
[0078] In this embodiment, the input layer is used to obtain semantic process parameters; the encoding layer is used for large language model feature extraction; the constraint layer is used for physical constraint verification and correction; and the output layer is used to output the predicted endpoint temperature.
[0079] For example, this four-layer architecture consists of an input layer, an encoding layer, a constraint layer, and an output layer. The input layer processes semantic process parameters, achieved through specialized processing of numerical features and fusion of multiple features. The core function of the encoding layer is deep feature extraction using a large language model, specifically implemented through an Llama model combined with LoRA fine-tuning technology, supplemented by feature projection to adapt to downstream tasks. The constraint layer is responsible for crucial physical constraint verification and correction, strictly adhering to three core constraints: first, a temperature range constraint (1550°C ≤ T ≤ 1750°C); second, a thermal balance constraint ensuring system energy conservation (input heat = output heat + accumulated heat); and third, a kinetic constraint controlling the rate of temperature change (dT / dt limit). Finally, the output layer specializes in predicting the final temperature. It uses a dedicated temperature prediction head to ensure all predicted outputs conform to the aforementioned physical constraints, thus obtaining a final result that is both accurate and physically reliable.
[0080] In some embodiments, the composite loss function is designed as follows: in This is due to the loss of accuracy in temperature prediction; Loss due to violation of physical constraints; For semantic consistency loss; physical constraint weight coefficients The value range is from 0.1 to 1.0. Under stable operating conditions, a smaller value (0.1-0.3) is recommended to prioritize prediction accuracy; while under abnormal operating conditions, a larger value (0.5-1.0) should be used to enhance the effect of physical constraints. This physical constraint weighting coefficient... It also supports adaptive adjustment based on the degree of deviation of input parameters. Semantic consistency weight coefficient. The value of is between 0.05 and 0.5. In the early stages of training, a larger value (0.3-0.5) is recommended to enhance semantic understanding; in the later stages of training, it should be gradually reduced to 0.05-0.2 to focus more on optimizing prediction performance. This coefficient ensures logical consistency between the input description and the prediction result.
[0081] Step S60: Perform knowledge distillation and dynamic quantization on the trained temperature prediction model to output an optimized lightweight temperature prediction model.
[0082] For example, knowledge distillation: 7B → 1.5B, while maintaining most of the model's performance; dynamic quantization: FP16 → INT8, significantly improving inference speed.
[0083] Step S70: Input the test data into the optimized lightweight temperature prediction model and output the final temperature prediction result.
[0084] In this embodiment, after substituting the data, the output result of the temperature prediction is as follows: Figure 2 As shown.
[0085] In a specific experiment, the above method was applied, and the accuracy of the predicted results with the actual results was greater than 99%. The test results are shown in Table 1.
[0086] Table 1. Predicted results of converter final temperature
[0087] like Figure 3 The figure shows the convergence curve of the loss function of the converter endpoint temperature prediction model based on a large language model provided in this embodiment of the application during the training process. This figure characterizes the learning process and convergence characteristics of the model on the training set. The horizontal axis represents the number of training steps, and the vertical axis represents the loss value. The figure contains two curves: the original curve reflects the real-time change of the loss value, and the smoothed curve is a smoothed version of the original data to more clearly observe the overall convergence trend. Combined with... Figure 3 Training loss convergence analysis reveals that the model's training loss convergence curve exhibits good learning characteristics. (1) Initial loss value: approximately 1.2 (2) Final training loss: 0.078795 (good convergence) (3) Training duration: approximately 4 hours (14276 seconds) (4) Total number of tokens processed: 9,872,280 The training loss decreased rapidly in the first 200 steps and then converged smoothly without overfitting, demonstrating the effectiveness of the LoRA fine-tuning strategy.
[0088] like Figure 2 As shown, the converter endpoint temperature prediction system interface displays the prediction results, including: Input process conditions display: (1) Pig iron composition: carbon content 4.50%, silicon content 0.40%, manganese content 0.34%, phosphorus content 0.14%, sulfur content 0.020%; (2) Material balance: 86.3 tons of pig iron in blast furnace and 10,497 kg of scrap steel; (3) Process parameters: oxygen blowing time 729 seconds, total oxygen consumption 4462 m³ / h 3 The temperature of pig iron is 1335℃.
[0089] Predicted endpoint temperature output: (1) Final temperature: 1625℃; (2) Temperature conditions: Standard endpoint temperature range; (3) Process assessment: Temperature control is good, suitable for steel tapping; (4) Response time: 3.33 seconds.
[0090] In summary, this application is the first to introduce large language models into the field of converter end-point temperature prediction, breaking through the limitations of traditional numerical modeling. Through core technological innovations such as semantic modeling of process parameters, physical constraint embedding mechanisms, and LoRA-specific fine-tuning, it achieves intelligent conversion of numerical process parameters into natural language descriptions. In the converter steelmaking process, operators, combining their metallurgical expertise, can utilize the intelligent prediction system designed with large language models to analyze process semantic information in real time and predict the end-point temperature, significantly improving the precision control and production efficiency of converter blowing. The accuracy of converter end-point temperature prediction directly affects steel quality and energy consumption control; therefore, accurate temperature forecasting is crucial.
[0091] After implementation, this method can effectively predict the converter's final temperature under different operating conditions, with a prediction accuracy of over 99%. Operators, combining their metallurgical expertise with a semantic prediction and decision-making system designed using a large language model, can understand process condition descriptions in real time, significantly improving the stability, product quality, and production efficiency of converter steelmaking, with a substantial improvement in accuracy compared to traditional methods.
[0092] This application innovatively integrates the semantic understanding capabilities of large language models with the converter steelmaking process, establishing a unified optimization goal of prediction accuracy, physical rationality, and semantic consistency. It achieves intelligent temperature prediction with high precision, high interpretability, and rapid response, which is conducive to improving the level of molten steel quality control, reducing energy consumption, and extending furnace lining life, providing a brand-new technical path for intelligent temperature control in the steel industry.
[0093] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.
Claims
1. A method for intelligent prediction of converter endpoint temperature, characterized in that, The method includes: The process data of the converter steelmaking process is collected and intelligently preprocessed to output preprocessed process data. The process data includes molten iron composition, molten iron temperature, molten iron charging amount, oxygen blowing time, total oxygen consumption, number of furnace turnings, slag splashing time, lance pressing time, scrap steel ratio, amount of various slag-forming materials added, furnace number, steel grade, work group, shift number, smelting timestamp, tapping timestamp, and final temperature. The intelligent preprocessing includes automatic feature type identification and classification, multi-dimensional feature distribution analysis, advanced correlation analysis and temperature feature relationship mining, converter process constraint definition and temperature boundary verification, key process parameter identification and temperature influencing factor analysis, intelligent missing value handling strategy, IQR outlier detection and processing, numerical coding of classification features, data standardization preprocessing, and temperature prediction-oriented data quality verification. The numerical parameters in the preprocessed process data are converted into structured natural language descriptions, and semantic process data is output. A pre-trained large language model is performed, and the semantic data is used for domain-specific fine-tuning to output the fine-tuned large language model. Thermodynamic constraints, kinetic constraints, and process feasibility constraints are embedded in the fine-tuned large language model to form a model framework with physical constraints. Based on the physical constraint model framework, a four-layer architecture model including an input layer, an encoding layer, a constraint layer, and an output layer is constructed, and the semantic process data and composite loss function are used for training to output a trained temperature prediction model. The trained temperature prediction model is subjected to knowledge distillation and dynamic quantization to output an optimized lightweight temperature prediction model. Input the test data into the optimized lightweight temperature prediction model and output the final temperature prediction result.
2. The intelligent prediction method for converter endpoint temperature according to claim 1, characterized in that, The automatic identification and classification of feature types includes: The timestamp data is identified as a time feature; wherein, the timestamp data includes smelting timestamps and steel tapping timestamps; Continuous variables are identified as numerical features; wherein, the continuous variables include molten iron composition, molten iron temperature, molten iron charging amount, oxygen blowing time, total oxygen consumption, number of furnace turnings, slag splashing time, lance pressing time, scrap steel ratio, and the amount of various slag-forming materials added; Discrete variables are identified as categorical features; wherein, the discrete variables include furnace number, steel grade, work group, and shift. The endpoint temperature is identified as the target feature.
3. The intelligent prediction method for converter endpoint temperature according to claim 1, characterized in that, The multidimensional feature distribution analysis includes: Statistical characteristic analysis: Calculate the mean of each feature. μ Standard deviation σ Minimum value (min), maximum value (max), quartiles; Skewness analysis: using skewness coefficient To assess the symmetry of the data distribution, the formula for calculating the skewness coefficient is: in E Represents the expected value, when Skew A value greater than 1 indicates that the data is right-skewed. Skew A skewness of less than -1 indicates a left-skewed distribution, and a skewness close to 0 indicates a relatively symmetrical distribution. The skewness coefficient is used to determine the distribution characteristics of temperature data. Distribution type detection: Shapiro-Wilk normality test was used, and the test statistic was: in x (i) These are the sorted sample values. a i The Shapiro-Wilk coefficient, The sample mean. i n is the sample number, and n is the number of samples; W The test statistic is set to a range of [0,1]. A value closer to 1 indicates that the data conforms more closely to a normal distribution. The calculated W value is compared with the theoretical distribution to obtain the corresponding p value. The p value is based on the W statistic and the sample size n, and is obtained with reference to its specific theoretical distribution. When the p value > 0.05, the normal distribution hypothesis is accepted; when the p value ≤ 0.05, the normal distribution hypothesis is rejected. Missing value proportion assessment: Calculate the missing value rate: in This represents the number of missing values. For the total number of samples, when Delete the feature when it reaches >90%. If the percentage is less than 80%, it will be retained for further processing.
4. The intelligent prediction method for converter endpoint temperature according to claim 1, characterized in that, The advanced correlation analysis and temperature feature relationship mining include: Pearson correlation analysis: Calculating the correlation coefficient r : in and Variables X and Y The mean, X The process parameters are represented by Y, the final temperature is represented by Y, and the correlation coefficient is represented by Y. r The value range of is [-1, 1]. and Representing the i Process parameters and endpoint temperature in each sample n The number of samples; Temperature correlation analysis: For the target of predicting the final temperature, key process parameters that are highly correlated with the final temperature are identified and classified according to their correlation level: |r| > 0.7 is highly correlated, |r| > 0.5 is strongly correlated, |r| > 0.3 is moderately correlated, and |r| ≤ 0.3 is weakly correlated. Highly correlated and strongly correlated features are given priority for use in the temperature prediction model. Strongly correlated feature pair identification: Construct a correlation matrix to automatically identify feature pairs with an absolute correlation coefficient greater than 0.7 for multicollinearity detection.
5. The intelligent prediction method for converter endpoint temperature according to claim 1, characterized in that, The converter process constraint definition and temperature boundary verification include: The composition range of molten iron was set as follows: pig iron C content (3.5, 4.8)%, Si content (0.3, 1.2)%, Mn content (0.3, 1.0)%, P content (0.05, 0.25)%, S content (0.01, 0.06)%; the final composition was C content (0.03, 0.25)%, Si content (0.01, 0.35)%, Mn content (0.10, 0.80)%, P content (0.005, 0.025)%, S content (0.005, 0.025)%. Based on the molten iron temperature range of (1280, 1400) °C, the endpoint temperature range 1 of (1580, 1680) °C, and the endpoint temperature range 2 of (1580, 1680) °C, the predicted results are ensured to meet physical feasibility through a constraint boundary clipping operation: in, This is the endpoint temperature value after physical constraint verification and correction. This represents the lower limit constraint value for temperature. This represents the upper limit constraint value for temperature. The endpoint temperature value predicted by the model; The time constraints are set as follows: oxygen blowing time range (600, 1200) seconds, slag splashing time range (60, 300) seconds, and lance pressing time range (10, 60) seconds; boundary replacement processing is performed when the time parameters are abnormal. Within the range of blast furnace pig iron production (60, 120) tons, scrap steel ratio (0.10, 0.25), and oxygen consumption per cubic meter (45, 65) m³. 3 Based on / t steel, ensure that the material balance meets process requirements: in, This refers to the total mass of materials used in the converter smelting process. To ensure the quality of the molten iron added to the converter, To ensure the quality of the scrap steel added to the converter, The quality of other added materials.
6. The intelligent prediction method for converter endpoint temperature according to claim 1, characterized in that, The identification of key process parameters and analysis of temperature influence factors include: Identify the raw material addition amount characteristics that begin with the first target character as the raw material parameter group. The raw material addition amount characteristics include slag-forming and alloying materials such as limestone, fluorite, and ferrosilicon, and calculate the total amount of raw materials. Identify the first component feature that begins with the second target character as the iron composition parameter group. The first component feature includes the content of C, Si, Mn, P and S elements. Identify the second component feature that begins with the third target character as the endpoint component parameter group to reflect the component control effect of the smelting process; Key operating parameters are identified as a group of process parameters, including oxygen blowing time, total oxygen consumption, number of furnace turnings, slag splashing time, and lance pressing time. Temperature-related parameters are identified as a group of thermal parameters, including pig iron temperature and final temperature.
7. The intelligent prediction method for converter endpoint temperature according to claim 1, characterized in that, The intelligent missing value handling strategy includes: Hierarchical missing value identification: Quantitative evaluation is performed using the missing value rate of each feature. The formula for calculating the missing value rate of each feature is as follows: in The feature missing rate, This represents the number of missing values. Total number of samples Missing value removal strategy: When the number of samples with missing values in the dataset is less than 10% and the missing values are randomly distributed, directly delete the rows containing missing values; if the missing values of a certain feature exceed 80%, delete that feature column to avoid features with insufficient information affecting the temperature prediction effect. Intelligent imputation of numerical features: For numerical features, if the data is normally distributed, the mean is used to impute missing values; when the data is skewed or has outliers, the median is used to impute missing values. Categorical feature imputation: The mode imputation method is used to handle missing discrete variables, and the category with the highest frequency is selected as the imputation value; Target variable handling: Directly delete sample rows with missing values for the endpoint temperature; The IQR outlier detection and processing includes: Outlier boundary calculation: The interquartile range is calculated using the following formula. : in, It is the first quartile. It is the third quartile; The criteria for outlier determination are: When process parameters A value is considered an outlier if it meets the outlier criteria. (2) Outlier Proportion Statistics: Calculate the outlier proportion for each feature using the following formula: in The number of outliers. The total number of samples; Conditional boundary replacement: percentage of outliers When the percentage is less than 5%, the boundary value replacement method is used to replace the outliers with the corresponding boundary values in order to maintain data continuity. Outlier retention strategy: Outlier percentage When the value is ≥ 5%, keep the original value to avoid information loss and prevent over-processing from causing the loss of important process information.
8. The intelligent prediction method for converter endpoint temperature according to claim 1, characterized in that, The domain knowledge-driven feature engineering includes: The total amount of raw materials is used as the comprehensive characteristic of raw materials; The temperature equilibrium characteristics are calculated using the following formula. : in For total input heat, For total heat loss, The total mass of the materials. This represents the average specific heat capacity. The process efficiency characteristic is calculated using the following formula. : in This represents the total oxygen consumption. Oxygen blowing time; The reaction intensity index is calculated using the following formula. : in The temperature at which molten iron enters the furnace. Total amount of materials; The numerical encoding of the classification features includes: Label encoding processing: Low-cardinality categorical features are numerically converted using LabelEncoder, transforming discrete variables such as furnace number, steel grade, and work group into continuous numerical values. Encoding consistency verification: Ensures the stability and reproducibility of encoding mapping, and guarantees that the same encoding rules are used during the training and deployment phases; Encoding feature extension: The processed features are added to the numerical feature set and processed in a unified manner with other continuous variables; The temperature prediction-oriented data quality verification includes Constraint violation detection: Count the number of data points that violate process constraints and calculate the violation rate. : in The number of data points that violate the constraints. Total number of data points Statistical consistency verification: Compare the statistical characteristics of the data before and after processing, and calculate the mean difference rate. : in, This represents the sample mean after data preprocessing. This represents the sample mean of the original data. Temperature distribution consistency check: The focus is on verifying the consistency of the distribution of temperature-related features to ensure that the preprocessed temperature data retains its original physical meaning and statistical regularity.
9. The intelligent prediction method for converter endpoint temperature according to claim 1, characterized in that, The thermodynamic constraints include: Temperature change pattern: Heat of reaction calculate: Thermal balance constraint: Input heat = Output heat + Accumulated heat; The dynamic constraints include: Relationship between decarbonization reaction and temperature: in The carbon content (%) in molten iron. For time, is the decarbonization rate function, representing the functional relationship of the rate of change of carbon content. Rate of temperature rise: in For temperature, is the decarbonization rate function, representing the functional relationship of the rate of change of carbon content. Minimum reaction time constraint; Temperature range constraint: 1550°C ≤ T ≤ 1750°C; The process feasibility constraints include the temperature distribution range based on historical data; The methods for embedding thermodynamic constraints, kinetic constraints, and process feasibility constraints in the fine-tuned large language model include soft constraints, hard constraints, and adaptive weights; wherein, the soft constraints include adding physical constraint terms to the loss function, the hard constraints include using a projection algorithm to ensure that the prediction results satisfy physical laws, and the adaptive weights include dynamically adjusting the constraint weights according to the characteristics of the operating conditions.
10. The intelligent prediction method for converter endpoint temperature according to claim 1, characterized in that, The composite loss function is expressed as follows: in For composite loss function, This is due to the loss of accuracy in temperature prediction; Loss due to violation of physical constraints; This represents the semantic consistency loss. These are the physical constraint weighting coefficients; This is the semantic consistency weight coefficient.