Machine learning based dynamic optimization method for converter blowing process

Abstract

The application discloses a dynamic optimization method for converter blowing process based on machine learning and relates to the technical field of converter blowing.The feature engineering method combining comprehensive scoring of Pearson correlation coefficient and variance inflation factor screening is used to eliminate feature redundancy and multicollinearity problems, select the optimal feature subset which is highly correlated with carbon content deviation and temperature deviation and independent of each other, and enhance the prediction stability of the model; the gradient boosting decision tree algorithm is used and the hyperparameters are iteratively optimized through five-fold cross-validation, so that the model has stronger nonlinear fitting capability, and the prediction accuracy of end-point carbon content and temperature reaches a high level; through the dynamic optimization algorithm combining the construction of a multi-objective weighted optimization function and single-variable disturbance search, the optimal set value is calculated in real time under the constraint condition of multiple operation variables, and the synchronous hit of carbon content and temperature is considered.

Description

Technical Field

[0001] This invention relates to the field of converter blowing technology, specifically to a dynamic optimization method for converter blowing processes based on machine learning. Background Technology

[0002] Converter steelmaking is one of the most important steelmaking methods in modern steel production. Its core process involves blowing high-pressure oxygen into a converter filled with molten iron. The heat generated by the oxidation reaction of oxygen with elements such as carbon, silicon, manganese, and phosphorus in the molten iron completes smelting tasks such as decarburization, dephosphorization, and temperature increase, ultimately yielding molten steel with the required composition and temperature at the end of the blowing process. The converter blowing process is characterized by high temperature, multiphase nature, strong coupling, and nonlinearity. The blowing time is typically only 12 to 20 minutes. Within such a short time, precise control of carbon content and temperature within the target range is crucial, placing extremely high demands on the accuracy and real-time performance of process control.

[0003] Currently, endpoint control in converter blowing processes mainly relies on two categories of methods: static models and dynamic models. Static models, based on initial conditions such as the composition of the molten iron, scrap ratio, target carbon content, and target temperature, calculate setpoints for operating parameters such as oxygen dosage, slag-forming material dosage, and coolant dosage using empirical formulas or regression models, and these values ​​are not adjusted during the entire blowing process. The advantages of this method are its simplicity and ease of implementation, but its biggest drawback is its inability to adjust for the actual reaction conditions during blowing. Due to factors such as fluctuations in the composition of the molten iron, changes in the scrap ratio, and varying degrees of furnace lining erosion, the actual blowing process often deviates significantly from the assumptions of the static model, resulting in a low endpoint hit rate. Typically, the hit rate for carbon content is only 70% to 80%, and the hit rate for temperature is approximately 80% to 85%.

[0004] To overcome the shortcomings of static models, dynamic control methods have emerged. Dynamic control methods acquire intermediate process information during the blowing process through means such as auxiliary lance detection or flue gas analysis, and then adjust subsequent operating parameters accordingly. Traditional dynamic control methods often employ rule-based expert systems or linear regression-based dynamic models. For example, some existing technologies use auxiliary lances to detect the carbon content and temperature of molten steel during the middle of the blowing process, and adjust the oxygen flow rate and coolant addition based on the deviation between the detected values ​​and the target values ​​according to preset correction coefficients. However, these methods have the following significant drawbacks: First, the correction rules are usually linear relationships summarized from human experience, which cannot accurately describe the complex nonlinear coupling relationship between carbon content, temperature, and various operating variables during the blowing process; second, most existing dynamic models only optimize for the two endpoints of carbon content and temperature, without fully considering the synergistic constraints between multiple operating variables such as oxygen injection rate, bottom blowing gas flow rate, oxygen lance position height, and the timing of auxiliary material addition; third, the selection of feature variables in traditional dynamic models often relies on human experience, lacking systematic feature screening and correlation analysis methods, resulting in redundant model input dimensions and limited prediction accuracy. Summary of the Invention

[0005] To address the aforementioned technical problems, this technical solution provides a dynamic optimization method for the converter blowing process based on machine learning, thus resolving the issues raised in the background section.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0007] A machine learning-based dynamic optimization method for converter blowing processes includes the following steps:

[0008] S100: During the converter blowing process, multi-source process data are collected at preset time intervals. The multi-source process data includes secondary lance detection data, flue gas analysis data, spectral analysis data, and oxygen lance operation parameter data. The multi-source process data is then time-stamped and outlier removed to construct the original training dataset.

[0009] S200: Perform feature extraction and feature selection on the original training dataset. Based on the correlation analysis results between each feature and the deviation of carbon content and temperature at the blowing endpoint, retain feature variables with correlation coefficients greater than the preset threshold to form the model input feature vector.

[0010] S300: Using the model input feature vector and the corresponding target value of the blowing endpoint, the gradient boosting decision tree algorithm is used to establish the converter blowing endpoint prediction model, and the hyperparameters of the model are iteratively optimized through the cross-validation method to obtain the trained endpoint prediction model.

[0011] S400: During the current heat blowing process, the multi-source process data collected in real time are processed by the same feature extraction and then input into the endpoint prediction model to obtain the predicted values ​​of carbon content and temperature of molten steel at the current moment.

[0012] S500: Based on the deviation between the predicted carbon content and temperature of molten steel and the corresponding target values, and combined with the dynamic optimization algorithm, the optimal set values ​​of each operating variable between the current time and the next sampling time are calculated. The operating variables include oxygen injection rate, bottom blowing gas flow rate, weight of each auxiliary material added, and oxygen lance position height.

[0013] S600: The optimal setpoint is sent to the converter blowing control system for execution, and steps S400 to S600 are repeated at the next sampling time until the blowing endpoint is reached.

[0014] Preferably, in step S100, the timestamp alignment and outlier removal processing of the multi-source process data specifically includes:

[0015] S101: Using the secondary gun detection time as the reference time node, the flue gas analysis data and spectral analysis data are linearly interpolated and aligned according to the time difference with the reference time node, so that all data have a unified timestamp on the time axis.

[0016] S102: For each aligned variable data point, outlier detection is performed using the three-standard-deviation method. Data points exceeding the mean plus or minus three standard deviations are marked as outliers. The outlier determination criteria are as follows:

[0017] ;

[0018] in, For the first The original data values ​​of each sampling point This is the arithmetic mean of all sampled data points for this variable. The standard deviation of all sampled data for this variable;

[0019] S103: For data points marked as outliers, replace them using a weighted average of two adjacent valid data points. The replacement value is calculated as follows:

[0020] ;

[0021] in, For the replaced number Data values ​​of each sampling point, and The first Two valid data points adjacent to each outlier point and These are the weights of adjacent valid data points, and... , , The time distance between an outlier and the previous valid data point. This represents the time distance between an outlier and the next valid data point.

[0022] S104: Perform sliding window smoothing on the replaced data, with a window length of 5 sampling points, and use the smoothed data as the original training dataset after cleaning.

[0023] Preferably, in step S200, based on the correlation analysis results between each feature and the deviation of carbon content and temperature at the blowing endpoint, feature variables with correlation coefficients greater than a preset threshold are retained, specifically including:

[0024] S201: Calculate the Pearson correlation coefficient between each characteristic variable and the deviation from the endpoint carbon content, and the Pearson correlation coefficient between each characteristic variable and the deviation from the endpoint temperature. The Pearson correlation coefficient is calculated as follows:

[0025] ;

[0026] in, For characteristic variables With target variable The Pearson correlation coefficient between them For the first The value of the feature variable for each sample For the first The target variable values ​​for each sample For characteristic variables The sample mean, For target variable The sample mean, The total number of samples;

[0027] S202: Summing the absolute values ​​of the two correlation coefficients for the same feature variable yields the comprehensive correlation score for that feature variable.

[0028] ;

[0029] in, For the first The overall relevance score of each feature variable. For the first Pearson correlation coefficients between the characteristic variables and the deviation from the endpoint carbon content. For the first Pearson correlation coefficients between the characteristic variables and the endpoint temperature deviation;

[0030] S203: Sort all feature variables from high to low according to their comprehensive correlation scores, and select the top few feature variables with a comprehensive correlation score greater than 0.3 as the candidate feature set;

[0031] S204: Perform variance inflation factor analysis on the characteristic variables in the candidate feature set. The variance inflation factor for each characteristic variable is calculated as follows:

[0032] ;

[0033] in, For the first The determination coefficients are obtained by performing linear regression with one feature variable as the dependent variable and the remaining feature variables as independent variables. When the variance inflation factor between any two feature variables is greater than 10, the feature variable with the lower comprehensive correlation score is removed until the variance inflation factor between all retained feature variables is no greater than 10, and finally the model input feature vector is formed.

[0034] Preferably, in step S300, the model hyperparameters are iteratively optimized using cross-validation, specifically including:

[0035] S301: The original training dataset is randomly divided into a training subset and a validation subset in a 7:3 ratio, where the training subset is used for model training and the validation subset is used for model evaluation.

[0036] S302: Define the hyperparameter search space for the gradient boosting decision tree algorithm. Hyperparameters include the number of trees. Learning rate Maximum tree depth Sampling ratio ;

[0037] S303: Use a random search method to randomly select a preset number of hyperparameter combinations from the hyperparameter search space, perform five-fold cross-validation on each hyperparameter combination, and calculate the average coefficient of determination of the five-fold validation results. and mean absolute percentage error The mean absolute percentage error is calculated as follows:

[0038] ;

[0039] in, For the first The actual endpoint carbon content or endpoint temperature value of each sample For the first The predicted endpoint carbon content or predicted endpoint temperature value for each sample. To verify the total number of samples in the subset;

[0040] S304: Screening for the coefficient of determination Greater than or equal to 0.95 and mean absolute percentage error Combinations of hyperparameters less than or equal to 5% are considered as candidate optimal hyperparameter sets;

[0041] S305: Select the mean absolute percentage error from the candidate optimal hyperparameter set. The smallest set of hyperparameters is used as the final hyperparameters. The model is then retrained using the entire original training dataset with the final hyperparameters to obtain the trained endpoint prediction model.

[0042] Preferably, in step S400, inputting the real-time collected multi-source process data into the endpoint prediction model after undergoing the same feature extraction process specifically includes:

[0043] S401: During the current heat refining process, starting from the middle of the refining process, carbon temperature and oxygen probes are launched at preset time intervals to obtain the actual carbon content and actual temperature of the molten steel at the current moment.

[0044] S402: Simultaneously collect flue gas analyzer data at the current moment to obtain the carbon monoxide and carbon dioxide content in the flue gas, and collect spectral analyzer data to obtain the furnace flame temperature;

[0045] S403: Compare the actual carbon content of molten steel measured by the carbon temperature-oxygen probe with the carbon content measured by the flue gas analyzer, and compare the actual temperature of molten steel measured by the carbon temperature-oxygen probe with the flame temperature measured by the spectrometer. When the deviations of both are within the preset range, the data is confirmed to be valid; otherwise, the data from the carbon temperature-oxygen probe shall be used for correction.

[0046] S404: Process the confirmed valid multi-source process data at the current moment using the same feature extraction method as in step S200 to generate feature vectors with the same format as those in the model training phase;

[0047] S405: Input the feature vector into the endpoint prediction model, and output the predicted values ​​of the carbon content and temperature of the molten steel at the current moment, as well as the prediction confidence intervals corresponding to each output value.

[0048] Preferably, in step S500, calculating the optimal setpoints of each operational variable between the current time and the next sampling time using a dynamic optimization algorithm specifically includes:

[0049] S501: Calculate the first deviation between the predicted carbon content of molten steel and the target carbon content. And the second deviation between the predicted steel temperature and the target temperature. :

[0050] ;

[0051] ;

[0052] in, This is the predicted carbon content of the molten steel at the current moment. The target endpoint is carbon content; This is the predicted temperature of the molten steel at the current moment. The target endpoint temperature;

[0053] S502: When Greater than the preset carbon content deviation threshold or When the temperature deviation exceeds a preset threshold, the objective function is to minimize the weighted sum of squares of the first and second deviations. Starting with the current values ​​of each operand, an iterative search is performed within the constraints of each operand. The objective function is:

[0054]

[0055] in, The objective function value, The weighting coefficient for the carbon content deviation. is the weighting coefficient for temperature deviation, and ;

[0056] S503: In each iteration, the oxygen injection rate is adjusted. Bottom-blown gas flow rate The weight of each auxiliary raw material added and oxygen lance position height Perform a univariate perturbation, calculate the change in the objective function value after the perturbation, and select the perturbation direction that causes the objective function value to decrease the most as the search direction;

[0057] S504: Move along the search direction in preset steps. Move the target function value. If the target function value decreases after the move, continue moving in that direction. If the target function value increases, back off and reduce the step size. Repeat this process until the target function value converges or the maximum number of iterations is reached.

[0058] S505: The values ​​of each operation variable after convergence are used as the optimal setpoints between the current time and the next sampling time.

[0059] Preferably, step S700 is included after step S600. Step S700 is a model self-learning update step, which specifically includes:

[0060] S701: Obtain the actual endpoint carbon content of this heat after the blowing endpoint is reached. and actual endpoint temperature Calculate the predicted deviation of carbon content for this furnace. Temperature prediction deviation :

[0061] ;

[0062] ;

[0063] in, This is the predicted carbon content of the molten steel at the final moment. This is the predicted temperature of the molten steel at the final moment.

[0064] S702: Determining the Prediction Deviation of Carbon Content Whether it is within the preset carbon content deviation range, temperature prediction deviation Whether the temperature deviation is within the preset range, whether the flue gas analysis data is within the preset flue gas composition range, when all three conditions are met, the multi-source process data and actual endpoint values ​​of this furnace will be added as new samples to the original training dataset.

[0065] S703: When the number of new samples reaches the preset cumulative number, model retraining is triggered. Hyperparameter optimization and model training are performed again according to the method in step S300, and the original model is replaced with the updated endpoint prediction model.

[0066] S704: If any of the three conditions is not met, the data of this furnace batch will not be added to the training dataset, and the correspondence between the set value of the operation variable and the actual endpoint value of this furnace batch will be recorded and stored as an abnormal furnace batch sample.

[0067] Preferably, in step S500, the operating variables also include the timing of adding each auxiliary raw material, and calculating the timing of adding each auxiliary raw material specifically includes:

[0068] S501: Calculate the rate of decrease in carbon content based on the trend of the predicted carbon content in molten steel over time, as output by the endpoint prediction model. :

[0069] ;

[0070] in, The current sampling time Predicted carbon content of molten steel The previous sampling time Predicted carbon content of molten steel;

[0071] S502: When the carbon content decreases at a certain rate Less than the preset rate threshold When the current stage is determined to be the weakening stage of the decarbonization reaction, the first batch of coolant is generated to trigger the addition of the first batch of coolant.

[0072] S503: Determine the total amount of coolant to be added based on the deviation between the predicted molten steel temperature and the target temperature. The total amount added will be allocated into multiple batches according to a preset ratio, and the amount added in each batch will not exceed the preset maximum value for a single batch. ;

[0073] S504: Determine the time interval for adding each batch of coolant based on the changing trend of carbon monoxide content in the flue gas analysis data. When the carbon monoxide content is increasing, shorten the time interval; when the carbon monoxide content is decreasing, extend the time interval.

[0074] S505: The timing and amount of coolant added in each batch are considered as part of the optimal setting value for the weight of auxiliary raw materials added.

[0075] Preferably, step S700 is followed by step S800, which is a splash warning and dynamic gun position adjustment step, specifically including:

[0076] S801: During the blowing process, the carbon monoxide content in the flue gas analysis data is acquired in real time. and oxygen content Calculate the ratio of carbon monoxide content to oxygen content. :

[0077] ;

[0078] S802: When the ratio Within the preset time window The slope of the slope continues to decrease and exceeds the preset descent slope threshold. When this is determined to be a pre-splashing state, a splashing warning signal is generated. The calculation method for the descent slope is as follows:

[0079] ;

[0080] in, The end of the time window The ratio, The start time of the time window The ratio when Splash warning triggered at any time;

[0081] S803: In response to the splash warning signal, the oxygen lance position is adjusted to the current position. Increase the preset gun raising range based on the existing parameters. :

[0082] ;

[0083] in, This refers to the height of the oxygen lance position after lifting the lance.

[0084] S804: Preset observation time after drawing the gun If the carbon monoxide and oxygen content in the flue gas analysis data both fall back to the safe range, the oxygen lance position will be gradually lowered back to the normal blowing lance position range; if they do not fall back, the lance will remain raised and a secondary splash warning will be generated.

[0085] S805: In the later stage of blowing, based on the deviation between the predicted carbon content of molten steel output by the endpoint prediction model and the target carbon content, calculate the step-by-step lance lowering amplitude during the carbon removal stage. The calculation method for each step of lowering the gun is as follows:

[0086] ;

[0087] in, This is the current gun position height. To determine the target carbon gun position height, This represents the total number of steps in the carbon extraction stage. This is the deviation adjustment coefficient. The current carbon content prediction bias, The target carbon content is determined until the gun position drops to the range of the carbon extraction gun position.

[0088] A machine learning-based dynamic optimization system for converter blowing processes includes:

[0089] Data Acquisition Module: During the converter blowing process, multi-source process data is collected at preset time intervals. The multi-source process data includes secondary lance detection data, flue gas analysis data, spectral analysis data, and oxygen lance operation parameter data. The multi-source process data is then processed by timestamp alignment and outlier removal to construct the original training dataset.

[0090] Feature extraction module: Performs feature extraction and feature selection on the original training dataset. Based on the correlation analysis results between each feature and the deviation of carbon content and temperature at the blowing endpoint, retains feature variables with correlation coefficients greater than the preset threshold to form the model input feature vector.

[0091] Model building module: Using the model input feature vector and the corresponding target value of the blowing endpoint, the gradient boosting decision tree algorithm is used to build a converter blowing endpoint prediction model, and the hyperparameters of the model are iteratively optimized through cross-validation to obtain the trained endpoint prediction model;

[0092] Prediction module: During the current heat blowing process, the multi-source process data collected in real time are processed by the same feature extraction and then input into the endpoint prediction model to obtain the predicted values ​​of carbon content and temperature of molten steel at the current moment.

[0093] Calculation module: Based on the deviation between the predicted carbon content and temperature of molten steel and the corresponding target values, the module uses a dynamic optimization algorithm to calculate the optimal setpoints of each operational variable between the current time and the next sampling time. The operational variables include oxygen injection rate, bottom blowing gas flow rate, weight of each auxiliary material added, and oxygen lance position height.

[0094] Execution module: Sends the optimal setpoint to the converter blowing control system for execution, and repeats the steps in the above module at the next sampling time until the blowing endpoint is reached.

[0095] Compared with existing technologies, this invention provides a dynamic optimization method for converter blowing processes based on machine learning, which has the following beneficial effects: 1. This invention employs a multi-level data cleaning process combining the three-standard-deviation method with weighted average replacement and sliding window smoothing to effectively remove outliers and noise from multi-source heterogeneous process data, significantly improving the quality of model training data and thus enhancing endpoint prediction accuracy. Furthermore, it utilizes a feature engineering method combining Pearson correlation coefficient comprehensive scoring and variance inflation factor screening to systematically eliminate feature redundancy and multicollinearity, selecting the optimal feature subset that is highly correlated with and independent of both carbon content and temperature deviations, thereby enhancing the model's predictive stability. 2. The gradient boosting decision tree algorithm is adopted and the hyperparameters are iteratively optimized through five-fold cross-validation, which makes the model have a stronger nonlinear fitting ability and the accuracy of the final carbon content and temperature prediction reaches a high level. 3. By constructing a multi-objective weighted optimization function and combining it with a dynamic optimization algorithm based on univariate perturbation search, the optimal setpoint is calculated in real time under multiple operational variable constraints, taking into account the simultaneous hit of carbon content and temperature; 4. A dynamic gun position adjustment strategy that links the splash precursor ratio slope criterion with the endpoint deviation is used to achieve rapid response and automatic recovery under abnormal working conditions. 5. By training a dedicated sub-model through daily data grouping and adaptively selecting the optimal model based on furnace conditions, the problem of insufficient model generalization ability caused by operating condition drift is effectively solved. 6. Through the model self-learning update mechanism, the system can continuously evolve with the accumulation of production data and maintain high-precision control capabilities over the long term. Attached Figure Description

[0096] Figure 1 This is a schematic diagram of the method flow of S100-S600 in this invention;

[0097] Figure 2 This is a schematic diagram of the method flow for S101-S104 in this invention;

[0098] Figure 3 This is a schematic diagram of the method flow for S201-S204 in this invention;

[0099] Figure 4 This is a schematic diagram of the method flow for S301-S305 in this invention;

[0100] Figure 5 This is a schematic diagram of the method flow for S401-S405 in this invention. Detailed Implementation

[0101] The following description is intended to disclose the invention and enable those skilled in the art to implement it. The preferred embodiments described below are merely examples, and other obvious variations will occur to those skilled in the art.

[0102] Example 1

[0103] Please refer to Figure 1 As shown, the dynamic optimization method for the converter blowing process based on machine learning includes the following steps:

[0104] S100: During the converter blowing process, multi-source process data are collected at preset time intervals. The multi-source process data includes secondary lance detection data, flue gas analysis data, spectral analysis data, and oxygen lance operation parameter data. The multi-source process data is then time-stamped and outlier removed to construct the original training dataset.

[0105] S200: Perform feature extraction and feature selection on the original training dataset. Based on the correlation analysis results between each feature and the deviation of carbon content and temperature at the blowing endpoint, retain feature variables with correlation coefficients greater than the preset threshold to form the model input feature vector.

[0106] S300: Using the model input feature vector and the corresponding target value of the blowing endpoint, the gradient boosting decision tree algorithm is used to establish the converter blowing endpoint prediction model, and the hyperparameters of the model are iteratively optimized through the cross-validation method to obtain the trained endpoint prediction model.

[0107] S400: During the current heat blowing process, the multi-source process data collected in real time are processed by the same feature extraction and then input into the endpoint prediction model to obtain the predicted values ​​of carbon content and temperature of molten steel at the current moment.

[0108] S500: Based on the deviation between the predicted carbon content and temperature of molten steel and the corresponding target values, and combined with the dynamic optimization algorithm, the optimal set values ​​of each operating variable between the current time and the next sampling time are calculated. The operating variables include oxygen injection rate, bottom blowing gas flow rate, weight of each auxiliary material added, and oxygen lance position height.

[0109] S600: The optimal setpoint is sent to the converter blowing control system for execution, and steps S400 to S600 are repeated at the next sampling time until the blowing endpoint is reached.

[0110] Those skilled in the art will understand that this invention employs a multi-level data cleaning process combining the three-standard-deviation method with weighted average replacement and sliding window smoothing to effectively remove outliers and noise from multi-source heterogeneous process data, significantly improving the quality of model training data and thus enhancing endpoint prediction accuracy. Through a feature engineering method combining Pearson correlation coefficient comprehensive scoring and variance inflation factor screening, it systematically eliminates feature redundancy and multicollinearity problems, selecting the optimal feature subset that is highly correlated with and independent of both carbon content and temperature deviations, thereby enhancing the model's predictive stability. Finally, it utilizes a gradient boosting decision tree algorithm and iteratively optimizes hyperparameters through five-fold cross-validation, enabling the model to possess stronger nonlinear fitting capabilities. The system achieves high accuracy in predicting force, final carbon content, and temperature. By constructing a multi-objective weighted optimization function and combining it with a dynamic optimization algorithm based on univariate perturbation search, the optimal setpoint is calculated in real-time under multiple operational variable constraints, ensuring simultaneous accuracy in both carbon content and temperature. A dynamic gun position adjustment strategy linking the splash precursor ratio slope criterion with the final deviation enables rapid response and automatic recovery under abnormal operating conditions. By training a dedicated sub-model using daily data groups and adaptively selecting the optimal model based on furnace conditions, the system effectively solves the problem of insufficient model generalization ability caused by operating condition drift. Through a model self-learning update mechanism, the system can continuously evolve with the accumulation of production data, maintaining high-precision control capabilities over the long term.

[0111] Please refer to Figure 2 As shown, in step S100, the multi-source process data undergoes timestamp alignment and outlier removal processing, specifically including:

[0112] S101: Using the secondary gun detection time as the reference time node, the flue gas analysis data and spectral analysis data are linearly interpolated and aligned according to the time difference with the reference time node, so that all data have a unified timestamp on the time axis.

[0113] S102: For each aligned variable data point, outlier detection is performed using the three-standard-deviation method. Data points exceeding the mean plus or minus three standard deviations are marked as outliers. The outlier determination criteria are as follows:

[0114] ;

[0115] in, For the first The original data values ​​of each sampling point This is the arithmetic mean of all sampled data points for this variable. The standard deviation of all sampled data for this variable;

[0116] S103: For data points marked as outliers, replace them using a weighted average of two adjacent valid data points. The replacement value is calculated as follows:

[0117] ;

[0118] in, For the replaced number Data values ​​of each sampling point, and The first Two valid data points adjacent to each outlier point and These are the weights of adjacent valid data points, and... , , The time distance between an outlier and the previous valid data point. This represents the time distance between an outlier and the next valid data point.

[0119] S104: Perform sliding window smoothing on the replaced data, with a window length of 5 sampling points, and use the smoothed data as the original training dataset after cleaning.

[0120] Please refer to Figure 3 As shown, in step S200, based on the correlation analysis results between each feature and the deviation of carbon content and temperature at the blowing endpoint, feature variables with correlation coefficients greater than a preset threshold are retained, specifically including:

[0121] S201: Calculate the Pearson correlation coefficient between each characteristic variable and the deviation from the endpoint carbon content, and the Pearson correlation coefficient between each characteristic variable and the deviation from the endpoint temperature. The Pearson correlation coefficient is calculated as follows:

[0122] ;

[0123] in, For characteristic variables With target variable The Pearson correlation coefficient between them For the first The value of the feature variable for each sample For the first The target variable values ​​for each sample For characteristic variables The sample mean, For target variable The sample mean, The total number of samples;

[0124] S202: Summing the absolute values ​​of the two correlation coefficients for the same feature variable yields the comprehensive correlation score for that feature variable.

[0125] ;

[0126] in, For the first The overall relevance score of each feature variable. For the first Pearson correlation coefficients between the characteristic variables and the deviation from the endpoint carbon content. For the first Pearson correlation coefficients between the characteristic variables and the endpoint temperature deviation;

[0127] S203: Sort all feature variables from high to low according to their comprehensive correlation scores, and select the top few feature variables with a comprehensive correlation score greater than 0.3 as the candidate feature set;

[0128] S204: Perform variance inflation factor analysis on the characteristic variables in the candidate feature set. The variance inflation factor for each characteristic variable is calculated as follows:

[0129] ;

[0130] in, For the first The determination coefficients are obtained by performing linear regression with one feature variable as the dependent variable and the remaining feature variables as independent variables. When the variance inflation factor between any two feature variables is greater than 10, the feature variable with the lower comprehensive correlation score is removed until the variance inflation factor between all retained feature variables is no greater than 10, and finally the model input feature vector is formed.

[0131] Please refer to Figure 4 As shown, in step S300, the model hyperparameters are iteratively optimized using cross-validation, specifically including:

[0132] S301: The original training dataset is randomly divided into a training subset and a validation subset in a 7:3 ratio, where the training subset is used for model training and the validation subset is used for model evaluation.

[0133] S302: Define the hyperparameter search space for the gradient boosting decision tree algorithm. Hyperparameters include the number of trees. Learning rate Maximum tree depth Sampling ratio ;

[0134] S303: Use a random search method to randomly select a preset number of hyperparameter combinations from the hyperparameter search space, perform five-fold cross-validation on each hyperparameter combination, and calculate the average coefficient of determination of the five-fold validation results. and mean absolute percentage error The mean absolute percentage error is calculated as follows:

[0135] ;

[0136] in, For the first The actual endpoint carbon content or endpoint temperature value of each sample For the first The predicted endpoint carbon content or predicted endpoint temperature value for each sample. To verify the total number of samples in the subset;

[0137] S304: Screening for the coefficient of determination Greater than or equal to 0.95 and mean absolute percentage error Combinations of hyperparameters less than or equal to 5% are considered as candidate optimal hyperparameter sets;

[0138] S305: Select the mean absolute percentage error from the candidate optimal hyperparameter set. The smallest set of hyperparameters is used as the final hyperparameters. The model is then retrained using the entire original training dataset with the final hyperparameters to obtain the trained endpoint prediction model.

[0139] Please refer to Figure 5 As shown, in step S400, inputting the real-time collected multi-source process data into the endpoint prediction model after undergoing the same feature extraction process specifically includes:

[0140] S401: During the current heat refining process, starting from the middle of the refining process, carbon temperature and oxygen probes are launched at preset time intervals to obtain the actual carbon content and actual temperature of the molten steel at the current moment.

[0141] S402: Simultaneously collect flue gas analyzer data at the current moment to obtain the carbon monoxide and carbon dioxide content in the flue gas, and collect spectral analyzer data to obtain the furnace flame temperature;

[0142] S403: Compare the actual carbon content of molten steel measured by the carbon temperature-oxygen probe with the carbon content measured by the flue gas analyzer, and compare the actual temperature of molten steel measured by the carbon temperature-oxygen probe with the flame temperature measured by the spectrometer. When the deviations of both are within the preset range, the data is confirmed to be valid; otherwise, the data from the carbon temperature-oxygen probe shall be used for correction.

[0143] S404: Process the confirmed valid multi-source process data at the current moment using the same feature extraction method as in step S200 to generate feature vectors with the same format as those in the model training phase;

[0144] S405: Input the feature vector into the endpoint prediction model, and output the predicted values ​​of the carbon content and temperature of the molten steel at the current moment, as well as the prediction confidence intervals corresponding to each output value.

[0145] In step S500, the optimal setpoints of each operational variable between the current time and the next sampling time are calculated using a dynamic optimization algorithm, specifically including:

[0146] S501: Calculate the first deviation between the predicted carbon content of molten steel and the target carbon content. And the second deviation between the predicted steel temperature and the target temperature. :

[0147] ;

[0148] ;

[0149] in, This is the predicted carbon content of the molten steel at the current moment. The target endpoint is carbon content; This is the predicted temperature of the molten steel at the current moment. The target endpoint temperature;

[0150] S502: When Greater than the preset carbon content deviation threshold or When the temperature deviation exceeds a preset threshold, the objective function is to minimize the weighted sum of squares of the first and second deviations. Starting with the current values ​​of each operand, an iterative search is performed within the constraints of each operand. The objective function is:

[0151]

[0152] in, The objective function value, The weighting coefficient for the carbon content deviation. is the weighting coefficient for temperature deviation, and ;

[0153] S503: In each iteration, the oxygen injection rate is adjusted. Bottom-blown gas flow rate The weight of each auxiliary raw material added and oxygen lance position height Perform a univariate perturbation, calculate the change in the objective function value after the perturbation, and select the perturbation direction that causes the objective function value to decrease the most as the search direction;

[0154] S504: Move along the search direction in preset steps. Move the target function value. If the target function value decreases after the move, continue moving in that direction. If the target function value increases, back off and reduce the step size. Repeat this process until the target function value converges or the maximum number of iterations is reached.

[0155] S505: The values ​​of each operation variable after convergence are used as the optimal setpoints between the current time and the next sampling time.

[0156] Step S600 is followed by step S700, which is a model self-learning update step, specifically including:

[0157] S701: Obtain the actual endpoint carbon content of this heat after the blowing endpoint is reached. and actual endpoint temperature Calculate the predicted deviation of carbon content for this furnace. Temperature prediction deviation :

[0158] ;

[0159] ;

[0160] in, This is the predicted carbon content of the molten steel at the final moment. This is the predicted temperature of the molten steel at the final moment.

[0161] S702: Determining the Prediction Deviation of Carbon Content Whether it is within the preset carbon content deviation range, temperature prediction deviation Whether the temperature deviation is within the preset range, whether the flue gas analysis data is within the preset flue gas composition range, when all three conditions are met, the multi-source process data and actual endpoint values ​​of this furnace will be added as new samples to the original training dataset.

[0162] S703: When the number of new samples reaches the preset cumulative number, model retraining is triggered. Hyperparameter optimization and model training are performed again according to the method in step S300, and the original model is replaced with the updated endpoint prediction model.

[0163] S704: If any of the three conditions is not met, the data of this furnace batch will not be added to the training dataset, and the correspondence between the set value of the operation variable and the actual endpoint value of this furnace batch will be recorded and stored as an abnormal furnace batch sample.

[0164] In step S500, the operational variables also include the timing of adding each auxiliary raw material. The calculation of the timing of adding each auxiliary raw material specifically includes:

[0165] S501: Calculate the rate of decrease in carbon content based on the trend of the predicted carbon content in molten steel over time, as output by the endpoint prediction model. :

[0166] ;

[0167] in, The current sampling time Predicted carbon content of molten steel The previous sampling time Predicted carbon content of molten steel;

[0168] S502: When the carbon content decreases at a certain rate Less than the preset rate threshold When the current stage is determined to be the weakening stage of the decarbonization reaction, the first batch of coolant is generated to trigger the addition of the first batch of coolant.

[0169] S503: Determine the total amount of coolant to be added based on the deviation between the predicted molten steel temperature and the target temperature. The total amount added will be allocated into multiple batches according to a preset ratio, and the amount added in each batch will not exceed the preset maximum value for a single batch. ;

[0170] S504: Determine the time interval for adding each batch of coolant based on the changing trend of carbon monoxide content in the flue gas analysis data. When the carbon monoxide content is increasing, shorten the time interval; when the carbon monoxide content is decreasing, extend the time interval.

[0171] S505: The timing and amount of coolant added in each batch are considered as part of the optimal setting value for the weight of auxiliary raw materials added.

[0172] Step S700 is followed by step S800, which is the splash warning and dynamic gun position adjustment step, specifically including:

[0173] S801: During the blowing process, the carbon monoxide content in the flue gas analysis data is acquired in real time. and oxygen content Calculate the ratio of carbon monoxide content to oxygen content. :

[0174] ;

[0175] S802: When the ratio Within the preset time window The slope of the slope continues to decrease and exceeds the preset descent slope threshold. When this is determined to be a pre-splashing state, a splashing warning signal is generated. The calculation method for the descent slope is as follows:

[0176] ;

[0177] in, The end of the time window The ratio, The start time of the time window The ratio when Splash warning triggered at any time;

[0178] S803: In response to the splash warning signal, the oxygen lance position is adjusted to the current position. Increase the preset gun raising range based on the existing parameters. :

[0179] ;

[0180] in, This refers to the height of the oxygen lance position after lifting the lance.

[0181] S804: Preset observation time after drawing the gun If the carbon monoxide and oxygen content in the flue gas analysis data both fall back to the safe range, the oxygen lance position will be gradually lowered back to the normal blowing lance position range; if they do not fall back, the lance will remain raised and a secondary splash warning will be generated.

[0182] S805: In the later stage of blowing, based on the deviation between the predicted carbon content of molten steel output by the endpoint prediction model and the target carbon content, calculate the step-by-step lance lowering amplitude during the carbon removal stage. The calculation method for each step of lowering the gun is as follows:

[0183] ;

[0184] in, This is the current gun position height. To determine the target carbon gun position height, This represents the total number of steps in the carbon extraction stage. This is the deviation adjustment coefficient. The current carbon content prediction bias, The target carbon content is determined until the gun position drops to the range of the carbon extraction gun position.

[0185] A machine learning-based dynamic optimization system for converter blowing processes includes:

[0186] Data Acquisition Module: During the converter blowing process, multi-source process data is collected at preset time intervals. The multi-source process data includes secondary lance detection data, flue gas analysis data, spectral analysis data, and oxygen lance operation parameter data. The multi-source process data is then processed by timestamp alignment and outlier removal to construct the original training dataset.

[0187] Feature extraction module: Performs feature extraction and feature selection on the original training dataset. Based on the correlation analysis results between each feature and the deviation of carbon content and temperature at the blowing endpoint, retains feature variables with correlation coefficients greater than the preset threshold to form the model input feature vector.

[0188] Model building module: Using the model input feature vector and the corresponding target value of the blowing endpoint, the gradient boosting decision tree algorithm is used to build a converter blowing endpoint prediction model, and the hyperparameters of the model are iteratively optimized through cross-validation to obtain the trained endpoint prediction model;

[0189] Prediction module: During the current heat blowing process, the multi-source process data collected in real time are processed by the same feature extraction and then input into the endpoint prediction model to obtain the predicted values ​​of carbon content and temperature of molten steel at the current moment.

[0190] Calculation module: Based on the deviation between the predicted carbon content and temperature of molten steel and the corresponding target values, the module uses a dynamic optimization algorithm to calculate the optimal setpoints of each operational variable between the current time and the next sampling time. The operational variables include oxygen injection rate, bottom blowing gas flow rate, weight of each auxiliary material added, and oxygen lance position height.

[0191] Execution module: Sends the optimal setpoint to the converter blowing control system for execution, and repeats the steps in the above module at the next sampling time until the blowing endpoint is reached.

[0192] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention. The scope of protection claimed by the appended claims and their equivalents is defined.

Claims

1. A dynamic optimization method for converter blowing process based on machine learning, characterized in that, Includes the following steps: S100: During the converter blowing process, multi-source process data are collected at preset time intervals. The multi-source process data includes secondary lance detection data, flue gas analysis data, spectral analysis data, and oxygen lance operation parameter data. The multi-source process data is then time-stamped and outlier removed to construct the original training dataset. S200: Perform feature extraction and feature selection on the original training dataset. Based on the correlation analysis results between each feature and the deviation of carbon content and temperature at the blowing endpoint, retain feature variables with correlation coefficients greater than the preset threshold to form the model input feature vector. S300: Using the model input feature vector and the corresponding target value of the blowing endpoint, the gradient boosting decision tree algorithm is used to establish the converter blowing endpoint prediction model, and the hyperparameters of the model are iteratively optimized through the cross-validation method to obtain the trained endpoint prediction model. S400: During the current heat blowing process, the multi-source process data collected in real time are processed by the same feature extraction and then input into the endpoint prediction model to obtain the predicted values ​​of carbon content and temperature of molten steel at the current moment. S500: Based on the deviation between the predicted carbon content and temperature of molten steel and the corresponding target values, and combined with the dynamic optimization algorithm, the optimal set values ​​of each operating variable between the current time and the next sampling time are calculated. The operating variables include oxygen injection rate, bottom blowing gas flow rate, weight of each auxiliary material added, and oxygen lance position height. S600: The optimal setpoint is sent to the converter blowing control system for execution, and steps S400 to S600 are repeated at the next sampling time until the blowing endpoint is reached.

2. The dynamic optimization method for converter blowing process based on machine learning according to claim 1, characterized in that, In step S100, the multi-source process data undergoes timestamp alignment and outlier removal processing, specifically including: S101: Using the secondary gun detection time as the reference time node, the flue gas analysis data and spectral analysis data are linearly interpolated and aligned according to the time difference with the reference time node, so that all data have a unified timestamp on the time axis. S102: For each aligned variable data point, outlier detection is performed using the three-standard-deviation method. Data points exceeding the mean plus or minus three standard deviations are marked as outliers. The outlier determination criteria are as follows: ； in, For the first The original data values ​​of each sampling point This is the arithmetic mean of all sampled data points for this variable. The standard deviation of all sampled data points for this variable; S103: For data points marked as outliers, replace them using a weighted average of two adjacent valid data points. The replacement value is calculated as follows: ； in, For the replaced number Data values ​​of each sampling point, and The first Two valid data points adjacent to each outlier point and These are the weights of adjacent valid data points, and... , , The time distance between an outlier and the previous valid data point. This represents the time distance between an outlier and the next valid data point. S104: Perform sliding window smoothing on the replaced data, with a window length of 5 sampling points, and use the smoothed data as the original training dataset after cleaning.

3. The dynamic optimization method for converter blowing process based on machine learning according to claim 2, characterized in that, In step S200, based on the correlation analysis results between each feature and the deviation of carbon content and temperature at the blowing endpoint, feature variables with correlation coefficients greater than a preset threshold are retained, specifically including: S201: Calculate the Pearson correlation coefficient between each characteristic variable and the deviation from the endpoint carbon content, and the Pearson correlation coefficient between each characteristic variable and the deviation from the endpoint temperature. The Pearson correlation coefficient is calculated as follows: ； in, For characteristic variables With target variable The Pearson correlation coefficient between them For the first The value of the feature variable for each sample For the first The target variable values ​​for each sample For characteristic variables The sample mean, For target variable The sample mean, The total number of samples; S202: Summing the absolute values ​​of the two correlation coefficients for the same feature variable yields the comprehensive correlation score for that feature variable. ； in, For the first The overall relevance score of each feature variable. For the first Pearson correlation coefficients between the characteristic variables and the deviation from the endpoint carbon content. For the first Pearson correlation coefficients between the characteristic variables and the endpoint temperature deviation; S203: Sort all feature variables from high to low according to their comprehensive correlation scores, and select the top few feature variables with a comprehensive correlation score greater than 0.3 as the candidate feature set; S204: Perform variance inflation factor analysis on the characteristic variables in the candidate feature set. The variance inflation factor for each characteristic variable is calculated as follows: ； in, For the first The determination coefficients are obtained by performing linear regression with one feature variable as the dependent variable and the remaining feature variables as independent variables. When the variance inflation factor between any two feature variables is greater than 10, the feature variable with the lower comprehensive correlation score is removed until the variance inflation factor between all retained feature variables is no greater than 10, and finally the model input feature vector is formed.

4. The dynamic optimization method for converter blowing process based on machine learning according to claim 3, characterized in that, In step S300, the model hyperparameters are iteratively optimized using cross-validation, specifically including: S301: The original training dataset is randomly divided into a training subset and a validation subset in a 7:3 ratio, where the training subset is used for model training and the validation subset is used for model evaluation. S302: Define the hyperparameter search space for the gradient boosting decision tree algorithm. Hyperparameters include the number of trees. Learning rate Maximum tree depth Sampling ratio ; S303: Use a random search method to randomly select a preset number of hyperparameter combinations from the hyperparameter search space, perform five-fold cross-validation on each hyperparameter combination, and calculate the average coefficient of determination of the five-fold validation results. and mean absolute percentage error The mean absolute percentage error is calculated as follows: ； in, For the first The actual endpoint carbon content or endpoint temperature value of each sample For the first The predicted endpoint carbon content or predicted endpoint temperature value for each sample. To verify the total number of samples in the subset; S304: Screening for the coefficient of determination Greater than or equal to 0.95 and mean absolute percentage error Combinations of hyperparameters less than or equal to 5% are considered as candidate optimal hyperparameter sets; S305: Select the mean absolute percentage error from the candidate optimal hyperparameter set. The smallest set of hyperparameters is used as the final hyperparameters. The model is then retrained using the entire original training dataset with the final hyperparameters to obtain the trained endpoint prediction model.

5. The dynamic optimization method for converter blowing process based on machine learning according to claim 4, characterized in that, In step S400, inputting the real-time collected multi-source process data into the endpoint prediction model after undergoing the same feature extraction process specifically includes: S401: During the current heat refining process, starting from the middle of the refining process, carbon temperature and oxygen probes are launched at preset time intervals to obtain the actual carbon content and actual temperature of the molten steel at the current moment. S402: Simultaneously collect flue gas analyzer data at the current moment to obtain the carbon monoxide and carbon dioxide content in the flue gas, and collect spectral analyzer data to obtain the furnace flame temperature; S403: Compare the actual carbon content of molten steel measured by the carbon temperature-oxygen probe with the carbon content measured by the flue gas analyzer, and compare the actual temperature of molten steel measured by the carbon temperature-oxygen probe with the flame temperature measured by the spectrometer. When the deviations of both are within the preset range, the data is confirmed to be valid; otherwise, the data from the carbon temperature-oxygen probe shall be used for correction. S404: Process the confirmed valid multi-source process data at the current moment using the same feature extraction method as in step S200 to generate feature vectors with the same format as those in the model training phase; S405: Input the feature vector into the endpoint prediction model, and output the predicted values ​​of the carbon content and temperature of the molten steel at the current moment, as well as the prediction confidence intervals corresponding to each output value.

6. The dynamic optimization method for converter blowing process based on machine learning according to claim 5, characterized in that, In step S500, the optimal setpoint values ​​of each operational variable between the current time and the next sampling time are calculated using a dynamic optimization algorithm, specifically including: S501: Calculate the first deviation between the predicted carbon content of molten steel and the target carbon content. And the second deviation between the predicted steel temperature and the target temperature. : ； ； in, This is the predicted carbon content of the molten steel at the current moment. The target endpoint is carbon content; This is the predicted temperature of the molten steel at the current moment. The target endpoint temperature; S502: When Greater than the preset carbon content deviation threshold or When the temperature deviation exceeds a preset threshold, the objective function is to minimize the weighted sum of squares of the first and second deviations. Starting with the current values ​​of each operand, an iterative search is performed within the constraints of each operand. The objective function is: ； in, The objective function value, The weighting coefficient for carbon content deviation. is the weighting coefficient for temperature deviation, and ; S503: In each iteration, the oxygen injection rate is adjusted. Bottom-blown gas flow rate The weight of each auxiliary raw material added and oxygen lance position height Perform a univariate perturbation, calculate the change in the objective function value after the perturbation, and select the perturbation direction that causes the objective function value to decrease the most as the search direction; S504: Move along the search direction in preset steps. Move the target function value. If the target function value decreases after the move, continue moving in that direction. If the target function value increases, back off and reduce the step size. Repeat this process until the target function value converges or the maximum number of iterations is reached. S505: The values ​​of each operation variable after convergence are used as the optimal setpoints between the current time and the next sampling time.

7. The dynamic optimization method for converter blowing process based on machine learning according to claim 6, characterized in that, Step S600 is followed by step S700, which is a model self-learning update step, specifically including: S701: Obtain the actual endpoint carbon content of this heat after the blowing endpoint is reached. and actual endpoint temperature Calculate the predicted deviation of carbon content for this furnace. Temperature prediction deviation : ； ； in, This represents the predicted carbon content of the molten steel at the final moment. This is the predicted temperature of the molten steel at the final moment; S702: Determining the Prediction Deviation of Carbon Content Whether it is within the preset carbon content deviation range, temperature prediction deviation Whether the temperature deviation is within the preset range, whether the flue gas analysis data is within the preset flue gas composition range, when all three conditions are met, the multi-source process data and actual endpoint values ​​of this furnace will be added as new samples to the original training dataset. S703: When the number of new samples reaches the preset cumulative number, model retraining is triggered. Hyperparameter optimization and model training are performed again according to the method in step S300, and the original model is replaced with the updated endpoint prediction model. S704: If any of the three conditions is not met, the data of this furnace batch will not be added to the training dataset, and the correspondence between the set value of the operation variable and the actual endpoint value of this furnace batch will be recorded and stored as an abnormal furnace batch sample.

8. The dynamic optimization method for converter blowing process based on machine learning according to claim 7, characterized in that, In step S500, the operational variables also include the timing of adding each auxiliary raw material. Calculating the timing of adding each auxiliary raw material specifically includes: S501: Calculate the rate of decrease in carbon content based on the trend of the predicted carbon content in molten steel over time, as output by the endpoint prediction model. : ； in, The current sampling time Predicted carbon content of molten steel The previous sampling time Predicted carbon content of molten steel; S502: When the carbon content decreases at a certain rate Less than the preset rate threshold When the current stage is determined to be the weakening stage of the decarbonization reaction, the first batch of coolant is generated to trigger the addition of the first batch of coolant. S503: Determine the total amount of coolant to be added based on the deviation between the predicted molten steel temperature and the target temperature. The total amount added will be allocated into multiple batches according to a preset ratio, and the amount added in each batch will not exceed the preset maximum value for a single batch. ; S504: Determine the time interval for adding each batch of coolant based on the changing trend of carbon monoxide content in the flue gas analysis data. When the carbon monoxide content is increasing, shorten the time interval; when the carbon monoxide content is decreasing, extend the time interval. S505: The timing and amount of coolant added in each batch are considered as part of the optimal setting value for the weight of auxiliary raw materials added.

9. The dynamic optimization method for converter blowing process based on machine learning according to claim 8, characterized in that, The process is followed by step S800 after step S700. Step S800 is a splash warning and dynamic gun position adjustment step, specifically including: S801: During the blowing process, the carbon monoxide content in the flue gas analysis data is acquired in real time. and oxygen content Calculate the ratio of carbon monoxide content to oxygen content. : ； S802: When the ratio Within the preset time window The slope of the slope continues to decrease and exceeds the preset descent slope threshold. When this is determined to be a pre-splashing state, a splashing warning signal is generated. The calculation method for the descent slope is as follows: ； in, The end of the time window The ratio, The start time of the time window The ratio when Splash warning triggered at any time; S803: In response to the splash warning signal, the oxygen lance position is adjusted to the current position. Increase the preset gun raising range based on the existing parameters. : ； in, This refers to the height of the oxygen lance position after lifting the lance. S804: Preset observation time after drawing the gun If the carbon monoxide and oxygen content in the flue gas analysis data both fall back to the safe range, the oxygen lance position will be gradually lowered back to the normal blowing lance position range; if they do not fall back, the lance will remain raised and a secondary splash warning will be generated. S805: In the later stage of blowing, based on the deviation between the predicted carbon content of molten steel output by the endpoint prediction model and the target carbon content, calculate the step-by-step lance lowering amplitude during the carbon removal stage. The calculation method for each step of lowering the gun is as follows: ； in, This is the current gun position height. To determine the target carbon gun position height, This represents the total number of steps in the carbon extraction stage. This is the deviation adjustment coefficient. The current carbon content prediction bias, The target carbon content is determined until the gun position drops to the range of the carbon extraction gun position.

10. A machine learning-based dynamic optimization system for converter blowing processes, used to implement the machine learning-based dynamic optimization method for converter blowing processes as described in any one of claims 1-9, characterized in that, include: Data Acquisition Module: During the converter blowing process, multi-source process data is collected at preset time intervals. The multi-source process data includes secondary lance detection data, flue gas analysis data, spectral analysis data, and oxygen lance operation parameter data. The multi-source process data is then processed by timestamp alignment and outlier removal to construct the original training dataset. Feature extraction module: Performs feature extraction and feature selection on the original training dataset. Based on the correlation analysis results between each feature and the deviation of carbon content and temperature at the blowing endpoint, retains feature variables with correlation coefficients greater than the preset threshold to form the model input feature vector. Model building module: Using the model input feature vector and the corresponding target value of the blowing endpoint, the gradient boosting decision tree algorithm is used to build a converter blowing endpoint prediction model, and the hyperparameters of the model are iteratively optimized through cross-validation to obtain the trained endpoint prediction model; Prediction module: During the current heat blowing process, the multi-source process data collected in real time are processed by the same feature extraction and then input into the endpoint prediction model to obtain the predicted values ​​of carbon content and temperature of molten steel at the current moment. Calculation module: Based on the deviation between the predicted carbon content and temperature of molten steel and the corresponding target values, the module uses a dynamic optimization algorithm to calculate the optimal setpoints of each operational variable between the current time and the next sampling time. The operational variables include oxygen injection rate, bottom blowing gas flow rate, weight of each auxiliary material added, and oxygen lance position height. Execution module: Sends the optimal setpoint to the converter blowing control system for execution, and repeats the steps in the above module at the next sampling time until the blowing endpoint is reached.

Images