A data-driven dynamic secondary cooling water distribution method for slab continuous casting

Abstract

The present invention discloses a data-driven dynamic secondary cooling water distribution method for slab continuous casting, which belongs to the field of metallurgical technology. The method realizes water quantity prediction of each section of the secondary cooling zone through temperature prediction and water distribution prediction. The temperature prediction is achieved by pre-processing characteristic parameter I and inputting it into the Transformer model in the order of time series. The model captures the time dependency in the sequence and the interaction between features by encoding characteristic parameter I and learning the self-attention mechanism, thereby realizing accurate prediction of the temperature of the position point in the secondary cooling zone. The water distribution prediction is achieved by using the inherent law and correlation of characteristic parameter II through ensemble learning of random forest to realize prediction of the water quantity of the section to be predicted in the secondary cooling zone, realize accurate distribution of water quantity in each section of the secondary cooling zone, optimize the continuous casting process, and improve the quality of the casting and production efficiency.

Description

Technical Field

[0001] The present invention belongs to the field of metallurgy technology, and in particular relates to a data-driven dynamic secondary cooling water distribution method for slab continuous casting. Background Art

[0002] Continuous casting is a crucial process for billet formation and quality assurance. It also serves as the core of the long steelmaking process, driving the dynamic, orderly, coordinated, and continuous operation of the entire process. Continuous casting is essentially the process by which molten steel releases superheat, latent heat, and physical sensible heat to solidify into a billet. The molten steel must release approximately 50% of its heat before fully solidifying as it passes through the mold, secondary cooling zone, and air cooling zone. The cooling intensity of the mold and secondary cooling zone has a significant impact on continuous casting efficiency and billet quality, particularly during non-steady-state conditions such as the start-up, final pour, and tundish changes, where casting speed and cooling water levels fluctuate significantly. During the continuous casting process, the intensity, method, and distribution and control of the secondary cooling are crucial factors affecting billet quality. Most surface defects occur in the mold, while other defects such as cracks, desquamation, and bulging occur in the secondary cooling stage. Molten steel forms a shell of a certain thickness in the mold to prevent the billet from leaking out. After entering the secondary cooling zone, the billet is sprayed with cooling water to accelerate heat transfer within the billet, promoting gradual solidification. The primary function of secondary cooling is to accelerate the solidification process of the ingot, further cooling the majority of the unsolidified ingot emerging from the mold. This ensures ingot quality and minimizes surface and internal defects caused by improper secondary cooling. Currently, water distribution in the secondary cooling zone during continuous casting control suffers from inaccurate water regulation and insufficient stability. The stability and accuracy of water distribution in the secondary cooling zone are closely related to ingot quality. Summary of the Invention

[0003] In view of the above-mentioned problems, the present invention discloses a data-driven dynamic secondary cooling water distribution method for slab continuous casting.

[0004] The present invention adopts the following technical solutions:

[0005] A data-driven dynamic secondary cooling water distribution method for slab continuous casting, the water distribution method is used to predict the water volume in each section of the secondary cooling zone, and the dynamic secondary cooling water distribution method includes a temperature prediction method and a secondary cooling water distribution prediction method;

[0006] According to the different distances between the second cooling zone and the meniscus, a plurality of position points are sequentially set on the second cooling zone, and the position points divide the second cooling zone into a plurality of segments, and a segment is formed between two adjacent position points. The position point of the current segment close to one end of the meniscus is the previous position point of the current segment, and the position point of the other end of the current segment is the next position point of the current segment;

[0007] The segment adjacent to the current segment in the meniscus direction is the previous segment, and the previous position point of the current segment is the next position point of the previous segment;

[0008] The segment adjacent to the current segment in the opposite direction of the meniscus is the next segment, and the next position point of the current segment is the previous position point of the next segment;

[0009] The temperature prediction method includes: first preprocessing feature parameter I, then inputting feature parameter I into a Transformer model in the order of a time series, wherein the Transformer model captures the temporal dependency and the interaction between features in the time series by encoding feature parameter I and learning a self-attention mechanism, thereby obtaining a predicted value of the temperature at the next position point in the current section, wherein feature parameter I is a parameter affecting the temperature of the second cold zone, and the predicted value of the temperature at the next position point is obtained by predicting the temperature at the previous position point in the current section using the temperature prediction method;

[0010] The secondary cooling water distribution prediction method includes: through the integrated learning of random forests, using the inherent rules and correlations of characteristic parameters II, to predict the water volume of the to-be-predicted section of the secondary cooling zone, the characteristic parameters II include section specifications, pulling speed and superheat, the characteristic parameters II also include the previous position point temperature and the next position point temperature corresponding to the to-be-predicted section of the secondary cooling zone, the next position point temperature is predicted by the temperature prediction method, and the water volume of the to-be-predicted section of the secondary cooling zone is predicted based on the characteristic parameters II.

[0011] Furthermore, the characteristic parameter I includes: cross-sectional specifications, pulling speed, superheat, distance, temperature and water volume, the cross-sectional specifications are the width and length of the ingot, the pulling speed is the speed at which the ingot is pulled out of the crystallizer, the superheat is the difference between the pouring temperature of the molten steel and its liquidus temperature, the distance is the distance between each section of the secondary cooling zone and the meniscus, the temperature is the temperature of the previous position point, the temperature of the previous position point is the center temperature of the ingot at that position point, and the water volume is the water volume of the secondary cooling zone section.

[0012] Furthermore, the temperature prediction method includes:

[0013] S1.1 Collect relevant characteristic parameters of the continuous casting machine production, including distance, temperature, and water volume. The distance is the distance between each section of the secondary cooling zone and the meniscus. The temperature is the center temperature of the position of the billet corresponding to the position in the secondary cooling zone. The water volume is the water volume in each section of the secondary cooling zone.

[0014] S1.2 Determine characteristic parameters of the temperature prediction method I;

[0015] S1.3 inputs the feature parameter I into the Transformer model in the order of the time series. The Transformer model captures the time dependency and interaction between features in the sequence by encoding the feature parameter I and learning the self-attention mechanism, and obtains the predicted value of the temperature at the next location point.

[0016] Furthermore, S1.3 includes:

[0017] S1.3.1 Data preprocessing: preprocess characteristic parameter I to ensure that characteristic parameter I meets the requirements of the model;

[0018] S1.3.2 Build a Transformer model, which includes an encoder and an output layer. The encoder uses multiple stacked Transformer encoding layers, each of which includes a self-attention mechanism and a feedforward neural network. The output layer converts the encoder output into a final predicted temperature value.

[0019] S1.3.3 Training process: The Transformer model is trained using a training dataset. The Transformer model first calculates the predicted value through forward propagation, compares it with the true value, calculates the loss function, and then updates the model parameters through the backpropagation algorithm to minimize the loss function.

[0020] S1.3.4 Model evaluation and tuning: Evaluate the performance of the Transformer model on the validation set or test set, use mean square error, root mean square error, and mean absolute error metrics to evaluate the model's prediction accuracy, and tune the model based on the evaluation results;

[0021] S1.3.5 Prediction and Application: Use the trained Transformer model to perform temperature prediction. In actual production, by continuously inputting new time series data, real-time updated prediction results are obtained.

[0022] Furthermore, the preprocessing described in S1.3.1 includes feature scaling and sequence filling.

[0023] Furthermore, the training data set described in S1.3.3 is composed of the related feature parameters.

[0024] Furthermore, the tuning described in S1.3.4 includes adjusting model parameters and changing the model structure.

[0025] Furthermore, the secondary cooling water distribution prediction method includes:

[0026] S2.1 Determine characteristic parameters II of the secondary cooling water distribution prediction method, which include section specifications, casting speed, superheat, previous point temperature, and next point temperature;

[0027] S2.2 inputs characteristic parameter II into the random forest model. Through the integrated learning of the random forest, the random forest model uses the inherent rules and correlations of the input data to obtain the water volume between the previous position point and the next position point, thereby realizing the prediction of the water volume in each section of the secondary cooling zone.

[0028] Furthermore, S2.2 includes:

[0029] S2.2.1 Data preprocessing: Preprocess feature parameter II. This preprocessing includes feature scaling and sequence filling to ensure that the input data meets the requirements of the model.

[0030] S2.2.2 Model construction: Build a random forest regression model in the Scikit-learn library and set parameters to control the behavior of the random forest;

[0031] S2.2.3 training the model, using the training data set to train the random forest regression model; the random forest regression model will construct multiple decision trees based on the samples in the training set and the value of the target variable, and perform feature selection and partitioning on each tree;

[0032] S2.2.4 Prediction results: Use the trained random forest regression model to predict the samples in the test set; the trained random forest regression model averages or weighted averages the prediction results of each decision tree to obtain the final regression prediction result;

[0033] S2.2.5 Model Evaluation and Tuning: Evaluate the performance of the trained random forest regression model on the validation set or test set, using the mean square error, root mean square error, and mean absolute error metrics to assess the prediction accuracy of the trained random forest regression model. Tune the model based on the evaluation results, including adjusting model parameters and changing the model structure.

[0034] S2.2.6 Prediction and application: Use the trained random forest regression model to predict the water volume in the second cooling zone.

[0035] Furthermore, the parameters described in S2.2.2 include: the number of decision trees, feature selection method and decision tree growth method.

[0036] Beneficial effects:

[0037] The present invention discloses a data-driven dynamic secondary cooling water distribution method for slab continuous casting, which makes full use of the data accumulated during the production process, mines the inherent laws and correlations in the production data, and realizes precise control of the continuous casting process. The present invention constructs a prediction method that first predicts the temperature of each position point based on the Transform regression prediction method, and then accurately predicts the water volume of each section of the secondary cooling zone based on the random forest regression prediction and the above-mentioned predicted temperature of each position point. The method is trained based on a large amount of actual production data, and can accurately predict the temperature of the position point of the secondary cooling zone and distribute the water volume of the secondary cooling zone section. In actual production, it can significantly improve the quality of the casting, and the accuracy and stability of the secondary cooling zone water distribution are significantly improved compared with the existing methods. At the same time, the prediction method is based on relevant characteristic parameters, characteristic parameters I and characteristic parameters II, is not limited by fixed parameters and expert experience, and can automatically adapt to changes in the production process. This method provides a reference for dynamic secondary cooling water distribution during continuous casting, can effectively optimize the continuous casting process, and improve the quality of casting and production efficiency.

[0038] The prediction model in the present invention adopts the leave-one-out method for cross-validation. This verification method maximizes the use of all data for training and verification, reducing the overfitting problem caused by the small data size. At the same time, each sample will be used as a validation set, thereby reducing the problem of unstable evaluation results caused by uneven sample division, and can more accurately evaluate the generalization ability of the model on the entire dataset. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0040] Figure 1 This is a flow chart of a data-driven dynamic secondary cooling water distribution method for slab continuous casting according to the present invention;

[0041] Figure 2 This is a schematic diagram of a temperature prediction process of a data-driven slab continuous casting dynamic secondary cooling water distribution method according to the present invention;

[0042] Figure 3 This is a schematic diagram of prediction results of a temperature prediction method for a data-driven slab continuous casting dynamic secondary cooling water distribution method according to the present invention;

[0043] Figure 4 It is a schematic diagram of prediction results of a secondary cooling water distribution prediction method of a data-driven slab continuous casting dynamic secondary cooling water distribution method according to the present invention. DETAILED DESCRIPTION

[0044] In order to better understand the technical solution of the present invention, the embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0045] It should be understood that the embodiments described are only a portion of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by persons of ordinary skill in the art without creative work are within the scope of protection of the present invention.

[0046] Example 1

[0047] A data-driven dynamic secondary cooling water distribution method for slab continuous casting, the water distribution method is used to predict the water volume in each section of the secondary cooling zone, and the dynamic secondary cooling water distribution method includes a temperature prediction method and a secondary cooling water distribution prediction method;

[0048] According to the different distances between the second cooling zone and the meniscus, a plurality of position points are sequentially set on the second cooling zone, and the position points divide the second cooling zone into a plurality of segments, and a segment is formed between two adjacent position points. The position point of the current segment close to one end of the meniscus is the previous position point of the current segment, and the position point of the other end of the current segment is the next position point of the current segment;

[0049] The segment adjacent to the current segment in the meniscus direction is the previous segment, and the previous position point of the current segment is the next position point of the previous segment;

[0050] The segment adjacent to the current segment in the opposite direction of the meniscus is the next segment, and the next position point of the current segment is the previous position point of the next segment;

[0051] The temperature prediction method includes: first preprocessing feature parameter I, then inputting feature parameter I into a Transformer model in the order of a time series, wherein the Transformer model captures the temporal dependency and the interaction between features in the time series by encoding feature parameter I and learning a self-attention mechanism, thereby obtaining a predicted value of the temperature at the next position point in the current section, wherein feature parameter I is a parameter affecting the temperature of the second cold zone, and the predicted value of the temperature at the next position point is obtained by predicting the temperature at the previous position point in the current section using the temperature prediction method;

[0052] The secondary cooling water distribution prediction method includes: through the integrated learning of random forests, using the inherent rules and correlations of characteristic parameters II, to predict the water volume of the to-be-predicted section of the secondary cooling zone, the characteristic parameters II include section specifications, pulling speed and superheat, the characteristic parameters II also include the previous position point temperature and the next position point temperature corresponding to the to-be-predicted section of the secondary cooling zone, the next position point temperature is predicted by the temperature prediction method, and the water volume of the to-be-predicted section of the secondary cooling zone is predicted based on the characteristic parameters II.

[0053] Furthermore, the characteristic parameter I includes: cross-sectional specifications, pulling speed, superheat, distance, temperature and water volume, the cross-sectional specifications are the width and length of the ingot, the pulling speed is the speed at which the ingot is pulled out of the crystallizer, the superheat is the difference between the pouring temperature of the molten steel and its liquidus temperature, the distance is the distance between each section of the secondary cooling zone and the meniscus, the temperature is the temperature of the previous position point, the temperature of the previous position point is the center temperature of the ingot at that position point, and the water volume is the water volume of the secondary cooling zone section.

[0054] Furthermore, the temperature prediction method includes:

[0055] S1.1 Collect relevant characteristic parameters of the continuous casting machine production, including distance, temperature, and water volume. The distance is the distance between each section of the secondary cooling zone and the meniscus. The temperature is the center temperature of the position of the billet corresponding to the position in the secondary cooling zone. The water volume is the water volume in each section of the secondary cooling zone.

[0056] S1.2 Determine characteristic parameters of the temperature prediction method I;

[0057] S1.3 Input the feature parameter I into the Transformer model in the order of the time series. The Transformer model encodes the feature parameter I and learns the self-attention mechanism to capture the time dependency and interaction between features in the sequence, and obtain the predicted value of the temperature at the next location. Figure 1 shown.

[0058] Furthermore, S1.3 includes:

[0059] S1.3.1 Data preprocessing: preprocess characteristic parameter I to ensure that characteristic parameter I meets the requirements of the model;

[0060] S1.3.2 Build a Transformer model, which includes an encoder and an output layer. The encoder uses multiple stacked Transformer encoding layers, each of which includes a self-attention mechanism and a feedforward neural network. The output layer converts the encoder output into a final predicted temperature value.

[0061] S1.3.3 Training process: The Transformer model is trained using a training dataset. The Transformer model first calculates the predicted value through forward propagation, compares it with the true value, calculates the loss function, and then updates the model parameters through the backpropagation algorithm to minimize the loss function.

[0062] S1.3.4 Model evaluation and tuning: Evaluate the performance of the Transformer model on the validation set or test set, use mean square error, root mean square error, and mean absolute error metrics to evaluate the model's prediction accuracy, and tune the model based on the evaluation results;

[0063] S1.3.5 Prediction and Application: Use the trained Transformer model to perform temperature prediction. In actual production, by continuously inputting new time series data, real-time updated prediction results are obtained.

[0064] Furthermore, the preprocessing described in S1.3.1 includes feature scaling and sequence filling.

[0065] Furthermore, the training data set described in S1.3.3 is composed of the related feature parameters.

[0066] Furthermore, the tuning described in S1.3.4 includes adjusting model parameters and changing the model structure.

[0067] Furthermore, the secondary cooling water distribution prediction method includes:

[0068] S2.1 Determine characteristic parameters II of the secondary cooling water distribution prediction method, which include section specifications, casting speed, superheat, previous point temperature, and next point temperature;

[0069] S2.2 Input the characteristic parameter II into the random forest model. Through the ensemble learning of random forest, the random forest model uses the inherent rules and correlations of the input data to obtain the water volume between the previous location point and the next location point, and realizes the prediction of the water volume of each section of the secondary cooling zone. Figure 1 shown.

[0070] Furthermore, S2.2 includes:

[0071] S2.2.1 Data preprocessing: Preprocess feature parameter II. This preprocessing includes feature scaling and sequence filling to ensure that the input data meets the requirements of the model.

[0072] S2.2.2 Model construction: Build a random forest regression model in the Scikit-learn library and set parameters to control the behavior of the random forest;

[0073] S2.2.3 training the model, using the training data set to train the random forest regression model; the random forest regression model will construct multiple decision trees based on the samples in the training set and the value of the target variable, and perform feature selection and partitioning on each tree;

[0074] S2.2.4 Prediction results: Use the trained random forest regression model to predict the samples in the test set; the trained random forest regression model averages or weighted averages the prediction results of each decision tree to obtain the final regression prediction result;

[0075] S2.2.5 Model Evaluation and Tuning: Evaluate the performance of the trained random forest regression model on the validation set or test set, using the mean square error, root mean square error, and mean absolute error metrics to assess the prediction accuracy of the trained random forest regression model. Tune the model based on the evaluation results, including adjusting model parameters and changing the model structure.

[0076] S2.2.6 Prediction and application: Use the trained random forest regression model to predict the water volume in the second cooling zone.

[0077] Furthermore, the parameters described in S2.2.2 include: the number of decision trees, feature selection method and decision tree growth method.

[0078] Example 2

[0079] A temperature prediction method based on a Transformer model, comprising:

[0080] S101: Collect relevant characteristic parameters of the continuous casting machine, collect production data of the continuous casting machine, obtain the distance between each section of the secondary cooling zone and the meniscus, and the corresponding temperature data during the casting process, which are used as training data for the model.

[0081] The distance parameters of each section of the secondary cooling zone from the meniscus are shown in Table 1. The temperature data of the start and end positions of each zone during the billet production process are shown in Table 2. The billet section specifications in this table are 1250mm wide, 15℃ overheated, and 1m / min pulling speed.

[0082] Table 1 Secondary cooling zone parameters

[0083]

[0084] Table 2 Corresponding temperature data extracted during the casting process

[0085]

[0086]

[0087] S102: Determine characteristic parameter I of the temperature prediction method and determine the input characteristic sequence of the temperature prediction model. Characteristic parameter I includes characteristics such as section specifications, temperature at the previous position, pulling speed, superheat, and water volume over a period of time. The model output is the temperature at the next position.

[0088] S103: The feature parameters I are input into the Transformer model in the order of the time series. The model encodes these feature parameters and learns the self-attention mechanism to capture the time dependency in the sequence and the interaction between features, thereby achieving accurate prediction of the temperature at the next position.

[0089] When constructing a temperature prediction model, the present invention uses a Transformer model, which is suitable for the regression prediction task in the present invention. In the fields of machine learning and data science, regression analysis is a basic method for predicting numerical target variables. Traditionally, models such as linear regression, decision tree regression, and random forest have been widely used to solve regression problems. However, with the rapid development of deep learning technology, especially the rise of the Transformer model, new possibilities have been brought to regression analysis. The Transformer has achieved great success in the field of natural language processing (NLP) with its excellent ability to process sequence data. The Transformer model was originally proposed by Vaswani et al. in 2017 to solve NLP tasks such as machine translation. Its core lies in the self-attention mechanism, which can capture the dependency between any two positions in the input sequence, thereby effectively modeling long-distance dependencies. The Transformer consists of two parts: an encoder and a decoder. However, in regression tasks, the encoder part is the main focus because it can convert the input sequence into a high-dimensional representation, which can then be used for regression prediction.

[0090] The core of the temperature prediction function lies in using the Transformer model to capture temporal dependencies and inter-feature correlations within the input feature sequence. The Transformer model captures sequence dependencies through a self-attention mechanism and employs a multi-layered attention mechanism between the encoder and decoder to better capture information in the input sequence. The temperature prediction model accepts a feature sequence as input and outputs a predicted temperature value for the next time point in the sequence.

[0091] Feature Parameter I includes section specifications, pulling speed, superheat, distance, temperature, and water volume characteristics of the secondary cooling zone over a period of time. After preprocessing, this data is input into the Transformer model in time series order. By encoding these features and learning a self-attention mechanism, the model captures long-term dependencies and interactions between features in the sequence. The model then outputs a predicted value, namely the temperature at the next location.

[0092] The specific steps of the temperature prediction method include:

[0093] 1) Data preprocessing: First, the feature parameter I is preprocessed, including feature scaling, sequence filling and other operations to ensure that the input data meets the requirements of the model.

[0094] 2) Model Construction: Build a Transformer model, consisting of an encoder and an output layer. The encoder uses multiple stacked Transformer encoder layers, each of which incorporates a self-attention mechanism and a feedforward neural network. The output layer converts the encoder output into the final predicted temperature value.

[0095] 3) Training Process: The model is trained using the training dataset. During training, the model calculates predicted values ​​through forward propagation, compares them with the true values, and calculates the loss function. The model parameters are then updated through the backpropagation algorithm to minimize the loss function. Training uses a leave-one-out cross-validation method to avoid issues caused by a small training set. This cross-validation method is suitable for small datasets.

[0096] 4) Model Evaluation and Tuning: Evaluate model performance on the validation or test set, using metrics such as mean squared error (MSE), root mean squared error (RMSE), and mean absolute error (MAE) to assess the model's prediction accuracy. Based on the evaluation results, fine-tune the model, including adjusting model parameters and modifying the model structure.

[0097] 5) Prediction and Application: Use the trained model to predict temperature. In practice, new time series data can be continuously input, and the model's prediction results can be updated in real time. The prediction results can be used to guide temperature control, early warning, and decision support in the production process.

[0098] In summary, the temperature prediction function based on the Transformer model achieves accurate prediction of future temperature by capturing the temporal dependency and inter-feature correlation in the input feature sequence. The modeling method adopts the Transformer architecture in deep learning and completes the entire modeling and prediction process through steps such as data preprocessing, model construction, training, evaluation and tuning, and prediction and application. The temperature prediction process is as follows: Figure 2 shown.

[0099] The model prediction results are as follows Figure 3 As shown in Figure 3, the superheat and casting speed combinations are (1) 15-1.0, (2) 15-1.1, (3) 15-1.2, and (4) 15-1.3, respectively. The cross-sectional specifications are all 1750 mm. The evaluation indicators of the training results are shown in Table 3. It can be seen from the results that the model prediction is accurate and can achieve accurate prediction of the temperature at each position in the secondary cooling zone.

[0100] Table 3 Evaluation indicators of temperature prediction results

[0101]

[0102] Secondary cooling water distribution prediction method includes:

[0103] S104: Determine the input characteristic parameters II of the dynamic secondary cooling water distribution model. Specifically, the input data includes cross-section specifications, temperature at the previous position, temperature at the next position obtained by the temperature prediction model, pulling speed, superheat and other characteristics. The model output is the water volume between the previous and next positions.

[0104] The water volume data between the start and end positions of each zone during the casting process are shown in Table 4. The casting section specifications in this table are 1250 mm wide, 15°C superheat, and 1 m / min casting speed.

[0105] Table 4 Extracted water volume data between the start and end positions of each zone during the casting process

[0106]

[0107]

[0108] S105: Input characteristic parameter II into the random forest model. Through the ensemble learning of random forest, the model can make full use of the inherent laws and correlations of the data to achieve accurate prediction of the water volume in each section, so as to optimize the continuous casting process and improve the quality of the ingot and production efficiency.

[0109] In constructing a dynamic secondary cooling water distribution model, the present invention utilizes the random forest algorithm. Random forest regression is an ensemble learning-based algorithm, belonging to the bagging type. It performs regression tasks by constructing multiple decision trees and integrating their predictions. In a random forest, each decision tree is trained independently on randomly selected subsamples, effectively reducing the risk of overfitting and ensuring high accuracy and generalization of the overall model. The random forest algorithm averages or weighted averages the predictions of multiple decision trees to obtain the final regression result. Its impressive performance is primarily attributed to its "random" and "forest" features: one makes it resistant to overfitting, the other enhances its accuracy. Its advantages include: it can train multiple decision trees simultaneously, each independently constructed, enabling parallel processing of large datasets and achieving high parallelization. By randomly selecting features and samples for training, it reduces the risk of overfitting in individual decision trees. It is robust to missing data and outliers, making it less susceptible to noise. It can also provide feature importance ranking, helping analysts understand the model's prediction process and providing strong interpretability.

[0110] Specifically, characteristic parameter II includes section specifications, location information, temperature difference between adjacent positions, pulling speed, and superheat, among which the temperature of the next position of the temperature difference between adjacent positions is predicted by the temperature prediction model, and the output is the water volume between adjacent positions.

[0111] Secondary cooling water distribution prediction methods include:

[0112] 1) Data preprocessing: First, the input data is preprocessed, including feature scaling, sequence padding and other operations to ensure that the input data meets the requirements of the model.

[0113] 2) Model Construction: Build a random forest: In the Scikit-learn library, use the RandomForestRegressor class to build a random forest regression model. Set parameters to control the behavior of the random forest, such as the number of decision trees, feature selection method, and decision tree growth method.

[0114] 3) Training the model: The random forest regression model is trained using the training set. The model constructs multiple decision trees based on the samples in the training set and the value of the target variable, and performs feature selection and partitioning on each tree.

[0115] 4) Prediction: Use the trained random forest regression model to predict the samples in the test set. The model will average or weighted average the prediction results of each decision tree to obtain the final regression prediction result.

[0116] 5) Model Evaluation and Tuning: Evaluate model performance on the validation or test set, using metrics such as mean squared error (MSE), root mean squared error (RMSE), and mean absolute error (MAE) to assess the model's prediction accuracy. Based on the evaluation results, fine-tune the model, including adjusting model parameters and modifying the model structure.

[0117] 6) Prediction and Application: Use the trained model to predict water flow in each section of the secondary cooling zone. In practice, new time series data can be continuously input, and the model's prediction results can be updated in real time. The prediction results can be used to guide water flow control, early warning, and decision support during the production process.

[0118] The model prediction results are as follows Figure 4 As shown in Figure 5, the superheat and casting speed combinations are (1) 15-1.0, (2) 15-1.1, (3) 15-1.15, and (4) 15-1.3, respectively. The cross-sectional specifications are all 1750 mm. The training also uses the leave-one-out method. The evaluation indicators of the training results are shown in Table 5. It can be seen from the results that the model prediction is accurate and can achieve accurate prediction of the water volume in each section of the secondary cooling zone.

[0119] Table 5 Evaluation indicators of temperature prediction results

[0120]

[0121] The embodiments of the present invention are described in detail above. Specific examples are used herein to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only used to help understand the method of the present invention and its core ideas. At the same time, for those skilled in the art, according to the ideas of the present invention, there may be changes in the specific implementation methods and application scopes. In summary, the contents of this specification should not be understood as limiting the present invention.

Claims

1. A data-driven dynamic secondary cooling water distribution method for slab continuous casting, the water distribution method is used to predict the water volume in each section of the secondary cooling zone, characterized in that: The dynamic secondary cooling water distribution method includes a temperature prediction method and a secondary cooling water distribution prediction method; According to the different distances between the second cooling zone and the meniscus, a plurality of position points are sequentially set on the second cooling zone, and the position points divide the second cooling zone into a plurality of segments, and a segment is formed between two adjacent position points. The position point of the current segment close to one end of the meniscus is the previous position point of the current segment, and the position point of the other end of the current segment is the next position point of the current segment; The segment adjacent to the current segment in the meniscus direction is the previous segment, and the previous position point of the current segment is the next position point of the previous segment; The segment adjacent to the current segment in the opposite direction of the meniscus is the next segment, and the next position point of the current segment is the previous position point of the next segment; The temperature prediction method includes: first preprocessing feature parameter I, then inputting feature parameter I into a Transformer model in the order of a time series, wherein the Transformer model captures the temporal dependency and the interaction between features in the time series by encoding feature parameter I and learning a self-attention mechanism, thereby obtaining a predicted value of the temperature at the next position point in the current section, wherein feature parameter I is a parameter affecting the temperature of the second cold zone, and the predicted value of the temperature at the next position point is obtained by predicting the temperature at the previous position point in the current section using the temperature prediction method; The secondary cooling water distribution prediction method includes: through the integrated learning of random forests, using the inherent rules and correlations of characteristic parameters II, to predict the water volume of the to-be-predicted section of the secondary cooling zone, the characteristic parameters II include section specifications, pulling speed and superheat, the characteristic parameters II also include the previous position point temperature and the next position point temperature corresponding to the to-be-predicted section of the secondary cooling zone, the next position point temperature is predicted by the temperature prediction method, and the water volume of the to-be-predicted section of the secondary cooling zone is predicted based on the characteristic parameters II.

2. The dynamic secondary cooling water distribution method according to claim 1, characterized in that: The characteristic parameter I includes: cross-sectional specifications, pulling speed, superheat, distance, temperature and water volume, the cross-sectional specifications are the width and length of the ingot, the pulling speed is the speed at which the ingot is pulled out of the crystallizer, the superheat is the difference between the pouring temperature of the molten steel and its liquidus temperature, the distance is the distance between each section of the secondary cooling zone and the meniscus, the temperature is the temperature of the previous position point, the temperature of the previous position point is the center temperature of the ingot at that position point, and the water volume is the water volume of the secondary cooling zone section.

3. The dynamic secondary cooling water distribution method according to claim 2, characterized in that: The temperature prediction method comprises: S1.1 Collect relevant characteristic parameters of the continuous casting machine production, including distance, temperature, and water volume. The distance is the distance between each section of the secondary cooling zone and the meniscus. The temperature is the center temperature of the position of the billet corresponding to the position in the secondary cooling zone. The water volume is the water volume in each section of the secondary cooling zone. S1.2 Determine characteristic parameters of the temperature prediction method I; S1.3 inputs the feature parameter I into the Transformer model in the order of the time series. The Transformer model captures the time dependency and interaction between features in the sequence by encoding the feature parameter I and learning the self-attention mechanism, and obtains the predicted value of the temperature at the next location point.

4. The dynamic secondary cooling water distribution method according to claim 3, characterized in that: S1.3 includes: S1.3.1 Data preprocessing: preprocess characteristic parameter I to ensure that characteristic parameter I meets the requirements of the model; S1.3.2 Build a Transformer model, which includes an encoder and an output layer. The encoder uses multiple stacked Transformer encoding layers, each of which includes a self-attention mechanism and a feedforward neural network. The output layer converts the encoder output into a final predicted temperature value. S1.3.3 Training process: The Transformer model is trained using a training dataset. The Transformer model first calculates the predicted value through forward propagation, compares it with the true value, calculates the loss function, and then updates the model parameters through the backpropagation algorithm to minimize the loss function. S1.3.4 Model evaluation and tuning: Evaluate the performance of the Transformer model on the validation set or test set, use mean square error, root mean square error, and mean absolute error metrics to evaluate the model's prediction accuracy, and tune the model based on the evaluation results; S1.3.5 Prediction and Application: Use the trained Transformer model to perform temperature prediction. In actual production, by continuously inputting new time series data, real-time updated prediction results are obtained.

5. The dynamic secondary cooling water distribution method according to claim 4, characterized in that: The preprocessing described in S1.3.1 includes feature scaling and sequence filling.

6. The dynamic secondary cooling water distribution method according to claim 4, characterized in that: S1.3.3 The training data set is composed of the relevant feature parameters.

7. The dynamic secondary cooling water distribution method according to claim 4, characterized in that: The tuning described in S1.3.4 includes adjusting model parameters and changing model structure.

8. The dynamic secondary cooling water distribution method according to claim 1, characterized in that: The secondary cooling water distribution prediction method comprises: S2.1 Determine characteristic parameters II of the secondary cooling water distribution prediction method, which include section specifications, casting speed, superheat, previous point temperature, and next point temperature; S2.2 inputs characteristic parameter II into the random forest model. Through the integrated learning of the random forest, the random forest model uses the inherent rules and correlations of the input data to obtain the water volume between the previous position point and the next position point, thereby realizing the prediction of the water volume in each section of the secondary cooling zone.

9. The dynamic secondary cooling water distribution method according to claim 8, characterized in that: S2.2 includes: S2.2.1 Data preprocessing: Preprocess feature parameter II. This preprocessing includes feature scaling and sequence filling to ensure that the input data meets the requirements of the model. S2.2.2 Model construction: Build a random forest regression model in the Scikit-learn library and set parameters to control the behavior of the random forest; S2.2.3 training the model, using the training data set to train the random forest regression model; the random forest regression model will construct multiple decision trees based on the samples in the training set and the value of the target variable, and perform feature selection and partitioning on each tree; S2.2.4 Prediction results: Use the trained random forest regression model to predict the samples in the test set; the trained random forest regression model averages or weighted averages the prediction results of each decision tree to obtain the final regression prediction result; S2.2.5 Model Evaluation and Tuning: Evaluate the performance of the trained random forest regression model on the validation set or test set, using the mean square error, root mean square error, and mean absolute error metrics to assess the prediction accuracy of the trained random forest regression model. Tune the model based on the evaluation results, including adjusting model parameters and changing the model structure. S2.2.6 Prediction and application: Use the trained random forest regression model to predict the water volume in the second cooling zone.

10. The dynamic secondary cooling water distribution method according to claim 9, characterized in that: The parameters described in S2.2.2 include: the number of decision trees, feature selection method and decision tree growth method.

Images