Converter steelmaking endpoint carbon temperature soft measurement method based on ensemble learning

Abstract

The application discloses a converter steelmaking endpoint carbon temperature soft measurement method based on integrated learning, and specific steps comprise the following steps: step 1, obtaining historical data, collecting historical data in actual steel converter steelmaking production process; step 2, constructing a diversity base model pool, the historical data is divided according to a plurality of different data division characteristics clustering algorithm, and a plurality of sample subsets with different division characteristics are obtained; step 3, fusing a to-be-measured sample and the model subset to obtain a prediction result, the to-be-measured sample is each data in a to-be-measured steel converter steelmaking production process, the local similarity of the to-be-measured sample and the historical data is converted into a corresponding weight, and a terminal carbon temperature prediction result is obtained through weighted fusion. The application constructs a diversity base model pool, iteratively optimizes the base models in the base model pool, and adaptively selects a model subset meeting the accuracy requirement. Higher-precision prediction results are obtained through the foregoing steps.

Description

Technical Field

[0001] This invention discloses a soft measurement method for the carbon temperature at the end point of converter steelmaking based on ensemble learning, which relates to the technical field of soft measurement methods for the carbon temperature at the end point of converter steelmaking. Background Technology

[0002] Converter steelmaking uses molten iron, scrap steel, and ferroalloys as main raw materials, relying on the heat generated by the physicochemical reactions between the molten steel and various raw materials in the molten pool to achieve heating and decarburization. Its key technology lies in endpoint control, meaning that at the end of oxygen blowing, the carbon content, the content of various metallic elements, and the temperature of the molten steel in the pool must simultaneously meet the requirements for tapping. Therefore, achieving accurate real-time measurement and control of carbon content and temperature is of great value and significance for improving the accuracy of single-pass tapping, reducing the consumption of raw materials and energy, lowering production costs, and improving product quality.

[0003] In existing technologies, methods for detecting carbon content and temperature during converter blowing can be categorized from a measurement principle perspective into traditional sensor-based detection methods, spectral radiometry detection, flame image processing and recognition detection, and process data-driven soft measurement methods. Traditional sensor-based detection methods mainly include sub-lance control technology and furnace gas analysis technology. Sub-lance control technology involves directly measuring the carbon content in the molten steel near the end point using a dedicated probe. Furnace gas analysis technology predicts the carbon content in the molten steel during blowing by analyzing the composition of the converter furnace gas. However, traditional sensors suffer from problems such as discontinuous endpoint carbon and temperature detection methods or endpoint judgments, high equipment costs hindering widespread adoption, or reliance on operator experience leading to significant subjectivity.

[0004] With the development of computer technology, optical technology, and imaging technology, some advanced endpoint detection technologies have been proposed to compensate for the shortcomings of traditional methods, such as low accuracy and expensive equipment. These include spectral radiometric detection and flame image processing and recognition detection. However, due to the harsh environment of actual converter steelmaking production and the presence of many interfering factors, the collected spectral radiometric and flame image information are affected, leading to significant errors in the final prediction results.

[0005] To reduce the aforementioned prediction errors, researchers in the field employ soft measurement methods for prediction. Among these, ensemble learning is a widely used learning strategy in complex industrial processes. However, the actual industrial process of converter steelmaking is a closed dynamic process. Traditional ensemble learning ignores the differences between the test samples and uses all sub-models for prediction. The varying performance of each sub-model leads to a decline in the overall performance of ensemble learning.

[0006] Content of this invention

[0007] The purpose of this invention is to provide a soft measurement method for the carbon temperature at the end point of converter steelmaking based on ensemble learning, thereby solving the problem of poor accuracy in predicting the end point of converter steelmaking using ensemble learning soft measurement in the prior art.

[0008] To achieve the above-mentioned technical objectives and effects, the invention is implemented through the following technical solution:

[0009] A soft measurement method for the final carbon temperature in converter steelmaking based on ensemble learning, comprising the following steps:

[0010] Step 1: Obtain historical data.

[0011] Collect historical data from the actual steel converter steelmaking process;

[0012] Step 2: Construct a pool of diverse base models.

[0013] The historical data is divided according to clustering algorithms with different data segmentation characteristics to obtain several sample subsets with different segmentation characteristics;

[0014] Step 3: Merge the test samples with the model subset to obtain the prediction results.

[0015] The test sample consists of various data from the steelmaking process in the converter. The local similarity between the test sample and the historical data is converted into corresponding weights, and the endpoint carbon temperature prediction result is obtained through weighted fusion.

[0016] Furthermore, step 1 specifically includes:

[0017] Step 101: Standardize the historical data using the maximum-minimum normalization principle to eliminate the influence of data units;

[0018] Step 102: Clean and select features from the standardized sample set;

[0019] Step 103: Divide the historical data and convert it into a sample set.

[0020] Furthermore, step 2 specifically includes:

[0021] Step 201: The clustering algorithms include Self-Organizing Maps (SOM), Fuzzy C-Means (FCM), Spectral Clustering (SPC), and Gaussian Mixture Model (GMM). The sample set is divided into four subsets with different partitioning characteristics using the four clustering algorithms.

[0022] Step 202: Using the internal evaluation metric of clustering, the silhouette coefficient (SC), measure the optimal number of clusters for each clustering algorithm.

[0023]

[0024] Where 'a' represents the average distance between the sample and all other points in the same cluster, and 'b' represents the average distance between the sample and all other points in the next nearest cluster. A larger silhouette coefficient indicates that samples belonging to the same class are closer together, and samples from different classes are farther apart.

[0025] Furthermore, step 3 specifically includes:

[0026] Step 301: Measure the global similarity between the test sample and the sample set using standard Euclidean distance (GlobalSED).

[0027]

[0028] Where V is the variance of the two feature vectors, and if the reciprocal of this variance is regarded as a weight, then this formula can be regarded as a weighted Euclidean distance.

[0029] Step 302: Sort the obtained global similarity GlobalSED from largest to smallest, and select the top 1500 samples to form similar regions S;

[0030] Step 303: Select α models that meet the similarity criteria, and then obtain their corresponding models for the test sample x. q The prediction results.

[0031] y pre,α =F α (x q ,ω);

[0032] Where F(·) represents the functional relationship of the model, and ω is the model parameter;

[0033] Step 304: The base model retained after selection is M. new = [M1, M2, ..., Mα], α ∈ [1, m], and the corresponding sample subset of the data is D. new =((x) 1,i ,y 1,i ),(x 2,i ,y 2,i ),…,(x α,i ,y α,i )};α∈[1,m],i=1,2,3,…,n m}

[0034] Step 305: For the sample to be tested (hereinafter referred to as x)q (Representative), using standard Euclidean distance to measure the distance to samples in each of the sample subsets, and using its mean to represent the distance from the test sample to the corresponding sample subset, and using it as the model weight w corresponding to each base model Mα. α .

[0035]

[0036] Step 306, Finally, the sample to be tested x q The weighted fusion endpoint carbon temperature prediction output is as follows:

[0037]

[0038] Where y test y represents the true value of the sample to be tested. prediction The predicted value is for the sample to be tested.

[0039] Furthermore, step 2 also includes step 203:

[0040] Step 203: Construct a base model pool for the data subset, and iteratively optimize several base models in the base model pool M to obtain an optimal model that meets the accuracy requirements.

[0041] Furthermore, step 203 includes steps 2031-2036.

[0042] Step 2031: First, the model must be initialized.

[0043]

[0044] Where F represents the GBDT model, F0 is the initial state of the model, N is the number of training samples, and y i y is the initial true value, c is the initial predicted value, and L(y) is the initial true value. i c) is the loss function of the model, that is, to find the minimum c that minimizes the error of all samples. The loss function commonly used in the GBDT regression algorithm is least squares regression.

[0045] Step 2032: After initialization, for dataset D′ m The training samples i = 1, 2, 3, ..., n m Calculate the response value of the m-th tree, which is the negative gradient of the loss function:

[0046]

[0047] Step 2033: Then, use the CART regression tree to fit the residuals to obtain the m-th regression tree, whose corresponding leaf node is R. m,jWhere m represents the m-th tree, and j = 1, 2, 3, ..., J is the number of leaf nodes in the tree; then, for each leaf node, the best-fit value is calculated:

[0048]

[0049] Step 2034: Then update the model, using m trees for prediction in the next iteration.

[0050]

[0051] Where δ is the learning rate, which is used to control the degree to which each boosting tree improves the GBDT model;

[0052] Step 2035, repeat steps 2032-2034, to build m classification regression trees (m = 1, 2, 3, ..., M).

[0053] Step 2036, the final GBDT model expression for the M classification and regression trees is:

[0054]

[0055] The first Σ represents the summation of the M boosting trees, and the second Σ represents the summation of the best-fit values ​​of all leaf nodes in each boosting tree. That is, to obtain the final GBDT regressor, the initial regressor F0(x) is used to continuously sum the best-fit values ​​of the leaf nodes of all M trees. Therefore, the process of building the GBDT model involves applying the best-fit values ​​of the clustered sample subset D′ to the model. m Add a new CART regression tree to fit the predicted values ​​c and the true values ​​y of all the previous trees. i The residuals between them, and the prediction result is the sum of the predictions of all the trees that have been built.

[0056] Furthermore, step 3 also includes step 3031:

[0057] Step 3031: Calculate the prediction accuracy (accu) of all base models in the base model pool M on the similar region S. S (M), the optimal sample subset is selected based on the calculation results.

[0058] Furthermore, step 3031 includes steps 30311-30312:

[0059] Step 30311: Calculate the prediction accuracy (accu) of all the base models in the base model pool M on the similar region S. S (M)

[0060]

[0061]

[0062] PE indicates whether the prediction accuracy of the final carbon temperature is within the error range (Th). If it is, it is a hit and recorded as 1; otherwise, it is recorded as 0. The prediction error range for carbon content is Th = 0.02%, and the prediction error range for temperature is Th = 10℃.

[0063] Step 30312: Sort the accuracy of the base models from largest to smallest. If the sum of the prediction accuracy of the first α (α∈m) models is greater than the set threshold, then select the first α models as the selected models for the test sample, that is, the sub-models that will participate in the integration and fusion later.

[0064]

[0065] Furthermore, the sample set consists of experimental data with 6 feature dimensions and 5500 samples. The 6-dimensional features of carbon content are oxygen pressure 31, iron mixing time, time from the end of iron mixing to the start of oxygen production, molten iron temperature, time from the start of iron mixing to tapping, and the amount of scrap steel charged. The 6-dimensional features of temperature are the amount of pig iron charged, molten iron P, lance position 22, oxygen pressure 29, oxygen pressure 11, and lance position 16. Of the 5500 samples, 5000 are used as training data and 500 are used as test data.

[0066] Furthermore, step 3 is repeated for the next sample to be tested to obtain the endpoint carbon temperature prediction results for all samples to be tested.

[0067] Beneficial effects:

[0068] This invention combines clustering algorithms with different data partitioning characteristics to divide the dataset, constructing a diverse base model pool. The base models in the pool are iteratively optimized, and an adaptive selection strategy is used to model similar regions related to the test samples. A subset of models meeting the accuracy requirements is adaptively selected from the base model pool. Through these steps, higher accuracy prediction results are obtained.

[0069] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0070] Figure 1 The flowchart shows the soft measurement method for the final carbon temperature in converter steelmaking based on MC-DAS ensemble learning.

[0071] Figure 2 The diagram shows the DAS ensemble learning algorithm in step 3.

[0072] Figure 3 A comparison chart of predicted and actual carbon content values ​​at the end point of converter steelmaking;

[0073] Figure 4 A comparison chart of predicted and actual final temperatures at the converter steelmaking endpoint; Detailed Implementation

[0074] To more clearly illustrate the technical solutions of the embodiments of the present invention, the present invention will be described in detail below with reference to the accompanying drawings.

[0075] Example 1

[0076] Existing technologies using ensemble learning algorithms for accurate prediction of endpoint carbon temperature do not take into account the special characteristics of the converter steelmaking process, resulting in poor overall prediction performance. Based on the specific characteristics of the converter steelmaking process, the applicant has developed a soft measurement method for endpoint carbon temperature with better prediction accuracy, while ensuring the diversity of the ensemble learning base model and the correlation between the test sample and prior data.

[0077] The soft measurement method for the endpoint carbon temperature in converter steelmaking based on ensemble learning described in this embodiment includes the following steps:

[0078] Step 1: Obtain historical data.

[0079] Collect historical data from the actual steel converter steelmaking process.

[0080] In some embodiments, in order to eliminate interference from invalid data or other unnecessary data, step 1 specifically includes:

[0081] Step 101: Standardize the historical data using the maximum-minimum normalization principle to eliminate the influence of data units.

[0082] Step 102: Clean and select features from the standardized sample set.

[0083] Step 103: Divide the historical data and convert it into a sample set.

[0084] In one specific embodiment, the sample set is divided into 6 feature dimensions and contains 5500 experimental data samples. The 6 features of carbon content are oxygen pressure 31, iron addition time, time from the end of iron addition to oxygen opening, molten iron temperature, time from the start of iron addition to tapping into the furnace, and amount of scrap steel charged. The 6 features of temperature are amount of pig iron charged, molten iron P, lance position 22, oxygen pressure 29, oxygen pressure 11, and lance position 16. Of the 5500 samples, 5000 are used as training data and 500 are used as test data.

[0085] By employing the minimax normalization principle, standardization, cleaning, and feature selection, the resulting sample set is more reliable and can improve accuracy in subsequent steps.

[0086] Step 2: Construct a pool of diverse base models.

[0087] The historical data is divided into several sample subsets with different division characteristics based on clustering algorithms with different data segmentation characteristics.

[0088] It is easy to understand that different clustering algorithms have different partitioning characteristics, which is beneficial for the invention of this embodiment to construct a diverse subset of samples. The prediction results obtained through such a subset of samples have better reliability.

[0089] In one specific embodiment, step 2 specifically includes:

[0090] Step 201: The clustering algorithms include Self-Organizing Maps (SOM), Fuzzy C-Means (FCM), Spectral Clustering (SPC), and Gaussian Mixture Model (GMM). The sample set is divided into four subsets with different partitioning characteristics using the four clustering algorithms.

[0091] Step 202: Using the internal evaluation metric of clustering, the silhouette coefficient (SC), measure the optimal number of clusters for each clustering algorithm.

[0092]

[0093] Where 'a' represents the average distance between the sample and all other points in the same cluster, and 'b' represents the average distance between the sample and all other points in the next nearest cluster. A larger silhouette coefficient indicates that samples belonging to the same class are closer together, and samples from different classes are farther apart.

[0094] The clustered data subsets obtained in steps 201-202 are obtained after measuring the silhouette coefficient (SC). Using these data subsets as the same evaluation criterion is beneficial to improving the prediction accuracy.

[0095] Step 3: Merge the test samples with the model subset to obtain the prediction results.

[0096] The test samples are various data from the steelmaking process in the converter. The local similarity between the test samples and the sample subset is converted into corresponding weights, and the endpoint carbon temperature prediction result is obtained through weighted fusion.

[0097] In one embodiment, step 3 specifically includes:

[0098] Step 301: Measure the global similarity between the test sample and the sample set using standard Euclidean distance (GlobalSED).

[0099]

[0100] Where V is the variance of the two feature vectors, if the reciprocal of this variance is regarded as a weight, then this formula can be regarded as a weighted Euclidean distance.

[0101] Step 302: Sort the obtained global similarity GlobalSED from largest to smallest, and select the top 1500 samples to form similar regions S.

[0102] Step 302 is actually a screening process for the sample subset, using the sample with the highest relevance as the basis for subsequent prediction.

[0103] Step 303: Select α sample models that meet the similarity criteria, and then obtain their corresponding test sample x. q The prediction results.

[0104] y pre,α =F α (x q ,ω);

[0105] Where F(·) represents the functional relationship of the model, and ω is the model parameter.

[0106] Step 304: The base model retained after selection is M. new = [M1, M2, ..., Mα], α ∈ [1, m], and the corresponding sample subset of the data is D. new ={(x 1,i ,y 1,i ),(x 2,i ,y 2,i ),…,(x α,i ,y α,i )};α∈[1,m],i=1,2,3,…,n m}

[0107] Step 305: For the sample to be tested (hereinafter referred to as x) q (Representative), using standard Euclidean distance to measure the distance to samples in each of the sample subsets, and using its mean to represent the distance from the test sample to the corresponding sample subset, and using it as the model weight w corresponding to each base model Mα. α .

[0108]

[0109] Step 306, Finally, the sample to be tested x qThe weighted fusion endpoint carbon temperature prediction output is as follows:

[0110]

[0111] Where y test y represents the true value of the sample to be tested. prediction The predicted value is for the sample to be tested.

[0112] The final prediction result is obtained through step 306, which takes into account the correlation between the sample subset and the sample to be tested, and optimizes the multi-feature sample subset to improve the accuracy of the result.

[0113] Example 2

[0114] In conjunction with the selection of multi-feature sample subsets described in Example 1, considering that selecting only highly correlated sample subsets may not yield ideal results, it is necessary to iteratively optimize the sample subsets to obtain the best optimal model result. Therefore, in this embodiment, step 2 further includes step 203:

[0115] Step 203: Construct a base model pool for the data subset, and iteratively optimize several base models in the base model pool M to obtain an optimal model that meets the accuracy requirements.

[0116] In one specific embodiment, step 203 includes steps 2031-2036.

[0117] Step 2031: First, the model must be initialized.

[0118]

[0119] Where F represents the GBDT model, F0 is the initial state of the model, N is the number of training samples, and y i y is the initial true value, c is the initial predicted value, and L(y) is the initial true value. i c) is the loss function of the model, that is, to find the minimum c that minimizes the error of all samples. The loss function commonly used in the GBDT regression algorithm is least squares regression.

[0120] Step 2032: After initialization, for dataset D′ m The training samples i = 1, 2, 3, ..., n m Calculate the response value of the m-th tree, which is the negative gradient of the loss function:

[0121]

[0122] Step 2033: Then, use the CART regression tree to fit the residuals to obtain the m-th regression tree, whose corresponding leaf node is R. m,jWhere m represents the m-th tree, and j = 1, 2, 3, ..., J is the number of leaf nodes in the tree; then, for each leaf node, the best-fit value is calculated:

[0123]

[0124] Step 2034: Then update the model, using m trees for prediction in the next iteration.

[0125]

[0126] Where δ is the learning rate, which is used to control the degree to which each boosting tree improves the GBDT model;

[0127] Step 2035, repeat steps 2032-2034, to build m classification regression trees (m = 1, 2, 3, ..., M).

[0128] Step 2036, the final GBDT model expression for the M classification and regression trees is:

[0129]

[0130] In this embodiment, the first Σ represents the summation of the M boosting trees, and the second Σ represents the summation of the best-fit values ​​of all leaf nodes of each boosting tree. That is, in order to obtain the final GBDT regressor, the best-fit values ​​of the leaf nodes of all M trees are continuously summed up using the initial regressor F0(x). Therefore, the process of building the GBDT model is based on the clustered sample subset D′ m Add a new CART regression tree to fit the predicted values ​​c and the true values ​​y of all the previous trees. i The residuals between them, and the prediction result is the sum of the predictions of all the trees that have been built.

[0131] Example 3

[0132] In conjunction with the content involved in Examples 1-2, the applicant also considered that the base model should be optimized by combining the actual results and the predicted results, and step 3 further includes step 3031:

[0133] Step 3031: Calculate the prediction accuracy (accu) of all base models in the base model pool M on the similar region S. S (M), the base model is optimized.

[0134] In one embodiment, step 3031 includes steps 30311-30312:

[0135] Step 30311: Calculate the prediction accuracy (accu) of all the base models in the base model pool M on the similar region S. S(M)

[0136]

[0137]

[0138] PE indicates whether the prediction accuracy of the final carbon temperature is within the error range (Th). If it is, it is a hit and recorded as 1; otherwise, it is recorded as 0. The prediction error range for carbon content is Th = 0.02%, and the prediction error range for temperature is Th = 10℃.

[0139] Step 30312: Sort the accuracy of the base models from largest to smallest. If the sum of the prediction accuracy of the first α (α∈m) models is greater than the set threshold, then select the first α models as the selected models for the test sample, that is, the sub-models that will participate in the integration and fusion later.

[0140]

[0141] In other embodiments, step 3 is repeated for the next test sample to obtain the endpoint carbon temperature prediction results for all test samples.

[0142] Example 4

[0143] To demonstrate the effectiveness of the methods described in Examples 1-3, this example compares the proposed method with traditional ensemble learning methods, using prediction accuracy (PA), root mean square error (RMSE), and mean absolute percentage error (MAPE) as evaluation metrics. Simulation results show that the proposed method exhibits better prediction performance than existing technologies. See Table 1 for details.

[0144] Table 1 Comparison between the present invention and traditional ensemble learning methods

[0145]

[0146]

[0147] The above are merely some of the embodiments of this application and are not intended to limit the application in any way. Any simple modifications, equivalent changes, and alterations made to the above embodiments shall still fall within the scope of protection of the technical solution of this application.

Claims

1. A soft measurement method for the final carbon temperature in converter steelmaking based on ensemble learning, characterized in that, The specific steps include: Step 1: Obtain historical data. Collect historical data from the actual steel converter steelmaking process; Step 2: Construct the base model pool. The historical data is divided according to clustering algorithms with different data segmentation characteristics to obtain several sample subsets with different segmentation characteristics; Step 3: Merge the test samples and the model subset to obtain the prediction results. The test sample is data from the steelmaking process in the converter. The local similarity between the test sample and the sample subset is converted into corresponding weights, and the endpoint carbon temperature prediction result is obtained through weighted fusion. Step 1 specifically includes: Step 101: Standardize the historical data using the maximum-minimum normalization principle to eliminate the influence of data units; Step 102: Clean and select features from the standardized sample set; Step 103: Divide the historical data and convert it into a sample set; Step 2 specifically includes: Step 201: The clustering algorithms include Self-Organizing Maps (SOM), Fuzzy C-Means (FCM), Spectral Clustering (SPC), and Gaussian Mixture Model (GMM). The sample set is divided into four data subsets with different partitioning characteristics using the four clustering algorithms. Step 202: Using the internal evaluation metric of clustering, the silhouette coefficient (SC), measure the optimal number of clusters for each clustering algorithm. ； Where a represents the average distance between the sample and all other points in the same cluster, and b represents the average distance between the sample and all other points in the next nearest cluster; the larger the silhouette coefficient, the closer the samples of the same class are, and the farther the samples of different classes are. Step 3 specifically includes: Step 301: Measure the global similarity between the test sample and the sample set using standard Euclidean distance. ; ； in It is the variance of two eigenvectors. If we consider the reciprocal of this variance as a weight, then this formula can be regarded as a weighted Euclidean distance. Step 302: Sort the obtained global similarity GlobalSED from largest to smallest, and select the top 1500 samples to form similar regions S; Step 303: Select those that meet the similarity criteria. A model is used to obtain the test sample. The prediction results; ； in Represents the functional relationships within the model. These are model parameters; Step 304: The base model retained after selection is... , The corresponding sample subset of the data is Step 305: For the test sample, use standard Euclidean distance to measure the distance to samples in each of the sample subsets, and use the mean of the mean distance to represent the distance from the test sample to the corresponding sample subset, and use it as the basis model for each sample subset. The corresponding model weights ; ； Step 306: Finally, the sample to be tested... The weighted fusion endpoint carbon temperature prediction output is as follows: ； in The true value of the sample to be tested. The predicted value for the sample to be tested; Step 3 further includes step 3031: Step 3031: Calculate the prediction accuracy of all base models in the base model pool M on the similar region S. The base model is then optimized. Step 3031 includes steps 30311-30312: Step 30311: Calculate the prediction accuracy of all base models in the base model pool M in the similar region S. ； ； PE indicates whether the prediction accuracy of the final carbon temperature is within the error range (Th). If it is, it is a hit and recorded as 1; otherwise, it is recorded as 0. Among them, the prediction error range of carbon content is Th=0.02%, and the prediction error range of temperature is Th=10℃. Step 30312: Sort the base models in descending order of accuracy. If the first... If the sum of the prediction accuracies of all models is greater than the set threshold, then the top model is selected. Each model is selected as the model for the test sample, which is also the sub-model that participates in the integration and fusion later. ； The base models in the base model pool were iteratively optimized. The model subset that meets the accuracy requirements was adaptively selected from the base model pool by combining the model model obtained by the adaptive selection strategy with the model model of similar regions related to the test sample.

2. The soft measurement method for the final carbon temperature of converter steelmaking based on ensemble learning according to claim 1, characterized in that, Step 2 further includes step 203: Step 203: Construct a base model pool for the data subset, and iteratively optimize several base models in the base model pool M to obtain an optimal model that meets the accuracy requirements.

3. The soft measurement method for the final carbon temperature of converter steelmaking based on ensemble learning according to claim 2, characterized in that, Step 203 includes steps 2031-2036. Step 2031: First, the model must be initialized. ； in Represents the GBDT model. That is, the initial state of the model. The number of training samples. It is the initial real value. The initial predicted value, Let this be the loss function of the model; we want to find the one that minimizes it. To minimize the error of all samples, the loss function commonly used in the GBDT regression algorithm is least squares regression. Step 2032: After initialization, for the dataset... training samples Calculate the first The response value of each tree, i.e., the negative gradient of the loss function: ； Step 2033: Then, use the CART regression tree to fit the residuals to obtain the m-th regression tree, whose corresponding leaf nodes are... ,in Indicates the first A tree, Let be the number of leaf nodes in the tree; then, for each leaf node, calculate the best-fit value: ； Step 2034: Then update the model and use m trees for prediction in the next iteration. ； in The learning rate is used to control the degree to which each boosting tree improves the GBDT model; Step 2035, repeat steps 2032-2034, to establish Classification and regression trees Step 2036, the final GBDT model expression for the M classification and regression trees is: 。 4. The soft measurement method for the final carbon temperature in converter steelmaking based on ensemble learning according to claim 1, characterized in that, The sample subset is divided into experimental data with 6 feature dimensions and a sample size of 5500. The 6 features of carbon content are oxygen pressure 31, iron addition time, time from the end of iron addition to oxygen opening, molten iron temperature, time from the start of iron addition to tapping into the furnace, and amount of scrap steel charged. The 6 features of temperature are amount of pig iron charged, molten iron P, lance position 22, oxygen pressure 29, oxygen pressure 11, and lance position 16. Of the 5500 experimental data samples, 5000 are used as training data and 500 are used as test data.

5. The soft measurement method for the final carbon temperature of converter steelmaking based on ensemble learning according to claim 1, characterized in that, Repeat step 3 for the next sample to be tested to obtain the endpoint carbon temperature prediction results for all samples to be tested.

Images