Prediction model and method of phosphorus content at converter smelting end point and model construction method

Through the combination of noise reduction autoencoder model, clustering model and prediction model, the problems of low hit rate and incomplete steel coverage of the phosphorus content at the end point of the converter smelting are solved, and accurate and real-time prediction is achieved, and automated steel refining is supported. and the development of digital factories.

CN118211191BActive Publication Date: 2024-11-12JIANGSU JINHENG INFORMATION TECH CO LTD
View PDF 5 Cites 0 Cited by

Patent Information

Application Number
CN202410626850.6
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-05-21
Publication Date
2024-11-12
Estimated Expiration
2044-05-21

AI Technical Summary

Technical Problem

The existing converter smelting endpoint phosphorus content prediction method has low hit rate, insufficient steel grade coverage, and lack real-time performance, making it difficult to support the automated steelmaking needs of digital factories.

Method used

The combination of noise reduction autoencoder model, clustering model and prediction model is adopted to achieve accurate prediction of the phosphorus content at the end point of the converter smelting through data preprocessing, feature extraction, clustering analysis and sub-prediction model integration.

Benefits of technology

The hit rate of the phosphorus content prediction at the end point of the converter smelting is improved, covering all steel grade data, and real-time prediction is achieved, supporting shortening of smelting cycles and automated steelmaking in digital plants.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118211191B_ABST
    Figure CN118211191B_ABST
Patent Text Reader

Abstract

The present application relates to the technical field of prediction of phosphorus content at the end point of converter smelting, and provides a prediction model, method and model construction method of phosphorus content at the end point of converter smelting, firstly collect historical production data, and perform data preprocessing to obtain an initial data set, and on the basis of the initial data set, train and save a denoising autoencoder, use the trained denoising autoencoder to denoise and compress the data to obtain a new data set, use K-Means clustering to divide the data set into multiple sub-data sets, and use an integrated model to train multiple sub-prediction models based on multiple sub-data sets, and finally deploy the denoising autoencoder model, K-Means clustering model and sub-prediction model online, and then input data to obtain the predicted value of phosphorus content at the end point through the model. The scheme covers data of all steel grades and improves the hit rate of prediction of phosphorus content at the end point of converter smelting, providing technical support for shortening the smelting cycle and realizing automated steelmaking in digital factories.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present application relates to the technical field of prediction of phosphorus content at the end point of converter smelting, and in particular to a prediction model and method for phosphorus content at the end point of converter smelting, and a method for constructing the model. Background Art

[0002] Converter steelmaking is a process that uses molten iron, scrap steel (pig iron blocks), and ferroalloys as the main raw materials. Without the help of external energy, it relies on the physical heat of the molten iron itself and the heat generated by the chemical reaction between carbon, silicon and other components in the molten iron to complete the steelmaking process in the converter. With the advancement of modern steel production technology, during the smelting process, the molten iron pretreatment process is generally used to treat impurities such as sulfur and phosphorus that are difficult to remove.

[0003] At present, accurate prediction of phosphorus content at the end point of converter smelting has become the core of converter end point control in modern process flow. Most domestic steel mills mainly control the converter steelmaking end point manually based on the operating experience of on-site operators, and the operating experience of on-site operators is only obtained through the summary and analysis of on-site production and production control data, which makes it difficult to achieve stable and accurate results.

[0004] With the development of intelligent control technology and information technology, the detection and collection of data information generated in the converter steelmaking production process have been realized, thus forming a huge converter steelmaking production data resource. Converter steelmaking is a complex nonlinear process. It is difficult to establish an accurate calculation model using only simple mechanism analysis, statistical regression and other methods. Therefore, some researchers have tried to introduce intelligent methods such as machine learning into the process of establishing the converter steelmaking endpoint prediction model. However, the survey results show that the existing converter smelting endpoint phosphorus content prediction method still has the problems of low hit rate and incomplete coverage of steel grades, and fails to meet the requirements of real-time prediction, making it difficult to provide technical support for the realization of automated steelmaking in digital factories. Summary of the invention

[0005] In order to solve the problems that the above-mentioned existing methods for predicting the phosphorus content at the end point of converter smelting still have low hit rate, incomplete coverage of steel grades, and fail to meet the requirements of real-time prediction, the first aspect of the present application provides a model for predicting the phosphorus content at the end point of converter smelting.

[0006] The prediction model of phosphorus content at the end point of converter smelting includes: denoising autoencoder model, clustering model and prediction model;

[0007] The output end of the denoising autoencoder model is connected to the input end of the clustering model, and the output end of the clustering model is connected to the input end of the prediction model;

[0008] The denoising autoencoder model comprises: an encoder, which is used to extract input data features, retain relevant information, remove noise and non-essential information, and compress the input data into a low-dimensional representation to obtain compressed features;

[0009] The clustering model includes k clustering categories, and the clustering model is used to receive the data processed by the denoising autoencoder model, and determine the clustering category of the processed data according to the data characteristics;

[0010] The prediction model includes k sub-prediction models, and the k sub-prediction models correspond one-to-one to the k cluster categories. The prediction model is used to receive data after the cluster model determines the cluster category, and input the data into the sub-prediction model corresponding to the cluster category to obtain the data prediction result.

[0011] In a feasible implementation method, the data features include: molten iron quantity, scrap steel quantity, total oxygen consumption, TSC-C content, TSC-temperature, TSO-C content, TSO-O content, TSO-temperature, C content of molten iron after desulfurization, Si content of molten iron after desulfurization, Mn content of molten iron after desulfurization, P content of molten iron after desulfurization, S content of molten iron after desulfurization, molten iron temperature after desulfurization, sludge ball slag forming agent, sintering return ore slag forming agent, heat generating agent slag forming agent, dolomite slag forming agent, lime slag forming agent, fine crushed powder slag forming agent, magnesium ball slag forming agent, limestone slag forming agent.

[0012] A second aspect of the present application provides a method for predicting the phosphorus content at the end point of converter smelting, which is used in any one of the above-mentioned prediction models for the phosphorus content at the end point of converter smelting, comprising the following steps:

[0013] Preprocess the test data to obtain initial data;

[0014] Inputting the initial data into the trained denoising autoencoder model to obtain new data after denoising and compression;

[0015] Inputting the new data into the clustering model, and obtaining the clustering category of the new data through the clustering model;

[0016] According to the clustering category, the new data is input into the sub-prediction model corresponding to the clustering category to obtain a data prediction result.

[0017] A third aspect of the present application provides a method for constructing a converter smelting endpoint phosphorus content prediction model, which is used to construct any one of the above converter smelting endpoint phosphorus content prediction models, comprising the following steps:

[0018] Collect the historical production data of each furnace derived from the converter smelting secondary system in the converter smelting production system;

[0019] Processing the historical production data to remove invalid data, abnormal data, and data with missing values ​​to obtain a data set after data preprocessing;

[0020] Construct a denoising autoencoder model, input the data set, train the denoising autoencoder, and complete the training until the loss of the denoising autoencoder model has converged, save the denoising autoencoder and output the trained data set;

[0021] According to the processed data set, the data is clustered using a K-Means clustering algorithm, k cluster centers are set by a silhouette coefficient method, the positions of the k cluster centers are optimized in an iterative manner, the processed data set is divided into different clusters according to the distances from the k cluster centers, k different sub-data sets and k cluster categories are obtained, and a clustering model is obtained;

[0022] A prediction model consisting of a base learner and a meta learner is established based on a Stacking strategy, and k sub-prediction models are obtained by training according to the k different sub-data sets;

[0023] A Python programming environment is installed on the industrial computer, and the Python programming environment is used to run Python programs on the industrial computer, load the denoising autoencoder model, clustering model and prediction model, and use the flask open source library to access or call the denoising autoencoder model, clustering model and prediction model.

[0024] In a feasible implementation, the denoising autoencoder model further includes: a decoder, the decoder being used to restore the compressed features generated by the encoder to the dimension of the original data, and compare with the original data to confirm the compression effect of the encoder;

[0025] The encoder compresses the input data into a low-dimensional representation, and the data dimension changes are: 22 dimensions to 64 dimensions, 64 dimensions to 64 dimensions, 64 dimensions to 32 dimensions, and 32 dimensions to 20 dimensions;

[0026] When the decoder restores the low-dimensional representation to the dimension of the original data, the data dimension changes are: 20 dimensions to 32 dimensions, 32 dimensions to 64 dimensions, 64 dimensions to 64 dimensions, and 64 dimensions to 22 dimensions;

[0027] The compression feature is obtained when the encoder compresses the data dimension to 20 dimensions.

[0028] In a feasible implementation, the parameters of the denoising autoencoder model are set as follows: the number of training rounds epoch is set to 500-5000, the learning rate lr is set to 0.001-0.01, the optimizer selects one of SGD, Adam and RMSprop, and the loss function selects mean square error.

[0029] In a feasible implementation, the steps of setting k cluster centers using the silhouette coefficient method include:

[0030] Determine a cluster number range of 2 to 10, and for each k value within the cluster number range, perform a K-Means clustering algorithm;

[0031] For each clustering result under the k value, the silhouette coefficient of each data point is calculated, and the silhouette coefficients of all data points under each k value are averaged to obtain the average silhouette coefficient under the k value;

[0032] With the k value as the horizontal axis and the average silhouette coefficient as the vertical axis, a silhouette coefficient curve graph is drawn;

[0033] Observe the silhouette coefficient curve and find the k value at which the silhouette coefficient reaches a maximum value or a local maximum value.

[0034] In a feasible implementation, the base learners include: deep forest regression, random forest regression, CatBoost regression and XGBoost regression; the meta learner includes linear regression.

[0035] In a feasible implementation, the historical production data is the actual production data of the latest N months stored in the converter smelting secondary system, where N is greater than or equal to 6.

[0036] In a feasible implementation, the historical production data includes at least: molten iron weight, chemical composition information and temperature information of molten iron after desulfurization, scrap steel weight, total oxygen consumption in the smelting process, slag-making agent consumption data and phosphorus content data at the end point of smelting; the invalid data includes at least: shift data, shift data and pig iron data; the abnormal data includes at least: data outside the business scope during the converter smelting process.

[0037] From the above content, it can be seen that the present application provides a prediction model, method and method for constructing the phosphorus content at the end point of converter smelting. First, historical production data is collected, and data preprocessing is performed to obtain an initial data set. On the basis of the initial data set, a denoising autoencoder is trained and saved. The trained denoising autoencoder is used to denoise and compress the data to obtain a new data set. The data set is divided into multiple sub-data sets using K-Means clustering. Multiple sub-prediction models are trained based on multiple sub-data sets using an integrated model. Finally, the denoising autoencoder model, K-Means clustering model and sub-prediction model are deployed online. After inputting data, the predicted value of the phosphorus content at the end point can be obtained through the model. The present application scheme covers data of all steel grades and improves the hit rate of the prediction of the phosphorus content at the end point of converter smelting, providing technical support for shortening the smelting cycle and realizing automated steelmaking in digital factories. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] The accompanying drawings herein are incorporated into and constitute a part of the specification, illustrate embodiments consistent with the implementation of the present invention, and together with the specification are used to explain the principles of the embodiments of the present invention. Obviously, the accompanying drawings described below are only some embodiments of the implementation of the present invention, and for ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without creative work.

[0039] Figure 1 This is a schematic diagram of the structure of a model for predicting phosphorus content at the end point of converter smelting, shown as an exemplary embodiment of the present application;

[0040] Figure 2 This is a flow chart of a method for predicting phosphorus content at the end point of converter smelting according to an exemplary embodiment of the present application;

[0041] Figure 3 A schematic flow chart of a method for constructing a model for predicting phosphorus content at the end point of converter smelting according to an exemplary embodiment of the present application;

[0042] Figure 4 This is a schematic diagram of the structure of a denoising autoencoder shown as an exemplary embodiment of the present application;

[0043] Figure 5 A schematic diagram showing that the loss of a denoising autoencoder model reaches convergence according to an exemplary embodiment of the present application;

[0044] Figure 6 A contour coefficient curve diagram showing an exemplary embodiment of the present application;

[0045] Figure 7 This is a schematic diagram of the structure of a prediction model shown in an exemplary embodiment of the present application. DETAILED DESCRIPTION

[0046] Example embodiments will now be described more fully with reference to the accompanying drawings. However, example embodiments can be implemented in a variety of forms and should not be construed as limited to the examples set forth herein; rather, these embodiments are provided so that the embodiments of the present invention will be more comprehensive and complete and fully convey the concepts of the example embodiments to those skilled in the art. The described features, mechanisms, or characteristics may be combined in any suitable manner in one or more embodiments. In the following description, many specific details are provided to provide a full understanding of the implementation of the embodiments of the present invention.

[0047] Converter steelmaking uses molten iron, scrap steel and ferroalloys as raw materials, and relies on their internal heat to complete the steelmaking process. In modern processes, molten iron pretreatment is used to remove impurities such as sulfur and phosphorus. However, accurately predicting the phosphorus content at the end of converter smelting remains a technical difficulty. At present, most steel mills rely on the experience of operators for manual control, but the effect is unstable. With the development of intelligent technology, converter steelmaking data can be collected to form a huge resource. However, due to its nonlinear complexity, it is difficult to establish an accurate model. Although attempts have been made to introduce intelligent methods, existing prediction methods still have problems such as low hit rate, incomplete coverage of steel grades, and lack of real-time performance, making it difficult to support the automated steelmaking needs of digital factories.

[0048] In order to solve the above problems, the first aspect of the embodiment of the present application provides a prediction model for the phosphorus content at the end point of converter smelting, referring to Figure 1 As shown, it includes: a denoising autoencoder model, a clustering model and a prediction model. The output end of the denoising autoencoder model is connected to the input end of the clustering model, and the output end of the clustering model is connected to the input end of the prediction model.

[0049] The denoising autoencoder model plays a key role in data preprocessing and feature extraction in the prediction process. The model includes an encoder. The encoder is responsible for feature extraction of the input data. Through learning and training, it can intelligently retain information related to phosphorus content prediction, while removing noise and non-essential information, compressing the original data into a low-dimensional representation, and forming compressed features. This step not only reduces the dimension of the data and the amount of calculation, but also improves the purity and effectiveness of the data.

[0050] The clustering model includes k clustering categories. The clustering model is used to receive the data processed by the denoising autoencoder model and determine the clustering category of the processed data according to the data characteristics.

[0051] Specifically, the clustering model plays a role in data classification and organization in the prediction process. The clustering model takes the data processed by the denoising autoencoder model as input and performs cluster analysis based on the characteristics of the data. Through the clustering algorithm, the model can cluster data with similar characteristics together to form different cluster categories. Each cluster category represents a different specific feature of the data, which helps the subsequent prediction model to make more accurate predictions for different categories of data. The output of the clustering model is to divide the processed data into different cluster categories, providing classified data input for the prediction model.

[0052] The prediction model includes k sub-prediction models. The prediction model is used to receive data after the clustering model determines the clustering category, and input the data into the sub-prediction model corresponding to the clustering category to obtain the data prediction result.

[0053] It can be understood that the prediction model is composed of multiple sub-prediction models, each of which corresponds to a cluster category. When the clustering model clusters the data into a certain cluster category, the prediction model will input the data into the corresponding sub-prediction model for prediction. Since each sub-prediction model is trained for a specific category of data, it can more accurately capture the characteristics and laws of this type of data, thereby improving the accuracy and reliability of the prediction. The output of the prediction model is the final phosphorus content prediction result, which can provide an important basis for the control and optimization of the converter smelting process.

[0054] According to the above embodiments, the prediction model for the phosphorus content at the end of converter smelting proposed in the embodiments of the present application realizes the accurate prediction of the phosphorus content at the end of converter smelting by combining the denoising autoencoder model, the clustering model and the prediction model. The denoising autoencoder model is responsible for data preprocessing and feature extraction, the clustering model is responsible for data classification and organization, and the prediction model is responsible for the final prediction of phosphorus content. This model structure not only improves the accuracy of the prediction, but also adapts to the complexity and nonlinear characteristics of the converter smelting process, providing strong support for the automation and intelligent development of the steel industry.

[0055] In some embodiments of the present application, data features include: molten iron quantity, scrap steel quantity, total oxygen consumption, TSC-C content, TSC-temperature, TSO-C content, TSO-O content, TSO-temperature, C content of molten iron after desulfurization, Si content of molten iron after desulfurization, Mn content of molten iron after desulfurization, P content of molten iron after desulfurization, S content of molten iron after desulfurization, molten iron temperature after desulfurization, sludge ball slag forming agent, sintering return ore slag forming agent, heat generating agent slag forming agent, dolomite slag forming agent, lime slag forming agent, fine crushed powder slag forming agent, magnesium ball slag forming agent, limestone slag forming agent.

[0056] These characteristic data in the prediction model of phosphorus content at the end of converter smelting play an important role in the prediction process. For example, the amount of molten iron directly determines the amount of basic materials in the converter smelting process, which has an important influence on the balance and reaction rate of the chemical reaction in the furnace, thereby indirectly affecting the change of phosphorus content; scrap steel is an important raw material in the smelting process. The amount of scrap steel will change the composition and proportion of the materials in the furnace, affect the temperature in the furnace, the chemical reaction rate, etc., and then affect the change of phosphorus content; the total oxygen consumption reflects the degree of oxidation reaction in the converter, which is closely related to the oxidation and removal process of phosphorus. By monitoring the total oxygen consumption, the degree of phosphorus oxidation and removal effect can be indirectly inferred.

[0057] The relevant parameters of TSC and TSO include: the C content, O content, temperature and other parameters in TSC and TSO reflect the state of the materials in the furnace and the reaction conditions, which directly affect the phosphorus removal efficiency and the final phosphorus content; the relevant parameters of hot metal after desulfurization include: C content, Si content, Mn content, P content, S content and temperature, which directly reflect the effect of hot metal desulfurization treatment, and the change of these parameters will affect the phosphorus removal process. And various slag-making agents, such as sludge balls, sintered ore, heating agent, magnesite, lime, fine crushed powder, magnesium balls, limestone and other materials, reflect the types and quantities of auxiliary materials added in the smelting process, which have an important influence on the balance of chemical reactions in the furnace and the phosphorus removal effect.

[0058] These characteristic data together constitute the material balance and reaction conditions in the converter smelting process, and their changes will directly or indirectly affect the phosphorus removal efficiency and the final phosphorus content. Through real-time monitoring and recording of these characteristic data, rich data support can be provided for the prediction model of the final phosphorus content of converter smelting, so that the model can more accurately capture the changing laws in the smelting process and improve the accuracy of the prediction.

[0059] The second aspect of the present application provides a method for predicting the phosphorus content at the end point of converter smelting, referring to Figure 2 As shown, the prediction method is used for the prediction model of phosphorus content at the end point of converter smelting in any one of the above embodiments, and comprises the following steps:

[0060] S10: Preprocess the test data to obtain initial data; the main purpose of preprocessing is to clean the data, remove noise and outliers, and perform necessary standardization or normalization on the data so that the data meets the input requirements of the model. Preprocessing can ensure the quality of data input into the model and improve the accuracy of prediction.

[0061] S20: Input the initial data into the trained denoising autoencoder model to obtain new data after denoising and compression; the denoising autoencoder model extracts features from the input data through its encoder part, removes noise and non-essential information, and compresses the original data into a low-dimensional representation. This process not only reduces the dimension of the data and the amount of calculation, but also improves the purity and effectiveness of the data.

[0062] S30: Input the new data into the clustering model, and obtain the clustering category of the new data through the clustering model.

[0063] S40: According to the clustering category, the new data is input into the sub-prediction model corresponding to the clustering category to obtain the data prediction result. Since each sub-prediction model is trained for a specific category of data, they can more accurately capture the characteristics and rules of that type of data. Through the processing of the sub-prediction model, we can get the final phosphorus content prediction result. This result is not only based on high-quality data after noise reduction and feature extraction, but also takes into account the clustering characteristics of the data, thereby improving the accuracy and reliability of the prediction.

[0064] The method for predicting the phosphorus content at the end point of converter smelting provided in the second aspect of the embodiment of the present application realizes accurate prediction of the phosphorus content at the end point of converter smelting by preprocessing the test data, denoising autoencoder processing, cluster analysis and sub-prediction model prediction. This method makes full use of the advantages of the denoising autoencoder model and the clustering model to improve the accuracy of the prediction.

[0065] The third aspect of the embodiment of the present application provides a method for constructing a prediction model for the phosphorus content at the end point of converter smelting, the construction method is used to construct a prediction model for the phosphorus content at the end point of converter smelting in any of the above embodiments, referring to Figure 3 As shown, the following steps are included:

[0066] S100: Collect the historical production data of each furnace derived from the converter smelting secondary system in the converter smelting production system; by collecting the historical production data in the converter smelting production system, a large number of real and reliable data samples can be obtained. These data contain various parameters and results in the smelting process, which are important basis for subsequent model training and prediction. In some embodiments, the historical production data is the real production data of the last N months stored in the converter smelting secondary system, where N is greater than or equal to 6, that is, at least half a year of production data is collected.

[0067] Taking a specific implementation example, this construction method exports 20,818 pieces of production data of three converters in the last six months from the secondary system of converter smelting. The table headers include: smelting number, shift, shift type, steel type, molten iron charge, scrap steel charge, pig iron, weight after tapping, weight before tapping, actual molten steel weight, etc.

[0068] S200: Process the historical production data, remove invalid data, abnormal data and data with missing values, and obtain a data set after data preprocessing, which is stored in an Excel table. Since there may be invalid, abnormal or missing values ​​in the original data, these data will interfere with the training of the model and reduce the prediction accuracy of the model. Therefore, it is necessary to clean and organize the data, remove these bad data, and obtain a clean and complete data set.

[0069] Among them, historical production data include at least: molten iron weight, chemical composition information and temperature information after molten iron desulfurization, scrap steel weight, total oxygen consumption in the smelting process, slag-making agent consumption data and phosphorus content data at the end of smelting. Invalid data include at least: shift data, shift data, pig iron data; abnormal data include at least: data outside the business scope during the converter smelting process.

[0070] Invalid data, abnormal data, and data with missing values ​​are meaningless for modeling and are not used as input items. The collected data is stored in an Excel table. Many of the data information corresponding to each furnace number in the table are invalid data, such as shift data, class data, pig iron data, etc. These invalid data are directly deleted; some furnace numbers have missing values ​​due to data collection reasons, for example, some furnace numbers have not collected endpoint phosphorus content data or slag-making agent consumption data, and these data with missing values ​​are deleted; some furnace numbers have data outside the business scope, such as the total oxygen consumption value of the smelting process is negative, which is obviously abnormal data. For example, the scrap steel loading volume is usually 0-40t, but some data is found to be 50t, so it is also abnormal data, so the entire data is deleted and these abnormal data are directly deleted; finally, an initial data set is obtained, which has a total of 11,826 data, each of which contains 22 features and 1 label information.

[0071] The features are used as input items as shown in the table below, and the label information is used as output items, and the label information is the endpoint phosphorus content.

[0072] Table 1 Data characteristics

[0073] .

[0074] S300: Build a denoising autoencoder model, input a data set, train the denoising autoencoder, and complete the training until the loss of the denoising autoencoder model has converged. Save the denoising autoencoder and output the trained data set. The purpose of building a denoising autoencoder is that due to the complex scene of the steel industry, the complex smelting process, and the complex data collection, the data stored in the system inevitably has noise. Building a denoising autoencoder model to reduce noise and compress the data improves the data quality and expression ability.

[0075] In some embodiments of the present application, reference Figure 4 As shown, the denoising autoencoder model also includes: a decoder, which is used to restore the compressed features generated by the encoder to the dimension of the original data and compare them with the original data to confirm the compression effect of the encoder.

[0076] When the encoder compresses the input data into a low-dimensional representation, the data dimension changes as follows: 22 dimensions to 64 dimensions, 64 dimensions to 64 dimensions, 64 dimensions to 32 dimensions, and 32 dimensions to 20 dimensions; among them, the encoder structure is: fully connected layer, ReLU activation layer, fully connected layer, ReLU activation layer, fully connected layer, ReLU activation layer, and fully connected layer. The structure of the decoder is the same as that of the encoder, but the data dimension changes differently. When the decoder restores the low-dimensional representation to the dimension of the original data, the data dimension changes as follows: 20 dimensions to 32 dimensions, 32 dimensions to 64 dimensions, 64 dimensions to 64 dimensions, and 64 dimensions to 22 dimensions; when the encoder compresses the data dimension to 20 dimensions, compressed features are obtained.

[0077] Specifically, in the process of compressing the input data into a low-dimensional representation, the encoder gradually extracts and refines the features of the data through a combination of multiple fully connected layers and ReLU activation layers. First, the original data passes through the first fully connected layer, and the data dimension increases from 22 dimensions to 64 dimensions. This step is to map the original data to a higher-dimensional space in order to better capture the intrinsic characteristics of the data. Subsequently, the ReLU activation layer performs a nonlinear transformation on the data, enhancing the expressive power of the model.

[0078] Next, the data passes through a fully connected layer again, and the dimension remains at 64 dimensions. This step is to further fuse and refine features. After that, the data passes through the third fully connected layer, and the dimension is reduced to 32 dimensions. This step achieves dimensionality reduction of the data, removes redundant information, and retains key features. Finally, after the fourth fully connected layer, the data dimension is further compressed to 20 dimensions, and the final compressed features are obtained. This dimensional change not only significantly reduces the amount of calculation, but also makes it easier for the model to learn the inherent laws and patterns of the data.

[0079] Corresponding to the encoder, the decoder has the same structural layers, but the direction of its data dimension change is opposite to that of the encoder. The role of the decoder is to restore the low-dimensional compressed features generated by the encoder to the dimensions of the original data for subsequent analysis and processing. During the decoding process, the data first passes through a fully connected layer, and the dimension increases from 20 dimensions to 32 dimensions. This step is a preliminary recovery of the compressed features. Then, through the nonlinear transformation of the ReLU activation layer, the features of the data are further enhanced and enriched.

[0080] Next, the data passes through two fully connected layers again, and the dimension is first increased to 64 dimensions and then remains unchanged. This step is to gradually restore the original structure of the data. Finally, after the last fully connected layer, the data dimension is restored to the original 22 dimensions, and the output with the same dimension as the original data is obtained. In this process, the decoder not only restores the dimension of the data, but also retains the useful information in the original data as much as possible, so that the subsequent model can accurately analyze and predict it.

[0081] In summary, the encoder and decoder of the denoising autoencoder model realize feature extraction, compression and recovery of input data through a combination of a series of fully connected layers and ReLU activation layers. In the process of data dimension change, the model gradually removes noise and non-essential information, retains information related to phosphorus content prediction, and provides a strong guarantee for the accuracy of subsequent clustering models and prediction models. This step constructs a denoising autoencoder model and trains it using a preprocessed data set. The denoising autoencoder can automatically extract key features in the data, while removing noise and non-essential information, thereby achieving denoising and compression of the data. Through training, a denoising autoencoder model with excellent performance can be obtained. In the embodiment of the present application, the denoising autoencoder model is mainly composed of an encoder and a decoder. The decoder is used to verify the results of the encoder training. Therefore, after completing the training of the denoising autoencoder model, when the phosphorus content of the data is predicted, the encoder part can be directly called to perform denoising and compression on the data.

[0082] In some embodiments, the parameters of the denoising autoencoder model are set as follows: the number of training rounds epoch is set to 500-5000, the learning rate lr is set to 0.001-0.01, the optimizer selects one of SGD, Adam and RMSprop, and the loss function selects mean square error.

[0083] Reference Figure 5As shown in the figure, the setting of the number of training rounds (epoch) determines the number of times the model traverses the entire data set during the training process. The number of training rounds is set in the range of 500-5000, which can be flexibly adjusted according to the actual data volume and model complexity. More training rounds help the model fully learn the characteristics of the data, but may also lead to overfitting; while fewer training rounds may cause the model to underfit. Therefore, according to the actual situation, the training can be stopped after the loss curve reaches convergence, or the training can be stopped after the number of training rounds is reached. The learning rate (lr) is the step size of the model to update the weights during the training process. The size of the learning rate directly affects the convergence speed and stability of the model. The learning rate is set in the range of 0.001-0.01, and can be fine-tuned according to the training situation of the model. A larger learning rate can accelerate the convergence of the model, but it may also cause the model to oscillate near the optimal solution; while a smaller learning rate can make the model converge to the optimal solution more stably, but the training speed may be slower.

[0084] SGD (Stochastic Gradient Descent), Adam, and RMSprop are commonly used optimizer algorithms. The SGD algorithm is simple and intuitive, but may converge slowly when faced with complex data; the Adam algorithm combines the first-order moment estimation and the second-order moment estimation of the gradient, can adaptively adjust the learning rate, and usually has good convergence performance; the RMSprop algorithm is an improvement on SGD, which adjusts the learning rate by introducing the exponential decay average of the square of the gradient.

[0085] Mean square error (MSE) is a commonly used loss function that calculates the average of the squared differences between the model's predicted value and the actual value. The smaller the mean square error, the better the model's prediction performance. Selecting mean square error as the loss function helps the model continuously reduce prediction errors and improve prediction accuracy during training. It is understandable that in actual operation, the parameters of the denoising autoencoder model can be selected according to actual conditions, and this application does not impose specific restrictions.

[0086] S400: Based on the processed data set, the K-Means clustering algorithm is used to cluster the data, k cluster centers are set by the silhouette coefficient method, and the positions of the k cluster centers are optimized in an iterative manner. The processed data set is divided into different clusters according to the distance from the k cluster centers, and k different sub-data sets and k cluster categories are obtained to obtain a clustering model.

[0087] Specifically, the steps of setting k cluster centers using the silhouette coefficient method include:

[0088] S410: Determine that the number of clusters ranges from 2 to 10, and for each k value within the range of the number of clusters, execute the K-Means clustering algorithm; this range enables the data to be divided into at least 2 subsets, and 10 cluster centers are usually sufficient to describe the structure of most data sets while avoiding overly complex divisions. Then, for each k value within this range, execute the K-Means clustering algorithm.

[0089] S420: For the clustering results under each k value, calculate the silhouette coefficient of each data point, average the silhouette coefficients of all data points under each k value, and obtain the average silhouette coefficient under the k value; the silhouette coefficient is an indicator to measure the clustering effect, which takes into account the similarity (cohesion) of the data point with other data points in its cluster and the similarity (separability) with data points in other clusters. For each data point, the value of its silhouette coefficient is between -1 and 1, and the larger the value, the better the clustering effect. We calculate the silhouette coefficients of all data points under each k value and average them to obtain the average silhouette coefficient under the k value. This average silhouette coefficient can reflect the overall situation of the clustering effect under the k value.

[0090] S430: Draw a contour coefficient curve graph with k value as the horizontal axis and the average contour coefficient as the vertical axis; Figure 6 As shown in the figure, different k values ​​and their corresponding average silhouette coefficients are plotted into a curve graph. The advantage of this is that the trend of the average silhouette coefficient can be observed intuitively as the k value changes. Usually, as the k value increases, the average silhouette coefficient will first increase and then decrease, forming a peak or several local peaks.

[0091] S440: Observe the silhouette coefficient curve and find the k value at which the silhouette coefficient reaches the maximum value or the local maximum value. The ideal k value is the k value that makes the silhouette coefficient reach the maximum value, because this means that the data clustering effect is best under this k value. If there are multiple local maxima in the curve graph, we may need to select the most appropriate k value based on the actual situation (such as the characteristics of the data, the purpose of clustering, etc.).

[0092] In the embodiment of the present application, in the converter smelting process, the data used does not distinguish between steel types, so it can be considered that the entire data set can be divided into different categories. Using the K-Means clustering model and the silhouette coefficient method, it can be intuitively concluded through the silhouette coefficient curve that the k value is 3 with a peak value, and finally determined that the data set is divided into 3 categories. Therefore, the new data set data_new is divided into data_new_0, data_new_1, and data_new_2.

[0093] S500: Based on the Stacking strategy, a prediction model consisting of a base learner and a meta learner is established, and k sub-prediction models are trained based on k different sub-datasets. Stacking is an integrated learning method that improves the overall prediction performance by combining the prediction results of multiple base learners. Based on the three sub-datasets obtained previously, three sub-prediction models are trained separately. These sub-prediction models will be used as base learners and combined and optimized in the subsequent Stacking process.

[0094] In some embodiments, reference Figure 7 As shown, the base learners include: deep forest regression, random forest regression, CatBoost regression and XGBoost regression; the meta learner includes linear regression.

[0095] S600: Install the Python programming environment on the industrial computer. The Python programming environment is used to run Python programs on the industrial computer, load the denoising autoencoder model, clustering model, and prediction model, and use the flask open source library to access or call the denoising autoencoder model, clustering model, and prediction model.

[0096] In the embodiment of the present application, a Python programming environment is installed to run Python programs on the industrial computer. At the same time, the trained denoising autoencoder model, clustering model and prediction model are loaded to ensure that the model can be quickly loaded and used when needed. In addition, the flask open source library is used to provide an interface for model access and calling, so that other systems or modules can easily call these models for predictive analysis.

[0097] After the application scheme was deployed and launched, the data of 2366 furnaces in the last two months were predicted, and the hit rate of this method was statistically analyzed. The hit rates of this method in the absolute error range of [-0.003, 0.003], [-0.005, 0.005] and [-0.007, 0.007] were 66.89%, 88.06% and 93.93% respectively, which is significantly superior to other methods. See the table below for details.

[0098] Table 2

[0099] .

[0100] According to the above content, the present application provides a prediction model, method and model construction method for the phosphorus content at the end point of converter smelting. First, historical production data is collected, and data preprocessing is performed to obtain an initial data set. On the basis of the initial data set, a denoising autoencoder is trained and saved, and the trained denoising autoencoder is used to denoise and compress the data to obtain a new data set. K-Means clustering is used to divide the data set into multiple sub-data sets, and multiple sub-prediction models are trained based on multiple sub-data sets using an integrated model. Finally, the denoising autoencoder model, K-Means clustering model, and sub-prediction model are deployed online, and the input data can be used to obtain the predicted value of the phosphorus content at the end point through the model. The present application scheme covers data of all steel grades and improves the hit rate of the prediction of the phosphorus content at the end point of converter smelting, providing technical support for shortening the smelting cycle and realizing automated steelmaking in digital factories.

[0101] It should be noted that the terms "include", "comprises" or any other variants thereof are intended to cover non-exclusive inclusion, so that a mechanism, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such mechanism, article or device. In the absence of further restrictions, an element defined by the phrase "includes..." does not exclude the presence of other identical elements in the mechanism, article or device including the element.

[0102] Those skilled in the art will readily appreciate other embodiments of the present disclosure after considering the specification and practicing the disclosure herein. This application is intended to cover any variations, uses or adaptations of the present disclosure, which follow the general principles of the present disclosure and include common knowledge or customary techniques in the art that are not disclosed in the present disclosure. The specification and examples are to be considered exemplary only, and the true scope and spirit of the present disclosure are indicated by the following claims.

Claims

1. A model for predicting phosphorus content at the end point of converter smelting, characterized in that: include: Denoising autoencoder models, clustering models, and prediction models; The output end of the denoising autoencoder model is connected to the input end of the clustering model, and the output end of the clustering model is connected to the input end of the prediction model; The denoising autoencoder model comprises: an encoder, which is used to extract input data features, retain relevant information, remove noise and non-essential information, and compress the input data into a low-dimensional representation to obtain compressed features; The clustering model includes k clustering categories, and the clustering model is used to receive the data processed by the denoising autoencoder model, and determine the clustering category of the processed data according to the data characteristics; The prediction model includes k sub-prediction models, and the k sub-prediction models correspond to the k cluster categories one by one. The prediction model is used to receive data after the cluster category is determined by the cluster model, and input the data into the sub-prediction model corresponding to the cluster category to obtain a data prediction result; The data features include: molten iron quantity, scrap steel quantity, total oxygen consumption, TSC-C content, TSC-temperature, TSO-C content, TSO-O content, TSO-temperature, C content of molten iron after desulfurization, Si content of molten iron after desulfurization, Mn content of molten iron after desulfurization, P content of molten iron after desulfurization, S content of molten iron after desulfurization, molten iron temperature after desulfurization, sludge ball slag forming agent, sintering return ore slag forming agent, exothermic agent slag forming agent, magnesite slag forming agent, lime slag forming agent, fine crushed powder slag forming agent, magnesium ball slag forming agent, limestone slag forming agent.

2. A method for predicting phosphorus content at the end point of converter smelting, characterized in that: The prediction model for the phosphorus content at the end point of converter smelting as claimed in claim 1 comprises the following steps: Preprocess the test data to obtain initial data; Inputting the initial data into the trained denoising autoencoder model to obtain new data after denoising and compression; Inputting the new data into the clustering model, and obtaining the clustering category of the new data through the clustering model; According to the clustering category, the new data is input into the sub-prediction model corresponding to the clustering category to obtain a data prediction result.

3. A method for constructing a prediction model for phosphorus content at the end point of converter smelting, characterized in that: The method for constructing the prediction model of phosphorus content at the end point of converter smelting as claimed in claim 1 comprises the following steps: Collect the historical production data of each furnace derived from the converter smelting secondary system in the converter smelting production system; Processing the historical production data to remove invalid data, abnormal data, and data with missing values ​​to obtain a data set after data preprocessing; Construct a denoising autoencoder model, input the data set, train the denoising autoencoder, and complete the training until the loss of the denoising autoencoder model has converged, save the denoising autoencoder and output the trained data set; According to the processed data set, the data is clustered using a K-Means clustering algorithm, k cluster centers are set by a silhouette coefficient method, the positions of the k cluster centers are optimized in an iterative manner, the processed data set is divided into different clusters according to the distances from the k cluster centers, k different sub-data sets and k cluster categories are obtained, and a clustering model is obtained; A prediction model consisting of a base learner and a meta learner is established based on a Stacking strategy, and k sub-prediction models are obtained by training according to the k different sub-data sets; A Python programming environment is installed on the industrial computer, and the Python programming environment is used to run Python programs on the industrial computer, load the denoising autoencoder model, clustering model and prediction model, and use the flask open source library to access or call the denoising autoencoder model, clustering model and prediction model.

4. The method for constructing a prediction model for phosphorus content at the end point of converter smelting according to claim 3, characterized in that: The denoising autoencoder model further includes: a decoder, the decoder being used to restore the compressed features generated by the encoder to the dimension of the original data and compare them with the original data to confirm the compression effect of the encoder; The encoder compresses the input data into a low-dimensional representation, and the data dimension changes are: 22 dimensions to 64 dimensions, 64 dimensions to 64 dimensions, 64 dimensions to 32 dimensions, and 32 dimensions to 20 dimensions; When the decoder restores the low-dimensional representation to the dimension of the original data, the data dimension changes are: 20 dimensions to 32 dimensions, 32 dimensions to 64 dimensions, 64 dimensions to 64 dimensions, and 64 dimensions to 22 dimensions; The compression feature is obtained when the encoder compresses the data dimension to 20 dimensions.

5. The method for constructing a prediction model for phosphorus content at the end point of converter smelting according to claim 3, characterized in that: The parameters of the denoising autoencoder model are set as follows: the number of training rounds epoch is set to 500-5000, the learning rate lr is set to 0.001-0.01, the optimizer selects one of SGD, Adam and RMSprop, and the loss function selects mean square error.

6. The method for constructing a prediction model for phosphorus content at the converter smelting endpoint according to claim 3, characterized in that: The steps of setting k cluster centers using the silhouette coefficient method include: Determine a cluster number range of 2 to 10, and for each k value within the cluster number range, perform a K-Means clustering algorithm; For each clustering result under the k value, the silhouette coefficient of each data point is calculated, and the silhouette coefficients of all data points under each k value are averaged to obtain the average silhouette coefficient under the k value; With the k value as the horizontal axis and the average silhouette coefficient as the vertical axis, a silhouette coefficient curve graph is drawn; Observe the silhouette coefficient curve and find the k value at which the silhouette coefficient reaches a maximum value or a local maximum value.

7. The method for constructing a prediction model for phosphorus content at the end point of converter smelting according to claim 3, characterized in that: The base learners include: deep forest regression, random forest regression, CatBoost regression and XGBoost regression; the meta learner includes linear regression.

8. The method for constructing a prediction model for phosphorus content at the end point of converter smelting according to claim 3, characterized in that: The historical production data is the actual production data of the last N months stored in the converter smelting secondary system, where N is greater than or equal to 6.

9. The method for constructing a prediction model for phosphorus content at the converter smelting endpoint according to claim 3, characterized in that: The historical production data includes at least: molten iron weight, chemical composition information and temperature information after molten iron desulfurization, scrap steel weight, total oxygen consumption in the smelting process, slag-making agent consumption data and phosphorus content data at the end of smelting; the invalid data includes at least: shift data, shift data and pig iron data; the abnormal data includes at least: data outside the business scope in the converter smelting process.

Citation Information

Patent Citations

  • Charging station electricity consumption prediction method based on deep clustering model

    CN112785029A

  • Steel surface quality detection method based on semi-supervised depth clustering

    CN113077461A

  • Method for predicting end-point carbon content and end-point temperature of small and medium-sized converter

    CN115456264A

  • Converter steelmaking process mode design method based on deep clustering generative adversarial network

    CN116127345A

  • Molten iron yield prediction method

    CN116777071A