A method for predicting tapping quantity of a converter with alloy baking equipment

Abstract

The present application belongs to the technical field of steel metallurgy, and particularly relates to a tapping quantity prediction method for a converter with an alloy baking device. The tapping quantity prediction method combines independent component analysis (ICA) with an LSTM long short-term memory recurrent neural network, screens and trains molten iron and scrap steel information that can be obtained before alloy baking and weighing, finds a nonlinear relationship between the molten iron, scrap steel information and the tapping quantity, and can quickly and accurately predict the tapping quantity of the converter before the alloy is baked and weighed after the raw materials are charged into the converter, in the early stage of the converter process, which is beneficial to narrow composition control of the molten steel composition after alloying in the steelmaking process and improves the quality stability of the molten steel.

Description

Technical Field

[0001] This invention belongs to the field of iron and steel metallurgy technology, specifically a method for predicting the steel output of a converter equipped with alloy baking equipment. Background Technology

[0002] Converter steelmaking is a complex physicochemical process, and alloying is a crucial step. The tapping rate is a key parameter for calculating the alloy content, but this can only be obtained through weighing after tapping is complete. Most converters are now equipped with alloy baking equipment, requiring the alloy to be baked in a baking chamber for 15-20 minutes. Therefore, after the molten iron and scrap steel are charged, the alloy must be weighed and added. This necessitates predicting the tapping rate to guide the alloy weighing. At this stage, the only available data is information about the charged molten iron and scrap steel. Therefore, predicting the tapping rate using this limited information is of great significance.

[0003] Current steel companies primarily rely on manual experience and full-parameter network model prediction to predict converter steel output. Manual prediction uses a linear empirical formula derived from the relationship between the quality of molten iron and scrap steel charged and the actual output. However, the converter blowing process is complex and prone to splashing, making it impossible to accurately describe the output and charge amounts linearly, resulting in low accuracy for manual prediction. Full-parameter network model prediction collects a large amount of data from the entire converter blowing process and uses intelligent network algorithms to fit a nonlinear mapping to predict the output. While this method is relatively accurate for centrally loaded converters without alloy baking, it is unusable for converters with alloy baking equipment, where the blowing process has just begun and subsequent parameters are unavailable. Summary of the Invention

[0004] To address the problems existing in the prior art, the main objective of this invention is to propose a method for predicting the steel output of a converter equipped with an alloy baking device, aiming to solve the current problem of difficulty in predicting the steel output before alloy baking and weighing in a converter equipped with an alloy baking device.

[0005] To address the aforementioned technical problems, according to one aspect of the present invention, the present invention provides the following technical solution:

[0006] A method for predicting the steel output of a converter equipped with alloy baking equipment includes the following steps:

[0007] S1. Establish a data transmission program for the manufacturer's server database;

[0008] S2. Establish a predictive model database and receive historical datasets and real-time data sent from the manufacturer's server database;

[0009] S3. Preprocess the historical dataset to filter and remove historical data with abnormal collection.

[0010] S4. Establish a real-time data filtering program and a real-time data manual input program;

[0011] S5. Perform ICA (Independent Component Analysis) on the preprocessed historical dataset;

[0012] S6. Establish an LSTM long short-term memory neural network converter steel output prediction model based on WOA whale optimization algorithm;

[0013] S7. The prediction model is trained and validated using a pre-processed and filtered historical dataset;

[0014] S8. Collect and filter real-time data from the on-site smelting process;

[0015] S9. Enable the manual input module for the abnormal data collected and filtered;

[0016] S10. Perform ICA extraction of independent components on real-time data;

[0017] S11. Input the dimensionality-reduced real-time data into the trained model to predict the steel quantity;

[0018] S12. After the prediction is completed, the input real-time data and the predicted steel output value are transmitted to the historical data in the model database;

[0019] S13. Periodically re-validate the trained model using data from historical furnaces at regular intervals.

[0020] As a preferred embodiment of the method for predicting the steel output of a converter with alloy baking equipment according to the present invention, in step S1, the data in the manufacturer's server database includes: furnace batch number, production date, steel type, ladle condition, molten iron ladle number, molten iron temperature, molten iron weight, molten iron element content (e.g., C, Si, Mn, P, S), scrap steel temperature, scrap steel weight, scrap steel category (self-produced scrap steel, DRI balls, pure scrap steel, magnetically separated slag steel, warm-cut head, crushed material, high-grade magnetic separation powder, sheared furnace charge, screened slag steel, tundish lump, heavy scrap, casting residue slag, self-produced steel slag), as well as the weight of each category, scrap steel ratio, and steel output weight, etc.

[0021] As a preferred embodiment of the steel output prediction method for a converter with alloy baking equipment described in this invention, in step S1, the method for establishing the manufacturer's server database data transmission program is as follows: install the MySQL database driver jar package on the server, connect to and operate the MySQL database through JDBC (Java Database Connectivity) to send and receive historical datasets and real-time data.

[0022] As a preferred embodiment of the method for predicting the steel output of a converter with an alloy baking device according to the present invention, in step S2, the model is written in MATLAB 2023a App Designer, and the model is packaged using the MATLABRuntime program so that it can be installed and run independently.

[0023] As a preferred embodiment of the steel output prediction method for a converter with alloy baking equipment described in this invention, in step S2, the model database type is also a MySQL database, used to receive data sent by the manufacturer's server database. On the computer with the MySQL database and the steel output prediction model installed, a data connection between the steel output prediction model and the model database is established by installing the ODBC (Open Database Connectivity) driver, avoiding direct connection of the model to the manufacturer's server database and improving connection security.

[0024] As a preferred embodiment of the method for predicting the steel output of a converter with alloy baking equipment according to the present invention, wherein: in step S3, the standardization of the historical dataset adopts the Z-standardization method, and the specific processing method is shown in equation (1):

[0025] (1)

[0026] In the formula, i Indicates the first i Type of parameters, j Indicates the first j One sample point, The raw data collected, For the first i The method for calculating the average value of the parameters is shown in equation (2). For the first i The standard deviation of the parameter is calculated as shown in equation (3).

[0027] (2)

[0028] (3)

[0029] In the formula, For the first i The number of sample points for each parameter.

[0030] In a preferred embodiment of the method for predicting the steel output of a converter with an alloy baking device according to the present invention, in step S3, the historical dataset is preprocessed to filter and remove historical data with abnormal collection, specifically as follows:

[0031] After deleting duplicate data, regression interpolation is performed on the abnormal and extreme abnormal data. The standardized data satisfies the standard normal distribution. According to the 3σ principle, the probability of the value distribution in (μ-σ,μ+σ) is 0.6826, the probability of the value distribution in (μ-2σ,μ+2σ) is 0.9545, and the probability of the value distribution in (μ-3σ,μ+3σ) is 0.9973. Therefore, the extreme abnormal data distributed outside (μ-3σ,μ+3σ) are deleted, and the abnormal data distribution between (μ-3σ,μ-2σ) and (μ+2σ,μ+3σ) is fitted and interpolated. The fitting function is the standard normal distribution function and the inverse operation of standardization, as shown in equations (4)-(5).

[0032] (4)

[0033] (5)

[0034] In the formula, For those in the range (μ-3σ, μ-2σ) and (μ+2σ, μ+3σ) Z-standardized anomaly probability density function value This is the normal Z-standardized value that originally corresponded to the probability density function value of this anomaly. Obtained after inverse standardization Interpolation value.

[0035] As a preferred embodiment of the method for predicting the steel output of a converter with an alloy baking device according to the present invention, in step S4, the 3σ principle of historical datasets is also used to judge abnormal data for real-time data. For abnormal and extreme abnormal data, the manual input module of the model is enabled, and the readings obtained directly from the primary system instruments can be filled into the model. The model performs prediction calculations based on the manually input parameter values.

[0036] As a preferred embodiment of the method for predicting the steel output of a converter with alloy baking equipment according to the present invention, wherein: in step S5, the independent variable analysis (ICA) is a data dimensionality reduction method used to extract independent features of the data, remove the correlation between input data features, and find independent components, the processing steps are as follows:

[0037] S51. Input standardized data;

[0038] S52. Whitening: Transform the data so that the covariance matrix becomes a diagonal matrix to remove correlation;

[0039] S53. Determine the number m of independent components to be estimated;

[0040] S54. Calculate the weight vector of each parameter within each independent component;

[0041] S55. Gradient-based optimization, updating the weight vector of each independent component using the expectation rule of a nonlinear g-function. w i ;

[0042] S56. Perform symmetric orthogonalization on the weight matrix W: If convergence is not achieved, return to step S55 and continue the update process until convergence is achieved and the dimensionality-reduced data is output.

[0043] As a preferred embodiment of the method for predicting the steel output of a converter with an alloy baking device according to the present invention, in step S6, the LSTM long short-term neural network model is mainly composed of LSTM units. The LSTM unit structure includes an input gate, an output gate, and a forget gate. The LSTM unit state includes a long memory state and a short memory state. The calculation formulas corresponding to the structure and state are shown in equations (6)-(11):

[0044] (6)

[0045] (7)

[0046] (8)

[0047] (9)

[0048] (10)

[0049] (11)

[0050] In the formula, It's an input gate. It is a sigmoid activation function. It is the hyperbolic tangent activation function. It is a weight matrix. yes t Short memory state at time -1 This is the current input. It is a deviation term. It's an output gate. It controls the forgetting gate. It is the gate to retrieve the forgotten information. It is an LSTM unit t The long memory state at time -1 It is an LSTM unit t The state of long-term memory at any given moment yes t Short-term memory state of a moment.

[0051] As a preferred embodiment of the method for predicting the steel output of a converter with an alloy baking device according to the present invention, in step S6, the hyperparameters of the LSTM long short-term neural network are determined by designing orthogonal experiments. The determined hyperparameters include the number of LSTM units, the number of layers, the sequence length, the batch size, the learning rate, the dropout rate, the activation function, the loss function, the initialization, the gradient pruning, the optimizer, the early stopping parameter, etc.

[0052] In a preferred embodiment of the steel output prediction method for a converter with alloy baking equipment described in this invention, in step S6, the initial weight matrix before the iteration begins is optimized using the WOA whale optimization algorithm. The WOA whale optimization algorithm is derived from biomimetic research on the hunting behavior of humpback whale pods. It simulates the unique bubble-net foraging method of humpback whales, dividing the search for the global optimal solution into three stages: prey encirclement, bubble-net foraging, and prey searching. The optimization of these three stages can be explained by the following mathematical formula:

[0053] Encirclement and capture phase:

[0054] (12)

[0055] In the formula, These are the coordinates of the optimal solution. These are the coordinates of the current solution. t It is the number of iterations. This is the Hadamard product operator. The absolute value operator is used for each element, where A and C are two vectors. The calculation formula is as follows:

[0056] (13)

[0057] (14)

[0058] In the formula, 、 A random vector between [0,1] a It is a constant that decreases linearly from 2 to 0 with the number of iterations. This is the dot product operator.

[0059] Search phase:

[0060] (15)

[0061] (16)

[0062] In the formula, The coordinates of the random solution. This represents the distance between the current solution and the random solution.

[0063] Predation phase:

[0064] (17)

[0065] (18)

[0066] In the formula, D This represents the distance between the coordinates of the current solution and the coordinates of the optimal solution. b For the spiral shape parameters, l It is a random number between [-1, 1].

[0067] There are two ways to approximate the optimal solution during the iteration process: contraction encirclement and spiral encirclement, and the mathematical formulas are shown below:

[0068] (19)

[0069] In the formula, p It is a random number between [0,1].

[0070] As a preferred embodiment of the method for predicting the steel output of a converter with an alloy baking device according to the present invention, in step S7, the preprocessed and filtered historical dataset is divided into a training set, a validation set, and a test set, with a division ratio of 3:1:1.

[0071] As a preferred embodiment of the method for predicting the steel output of a converter with alloy baking equipment according to the present invention, in step S8, the real-time data collected on site comes from the database of the manufacturer's server MES system.

[0072] As a preferred embodiment of the steel output prediction method for a converter with alloy baking equipment according to the present invention, in step S12, the number of historical dataset entries used for training, verifying, and testing the prediction model is fixed at 2000, which are stored separately in an xlsx file using the algorithm written in the model. Whenever the data collection and prediction of a new furnace are completed, the data of that furnace is written to the last line of the xlsx file, and the data lines except the first line are shifted up one line, and the data in the first line is excluded from the xlsx file.

[0073] As a preferred embodiment of the method for predicting the steel output of a converter with alloy baking equipment according to the present invention, in step S13, the .xlsx file of the historical dataset is updated every 500 heats, and the prediction model is trained, verified and tested using the new historical dataset to ensure that the prediction model adapts to the furnace condition and actual situation of the converter, and achieves the self-updating and self-learning functions of the model.

[0074] The beneficial effects of this invention are as follows:

[0075] This invention proposes a method for predicting the steel output of a converter equipped with an alloy baking device. It combines Independent Component Analysis (ICA) with an LSTM (Long Short-Term Memory) recurrent neural network to screen and train information on molten iron and scrap steel that can be obtained before alloy baking and weighing. This method finds the nonlinear relationship between molten iron and scrap steel information and the steel output, enabling rapid and accurate prediction of the converter steel output in the early stage of the converter process, after the raw materials are loaded and before alloy baking and weighing. This facilitates narrow composition control of molten steel after alloying and improves the stability of molten steel quality. Attached Figure Description

[0076] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the structures shown in these drawings without creative effort.

[0077] Figure 1 This is a schematic flowchart of the method for predicting the steel output of a converter equipped with an alloy baking device according to the present invention.

[0078] Figure 2 Flowchart of ICA independent component extraction method;

[0079] Figure 3 Flowchart for WOA whale optimization method;

[0080] Figure 4 This is a schematic diagram of the LSTM (Long Short-Term Memory) neural network structure.

[0081] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0082] The technical solutions described below in conjunction with the embodiments will be clearly and completely described. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0083] Currently, there is no effective method for accurately predicting the tapping weight of converters equipped with alloy baking equipment before weighing the alloy. The main difficulty lies in the fact that the prediction time point is in the early stage of blowing, when a large number of blowing parameters are unavailable, resulting in insufficient known information. Steel plants are now further refining the information of the raw materials entering the furnace, leading to technologies such as hot metal grading and scrap steel grading. The emergence of Long Short-Term Memory (LSTM) neural networks has also made it possible for the prediction network model to not rely on blowing information not available for the current heat. Abundant and detailed raw material information and LMST (Long Short-Term Memory) have created conditions for accurately predicting the tapping weight before alloy baking and weighing.

[0084] Considering the complex reactions within the converter during the actual smelting process, it is difficult to obtain accurate predictions using conventional linear fitting. Therefore, LSTM neural network fitting is chosen to obtain the nonlinear relationship between the input parameters and the steel output, which has excellent approximation performance and generalization ability.

[0085] Considering that the only information available before alloy baking and weighing includes parameters of the incoming molten iron and scrap steel, the LSTM (Long Short-Term Memory) neural network model was chosen. This structure introduces a gating mechanism to effectively solve the gradient vanishing problem, thus enabling it to handle longer sequence data. Unlike general neural network structures, LSTM, through cell states and gating mechanisms, can better capture long-term dependencies in sequence data. LSTM has excellent memory performance, retaining distant contextual information when processing sequence data. LSTM is also time-sensitive, able to learn patterns and features in time-series data. Therefore, the LSTM network structure is more suitable for predicting steel output in converters with alloy baking when fewer parameters are available and blowing process parameters are not yet obtained. Instead of directly using blowing process parameters that cannot be obtained before alloy baking and weighing, the molten iron and scrap steel parameters are used as prediction parameters. After obtaining the predicted steel output value, the blowing process parameters are then used as memory parameters to continuously correct the accuracy of the steel output in subsequent heats.

[0086] The historical dataset is processed using Independent Component Analysis (ICA) to extract independent parameters, thereby improving the model's training efficiency and prediction accuracy. The Whale Optimization Algorithm (WOA) is used to optimize the initial weight matrix before iteration, reducing fitting errors caused by improper weight initialization.

[0087] The technical solution of the present invention will be further described below with reference to specific embodiments.

[0088] Example 1

[0089] A steel plant has a 50t top-blown converter with an alloy baking bin. The average tap weight is 53.49t, but the actual tap weight fluctuates between 45.06 and 65.74t. After the converter is charged with molten iron and scrap, operators weigh the alloy from the alloy bin and send it to the alloy baking bin for preheating and de-crystallization. The alloy weighing quantity needs to be calculated based on the alloy element requirements of the steel grade and the tap weight. At this stage, with the blowing process just beginning, only the parameters of the molten iron and scrap are available. A linear formula fitting the total weight of molten iron, total weight of scrap, and actual tap weight is used for calculation. However, due to the large fluctuation range of the actual tap weight, the predicted calculation results deviate significantly from the actual values. The plant's steelmaking process strictly controls the range of alloy element composition in the molten steel leaving the station. Within 0.02%, a predicted steel output exceeding the actual value leads to excessive alloy baking feed, resulting in substandard molten steel composition and increased production costs. Conversely, a predicted output below the actual value results in insufficient alloy baking feed, necessitating additional alloy feed to the ladle, causing a drop in molten steel temperature and an increase in hydrogen content. Furthermore, parameters acquired during empirical prediction based on data prior to alloy baking are not fully utilized. Iron element content and temperature, scrap steel classification, and the weight of various scrap types all influence the actual steel output. Parameters from the blowing process are not yet available at this stage and cannot be used. Therefore, accurate prediction of the steel output from converters with alloy baking, based on collected parameters before weighing the baking alloy, is crucial for smooth converter production, cost savings, and improved molten steel composition stability.

[0090] Production data from 2508 heats of the 50t converter at the plant were collected. This initial dataset was standardized and filtered. Based on the 3σ principle, extreme outliers outside the range (μ-3σ, μ+3σ) were removed. Outliers between (μ-3σ, μ-2σ) and (μ+2σ, μ+3σ) were fitted and interpolated. The standardization formula, fitting function, and inverse standardization operation are shown below:

[0091] (1)

[0092] (2)

[0093] (3)

[0094] (4)

[0095] (5)

[0096] After screening, historical heat data for 2398 heats were obtained. The data was then Z-standardized again, and ICA independent correlation analysis was performed on the standardized data for the following 30 parameters corresponding to the 2398 sample points: heat batch number, production date, steel grade, ladle condition, hot metal ladle number, hot metal temperature, hot metal weight, hot metal element content (C, Si, Mn, P, S), scrap temperature, scrap weight, scrap category (self-produced scrap, DRI balls, pure scrap, magnetically separated slag steel, warm-cut head, crushed material, high-grade magnetic separation powder, sheared furnace charge, screened slag steel, tundish lump, heavy scrap, casting residue slag, self-produced steel slag), as well as the weight of each category, scrap ratio, and tapped weight.

[0097] Sixteen independent components were extracted from all independent parameters and used as input variables for the LSTM (Long Short-Term Neural Network). The network structure and the values ​​of its hyperparameters were determined using orthogonal experimentation. 2398 standardized independent components were input into the model. The first 1438 sets of data were selected as the training set, the next 480 sets as the validation set to prevent overfitting, and the last 480 sets as the test set to test the model's predictive ability.

[0098] The coefficient of determination (R²) is used to determine the model's predictions. 2 ), Root Mean Square Error (RMSE), Mean Absolute Error (MAE), Mean Percentage Error (MAPE), Error Inside hit rate ( The evaluation was conducted using five parameter standards, and the detailed evaluation results are shown in Table 1.

[0099] Table 1. Hit rates of test set error evaluation parameters and prediction errors within 3, 2, and 1 ton, respectively.

[0100]

[0101] As can be seen from the above error evaluation parameters, the ICA-LSTM model established through training and validation uses only the parameters of molten iron and scrap steel available before alloy baking, without using parameters of the blowing process, to make a relatively accurate prediction of the steel output. On the test set, all error parameters reach a good level, and the error between the predicted and actual steel output values ​​is within a certain range. Furnaces with a capacity of less than 1 ton accounted for 89.51%.

[0102] The established ICA-LSTM converter steel output prediction model with alloy baking was encapsulated into computer operating software using MATLAB Runtime. Simultaneously, JDBC connections were established between the computer database and the manufacturer's MES server database, and ODBC connections were established between the model and the computer database. Before weighing and baking the alloy in a heat, 30 parameters of molten iron and scrap steel were transmitted to the model. The model underwent preprocessing standardization, anomaly detection, and fitting interpolation. For extreme anomalies, manual data entry was initiated, with operators inputting parameters from primary equipment instruments into the model. After this operation, the model predicted the steel output. The predicted steel output value was obtained after the molten iron and scrap steel were loaded but before the alloy baking weighing was completed, serving as a crucial basis for guiding alloy baking weighing. After the steel tapping process was completed and the actual steel output was obtained through weighing, all data from that heat was collected and input into a historical dataset. After 500 heats, the ICA-LSTM model was retrained using the same method based on the new historical dataset, continuously reducing the steel output prediction error.

[0103] This invention combines Independent Component Analysis (ICA) with an LSTM (Long Short-Term Memory) neural network. Considering that converters with alloy baking require steel output prediction after charging, and many blowing parameters are unavailable due to limited known information, this invention fully utilizes the various parameters of molten iron and scrap steel grading obtained before alloy baking and weighing. These parameters are then analyzed using ICA and used as input to the LSTM neural network. This combines the advantages of both methods: ICA performs independent analysis and extraction of data, improving the efficiency and accuracy of network training; and the LSTM neural network treats each heat of the ongoing steelmaking process as a continuous event. By using parameters obtained before alloy baking to predict steel output before alloy weighing, and simultaneously recording subsequent blowing parameters, the LSTM memory mechanism corrects subsequent predictions of steel output based on molten iron and scrap steel parameters. This achieves accurate steel output prediction before alloy baking, overcoming the shortcomings of inaccurate empirical predictions and the lack of reference value in traditional prediction models due to their late prediction time points. Meanwhile, the model also has the ability to self-update the dataset. After a certain smelting cycle, the ICA-LSTM network is retrained based on the changes in the dataset to prevent changes in furnace conditions from affecting the prediction accuracy.

[0104] The above description is merely a preferred embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural transformations made using the contents of the present invention under the inventive concept of the present invention, or direct / indirect applications in other related technical fields, are included within the patent protection scope of the present invention.

Claims

1. A method for predicting the steel output of a converter equipped with alloy baking equipment, characterized in that, Includes the following steps: S1. Establish a data transmission program for the manufacturer's server database; S2. Establish a predictive model database and receive historical datasets and real-time data sent from the manufacturer's server database; S3. Preprocess the historical dataset to filter and remove historical data with abnormal collection. S4. Establish a real-time data filtering program and a real-time data manual input program; S5. Perform ICA (Independent Component Analysis) on the preprocessed historical dataset; S6. Establish an LSTM long short-term memory neural network converter steel output prediction model based on WOA whale optimization algorithm; S7. The prediction model is trained and validated using a pre-processed and filtered historical dataset; S8. Collect and filter real-time data from the on-site smelting process; S9. Enable the manual input module for the abnormal data collected and filtered; S10. Perform ICA extraction of independent components on real-time data; S11. Input the dimensionality-reduced real-time data into the trained model to predict the steel quantity; S12. After the prediction is completed, the input real-time data and the predicted steel output value are transmitted to the historical data in the model database; S13. Periodically re-validate the trained model using data from historical furnaces at regular intervals; In step S3, the historical dataset is standardized using the Z-standardization method, and the specific processing method is shown in equation (1): （1） In the formula, i Indicates the first i Type of parameters, j Indicates the first j One sample point, The raw data collected, For the first i The method for calculating the average value of the parameters is shown in equation (2). For the first i The standard deviation of the parameter is calculated as shown in equation (3). （2） （3） In the formula, For the first i The number of sample points for each parameter; In step S3, the historical dataset is preprocessed to filter and remove historical data with abnormal collection conditions, specifically as follows: After deleting duplicate data, regression interpolation is performed on the abnormal and extreme abnormal data. The standardized data satisfies the standard normal distribution. According to the 3σ principle, the probability of the value distribution in (μ-σ,μ+σ) is 0.6826, the probability of the value distribution in (μ-2σ,μ+2σ) is 0.9545, and the probability of the value distribution in (μ-3σ,μ+3σ) is 0.9973. Therefore, the extreme abnormal data distributed outside (μ-3σ,μ+3σ) are deleted, and the abnormal data distribution between (μ-3σ,μ-2σ) and (μ+2σ,μ+3σ) is fitted and interpolated. The fitting function is the standard normal distribution function and the inverse operation of standardization, as shown in equations (4)-(5). （4） （5） In the formula, For those in the range (μ-3σ, μ-2σ) and (μ+2σ, μ+3σ) Z-standardized anomaly probability density function value This is the normal Z-standardized value that originally corresponded to the probability density function value of this anomaly. Obtained after inverse standardization interpolation value; In step S5, ICA is a data dimensionality reduction method used to extract independent features of data, remove correlations between input data features, and find independent components. The processing steps are as follows: S51. Input standardized data; S52. Whitening: Transform the data so that the covariance matrix becomes a diagonal matrix to remove correlation; S53. Determine the number m of independent components to be estimated; S54. Calculate the weight vector of each parameter within each independent component; S55. Gradient-based optimization, updating the weight vector of each independent component using the expectation rule of a nonlinear g-function. ; S56. Perform symmetric orthogonalization on the weight matrix W: If convergence is not achieved, return to step S55 and continue the update process until convergence is achieved and the dimensionality-reduced data is output.

2. The method for predicting the steel output of a converter with alloy baking equipment according to claim 1, characterized in that, In step S1, the manufacturer's server database data transmission program is established by installing a MySQL database driver jar package on the server, connecting to and operating the MySQL database via JDBC to send and receive historical datasets and real-time data.

3. The method for predicting the steel output of a converter with alloy baking equipment according to claim 1, characterized in that, In step S2, the model is written in MATLAB 2023a App Designer, and the MATLAB Runtime program is used to encapsulate the model so that it can be installed and run independently.

4. The method for predicting the steel output of a converter with alloy baking equipment according to claim 2, characterized in that, In step S2, the model database type is also a MySQL database, used to receive data sent by the manufacturer's server database. On the computer where the MySQL database and the steel output prediction model are installed, an ODBC driver is installed to establish a data connection between the steel output prediction model and the model database, avoiding direct connection of the model to the manufacturer's server database and improving connection security.

5. The method for predicting the steel output of a converter with alloy baking equipment according to claim 1, characterized in that, In step S4, the 3σ principle of historical datasets is also used to judge abnormal data for real-time data. For abnormal and extreme abnormal data, the model manual input module is enabled to fill the readings obtained directly from the primary system instruments into the model. The model performs prediction calculations based on the manually input parameter values.

6. The method for predicting the steel output of a converter equipped with alloy baking equipment according to claim 1, characterized in that, In step S6, the LSTM long short-term neural network model is mainly composed of LSTM units. The structure of the LSTM unit includes an input gate, an output gate, and a forget gate. The LSTM unit states include long memory states and short memory states. The calculation formulas corresponding to its structure and state are shown in equations (6)-(11): （6） （7） （8） （9） （10） （11） In the formula, It's an input gate. It is a sigmoid activation function. It is the hyperbolic tangent activation function. It is a weight matrix. yes t Short memory state at time -1 This is the current input. It is a deviation term. It's an output gate. It controls the forgetting gate. It is the gate to retrieve the forgotten information. It is an LSTM unit t-1 The state of long-term memory at any given moment It is an LSTM unit t The state of long-term memory at any given moment yes t Short-term memory state of a moment.

7. The method for predicting the steel output of a converter equipped with alloy baking equipment according to claim 1, characterized in that, In step S6, the initial weight matrix before the iteration begins is optimized using the WOA whale optimization algorithm. The WOA whale optimization algorithm is derived from the biomimetic study of humpback whale pod hunting behavior. It simulates the humpback whale's unique bubble net foraging method and divides the search for the global optimum into three stages: prey encirclement, bubble net foraging, and prey searching. The optimization of the three stages can be explained by the following mathematical formula: Encirclement and capture phase: （12） In the formula, These are the coordinates of the optimal solution. These are the coordinates of the current solution. t It is the number of iterations. This is the Hadamard product operator. The absolute value operator is used for each element, where A and C are two vectors. The calculation formula is as follows: （13） （14） In the formula, 、 A random vector between [0,1] a It is a constant that decreases linearly from 2 to 0 with the number of iterations. This is the dot product operator. E It is a unit vector; Search phase: （15） （16） In the formula, The coordinates of the random solution. The distance between the current solution and the random solution; Predation phase: （17） （18） In the formula, D This represents the distance between the coordinates of the current solution and the coordinates of the optimal solution. b For the spiral shape parameters, l A random number between [-1, 1]; There are two ways to approximate the optimal solution during the iteration process: contraction encirclement and spiral encirclement, and the mathematical formulas are shown below: （19） In the formula, p It is a random number between [0,1].

Images