Error correction method and system for LF refining furnace molten steel temperature prediction

Abstract

The application discloses an error correction method and system for LF refining furnace molten steel temperature prediction, comprising: a single hidden layer back propagation neural network model, through collecting first smelting data corresponding to factors affecting the LF refining molten steel temperature as a data set, training the model, and constructing an LF furnace refining molten steel temperature prediction model; through collecting second smelting data of the current new furnace, obtaining the current initial molten steel temperature prediction value according to the prediction model; according to the similarity of the second smelting data and the first smelting data, obtaining the molten steel temperature prediction value of the first smelting data corresponding to the historical furnace close to the current new furnace, comparing with the real value of the molten steel temperature, obtaining the prediction error, and correcting the initial molten steel temperature prediction value; the application combines the advantages of case-based reasoning incremental learning and other artificial intelligence algorithms fitting nonlinear relationship, thereby improving the model prediction accuracy.

Description

Technical Field

[0001] This invention relates to the field of intelligent smelting processes and equipment, and more specifically, to an error correction method and system for predicting the temperature of molten steel in an LF refining furnace. Background Technology

[0002] LF refining is an important and commonly used steel refining method in steel production. The goal of LF refining is to ensure that the temperature and composition of the molten steel meet the requirements of the continuous casting process, coordinate the steelmaking-continuous casting production rhythm, and improve production efficiency and product quality. However, due to limitations in on-site working conditions and measuring instruments, continuous and timely measurement of the molten steel temperature cannot be achieved during the refining process. On-site operators cannot obtain timely and accurate temperature information of the molten steel, making precise temperature control difficult. Currently, the refining process requires multiple power outages for offline temperature measurement, inevitably resulting in heat loss. This temperature drop further increases power consumption and correspondingly increases the labor intensity of furnace operators. Therefore, the research and development of LF refining furnace molten steel temperature prediction models is of great significance for improving LF refining control, stabilizing product quality in steel enterprises, and reducing production costs. Currently, prediction models for LF molten steel temperature can be mainly divided into the following categories: mechanistic models, data-driven models, and hybrid models.

[0003] Mechanism models are primarily process models derived from the law of conservation of energy, thermodynamics, and kinetics. For example, Wu Yongjun studied the LF molten steel temperature prediction model using the overall heat balance method. He utilized the law of conservation of energy, taking molten steel and slag as research objects, analyzed the heat balance during the refining process, and derived a model for the steel heating rate. Regarding the heat transfer mechanism between the ladle wall and bottom in the model, he established one-dimensional unsteady-state heat conduction differential equations in cylindrical and rectangular coordinate systems, respectively, and solved them using the finite difference method. However, in actual refining processes, many factors affect the temperature of molten steel, and there are complex nonlinear relationships between them, making it extremely difficult to establish a complete mechanism model. Therefore, many simplifications and assumptions are introduced into the mechanism model, thus sacrificing the model's accuracy.

[0004] Data-driven models often rely on artificial intelligence algorithms to process historical smelting data, determine input and output terms, and then use algorithms to solve for the complex relationships between each input parameter and molten steel temperature. For example, Tian Huixin et al. used backpropagation neural networks, extreme learning machines, and other algorithms, combined with the principles of ensemble learning, to integrate multiple sub-models using an improved adaptive enhancement algorithm to achieve LF molten steel temperature prediction. The accuracy of data-driven models largely depends on the similarity between the model training set and the distribution of actual smelting data. However, due to the numerous complex physical and chemical reactions involved in LF refining and the differences in process parameters between different heats, discrepancies inevitably exist between the model training set and the actual smelting data. Therefore, when data-driven models are applied to new heats, their performance may inevitably decline because the model cannot adapt to these differences.

[0005] A hybrid model is a model that combines multiple mechanistic models and data-driven models. For example, He Fei et al., when predicting the temperature of molten steel during the steelmaking process, considered the influence of the ladle's thermal state on the steel temperature. They established a ladle heat tracking model to calculate the heat loss of the ladle throughout the steelmaking process, and combined the temperature predicted by the backpropagation neural network model with the ladle's thermal state compensation temperature to obtain a hybrid model for predicting the molten steel temperature. However, regardless of the hybrid structure used, the performance of the hybrid model can only be guaranteed if the mechanistic models and data-driven models within it are sufficiently accurate.

[0006] In existing technologies, to improve the accuracy of data-driven models, some scholars have proposed online model training methods. When new data is encountered during model application, the model is retrained using data that is similar to the new data in time or data space to adapt to the new application scenario. For example, Gu Maoqiang et al., in order to achieve dynamic prediction of the molten steel temperature during the second blowing stage of a converter, found cases similar to the new case in a historical case database, and used the process parameters of the second blowing stage of similar cases to establish a long short-term memory neural network model to predict the molten steel temperature change during the second blowing stage of the new case. However, retraining the model is often constrained by training time, training data volume, model structure, and model parameters, which is not conducive to practical applications.

[0007] Some scholars have proposed error correction methods, which involve building an error correction model after the prediction model to predict the errors generated during the prediction process and correct the model's prediction results. For example, Xu et al. applied an error correction method to improve the prediction accuracy of wind speed at wind power plants, based on the Weather Research and Forecasting Model. They used wind speed-related features and historical prediction errors of the weather forecasting model to build a long short-term memory neural network to correct the errors in the weather forecasting model. However, existing error correction models are still trained based on historical data, and their performance still suffers from a decline due to differences between the model training set and the actual data.

[0008] In summary, addressing the decline in model accuracy caused by discrepancies between training and real-world data is crucial for improving predictive accuracy. Furthermore, avoiding retraining when encountering new data facilitates the practical application of the model. Summary of the Invention

[0009] In view of the problems existing in the prior art, in order to overcome the decrease in accuracy of the prediction model in actual application due to the distribution difference between the actual data and the model training data, and to avoid repeated training of the model, the purpose of this invention is to provide an error correction method based on case reasoning (EC_CBR) to improve the accuracy of the prediction model for LF furnace refining molten steel. The method calculates the error of the new furnace by using the error of the prediction model of the furnaces similar to the current new furnace in the historical furnaces, and corrects the predicted value of the new furnace in the prediction model accordingly.

[0010] To achieve the above technical objectives, this application provides an error correction method for predicting the temperature of molten steel in an LF refining furnace, comprising the following steps:

[0011] Construct a backpropagation neural network model based on a single hidden layer;

[0012] Historical smelting data of the LF refining furnace were collected, and the first smelting data corresponding to the factors affecting the temperature of LF refined molten steel were obtained as the dataset. The model was trained to construct an LF furnace refining molten steel temperature prediction model for predicting the temperature of LF refined molten steel.

[0013] Based on the LF furnace refining steel temperature prediction model, the initial steel temperature prediction value for the current new furnace is obtained by collecting the second smelting data of the current new furnace.

[0014] Based on the similarity between the second smelting data and the first smelting data, the predicted steel temperature corresponding to the first smelting data of historical furnaces similar to the current new furnace is obtained. By comparing it with the actual steel temperature of historical furnaces, the prediction error is obtained, the initial predicted steel temperature is corrected, and the predicted steel temperature of the current new furnace is generated.

[0015] Preferably, in the process of acquiring the first smelting data, based on the smelting operation flow of the LF process, the first smelting data corresponding to the influencing factors of the LF refined steel temperature are obtained through energy balance analysis of the refining process.

[0016] Preferably, in the process of constructing the dataset, the first smelting data is preprocessed before the dataset is constructed. The preprocessing process includes: deleting blank values, deleting outliers, and data normalization.

[0017] Preferably, in the process of obtaining similarity, the similarity between the second smelting data and the first smelting data is measured according to Euclidean distance similarity.

[0018] Similarity is expressed as:

[0019]

[0020]

[0021] Where d(X) i X k S represents Euclidean distance. k This represents the similarity score, where m is the number of influencing factors, and x... ij and x kj Let w represent the j-th influencing factor in new furnace data and historical furnace data, respectively. j This represents the weight of the j-th influencing factor.

[0022] Preferably, in the process of obtaining similarity, the weight w of the j-th influencing factor is obtained through the Grey Wolf Optimization Algorithm. j And obtain the number of historical furnace runs;

[0023] Based on the number of data points, the prediction errors of several historical heats are obtained. The error of the initial molten steel temperature prediction value is obtained by weighted averaging. The initial molten steel temperature prediction value is then corrected to generate the molten steel temperature prediction value for the current heat.

[0024] The dataset is updated based on the predicted steel temperature of the current new furnace corresponding to the second smelting data.

[0025] Preferably, the optimization process of the Grey Wolf optimization algorithm conforms to the following constraints:

[0026] The sum of the weights of all factors is 1;

[0027] The weight of any one factor is greater than 0;

[0028] The number of similar furnace cycles ranges from 1 to 100.

[0029] Preferably, in the process of obtaining the predicted value of the molten steel temperature for the current heat, the predicted value of the molten steel temperature for the current heat is expressed as:

[0030]

[0031] Where, f(X) i ) represents the predicted value of the current heat cycle using the steel temperature prediction model, N represents the number of similar historical heat cycles, and q k ε represents the weight of furnace k. k This represents the error of furnace number k in the prediction model.

[0032] Preferably, in the process of obtaining the predicted value of molten steel temperature, the weight q of the heat k is... k Represented as:

[0033]

[0034] Preferably, in the process of obtaining the error of the initial molten steel temperature prediction value, the error of the initial molten steel temperature prediction value is expressed as:

[0035]

[0036] This invention discloses an error correction system for predicting the temperature of molten steel in an LF refining furnace, comprising:

[0037] The data acquisition module is used to acquire the second smelting data for the current new furnace based on the first smelting data corresponding to the factors affecting the temperature of molten steel in the LF refining furnace, which are obtained by acquiring historical smelting data of the LF refining furnace.

[0038] The temperature prediction module is used to obtain the initial predicted temperature of molten steel for the current furnace refining process based on the LF furnace refining steel temperature prediction model and second smelting data. Specifically, it employs a backpropagation neural network model with a single hidden layer, trained using first smelting data to construct the LF furnace refining steel temperature prediction model.

[0039] The temperature correction module is used to obtain the predicted steel temperature corresponding to the first smelting data of historical furnaces similar to the current new furnace based on the similarity between the second smelting data and the first smelting data. By comparing it with the actual steel temperature of historical furnaces, the prediction error is obtained, the initial predicted steel temperature is corrected, and the predicted steel temperature of the current new furnace is generated.

[0040] The present invention discloses the following technical effects:

[0041] This method combines the advantages of incremental learning in case-based reasoning with the advantages of fitting nonlinear relationships in other artificial intelligence algorithms, thereby improving the model's prediction accuracy.

[0042] This invention can effectively calculate the optimal weight allocation and number of cases, thereby helping to accurately retrieve similar furnace times and improve the correction effect on the model. Attached Figure Description

[0043] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments 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 these drawings without creative effort.

[0044] Figure 1 This is a flowchart illustrating the application process of the EC_CBR method described in this invention.

[0045] Figure 2 This is a schematic diagram of the method described in this invention. Detailed Implementation

[0046] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. The components of the embodiments of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely represents selected embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0047] like Figure 1-2 As shown in Example 1: This invention provides an error correction method for predicting the temperature of molten steel in an LF refining furnace, comprising the following steps:

[0048] Step 1: Obtain historical smelting data from the LF refining furnace and preprocess the historical smelting data to obtain a dataset for model training and building a case library;

[0049] Step 2: Establish a prediction model for LF furnace refining molten steel temperature and a production data - molten steel temperature case library;

[0050] Step 3: Collect the smelting data of the current new furnace, calculate the similarity between the new furnace and the historical furnaces in the case library according to formulas (1) and (2), obtain the smelting data and molten steel temperature of the N most similar historical furnaces, and calculate the weight q of each similar furnace according to formula (3). k ;

[0051] In a preferred embodiment of the present invention, Euclidean distance similarity is used to measure the similarity between data, and the Euclidean distance d(X) between the new furnace and case k is... i X k and similarity S k As shown in the following formula:

[0052]

[0053]

[0054] In the formula, m is the number of influencing factors, and x ij and x kj Let w represent the j-th influencing factor in the new furnace data and the historical furnace data in the case library, respectively. j This represents the weight of the j-th influencing factor.

[0055] The weights of influencing factors and the number of similar historical furnaces were calculated using the Grey Wolf optimization algorithm.

[0056] q k The size of this value is related to the similarity between new furnace data and historical furnace data, and can be expressed by the following formula:

[0057]

[0058] Step 4: Input the smelting data of the new furnace and the smelting data of each similar case into the steel temperature prediction model established in Step 2, and obtain the initial predicted value of the steel temperature of the new furnace and the predicted value of the steel temperature of each similar case.

[0059] Step 5: Compare the predicted steel temperature of each similar heat with its actual value, and calculate the prediction error of each similar heat.

[0060] Step 6: According to formulas (4), (5), and (6), calculate the initial prediction error of the new furnace based on the prediction error of similar furnaces, and correct the initial prediction value of the molten steel temperature of the new furnace to obtain the predicted value of the molten steel temperature of the new furnace.

[0061] The weighted average method is used to calculate the error of the new furnace in the prediction model. Based on the similarity between the retrieved similar historical furnaces and the new furnace, a weighted average of the errors of the similar furnaces is calculated to obtain the error ε of the new furnace in the prediction model. i .

[0062]

[0063]

[0064] In the formula, T i The corrected temperature of molten steel for the new furnace, f(X) i ) represents the predicted value of the molten steel temperature prediction model, N represents the number of similar historical heats, and q k q represents the weight of furnace k, and εx represents the error of furnace k in the prediction model. k The size of this value is related to the similarity between new furnace data and historical furnace data, and can be expressed by the following formula:

[0065]

[0066] Step 7: Save the smelting data and measured steel temperature of the new furnace in the case library.

[0067] Example 2: The error correction method for a case-based reasoning model for predicting molten steel temperature in an LF refining furnace disclosed in this invention includes the following technical processes in practical applications:

[0068] (1) Data Preprocessing: Temperature prediction modeling and error correction were performed using actual data from the first quarter of 2021 of a steel plant during the LF process of smelting SPHC11 steel. The main smelting operations in the LF process of this steel plant include electric arc heating, argon blowing and stirring, slag formation, alloy addition, and wire feeding. Based on this smelting data and combined with the energy balance analysis of the refining process, the factors affecting the temperature of molten steel in the LF refining process were obtained from historical production data, including molten steel weight, molten steel arrival temperature, smelting time, power consumption, bottom blowing argon volume, alloy addition (addition of carbonizer, high-carbon ferromanganese, medium-carbon ferromanganese, aluminum granules, and aluminum slag), slag forming agent addition (addition of quicklime, composite deoxidizing slag forming agent, fluorite, and slag flux), and wire feeding length (length of aluminum wire and length of solid calcium-aluminum cored wire). Necessary preprocessing was performed on the LF refining data of this steel plant, such as deleting blank values, deleting outliers, and data normalization, leaving 1495 sets of data. The statistical analysis results of each influencing factor data item and molten steel temperature are shown in Table 1. All influencing factors were used as inputs to the prediction model and as matching attributes for similar heat searches.

[0069] Table 1

[0070]

[0071]

[0072] (2) Temperature prediction model establishment:

[0073] A backpropagation neural network (BPNN) with a single hidden layer was used to predict the temperature of molten steel in an LF furnace. The input to the model consisted of the factors affecting the molten steel temperature analyzed above, and the output was the molten steel temperature. Of the data presented, 80% was randomly selected as the training set and initial case set, while the remaining 20% ​​served as the test set.

[0074] The prediction accuracy of the model is shown in Table 2. The table uses the root mean square error (RMSE) to describe the accuracy of the model, and its calculation method is as follows:

[0075]

[0076] In the formula, T predicted T represents the model's predicted value. actual is the true value; m is the number of predicted samples; j is the starting value of the summation function, i.e., the first batch in the predicted samples.

[0077] Table 2

[0078]

[0079] (3) Search for similar furnaces to the new furnace.

[0080] Euclidean distance similarity is used to measure the similarity between data points. The Euclidean distance d(X) between the new batch of data and case k is... i X k and similarity S k As shown in the following formula:

[0081]

[0082]

[0083] In the formula, m is the number of influencing factors, and x ij and x kj Let w represent the j-th influencing factor in the new furnace data and the historical furnace data in the case library, respectively. j This represents the weight of the j-th influencing factor.

[0084] In this step, a machine learning algorithm—the Grey Wolf Optimization Algorithm—is used to calculate the weights w of each influencing factor in formula (3). j The number of reused similar cases was also calculated. The weights of each influencing factor are shown in Table 3.

[0085] Table 3

[0086]

[0087] When calculating the weights of each influencing factor using the Grey Wolf optimization algorithm, the value of the number of similar furnace cycles N is simultaneously substituted into the algorithm for optimization calculation. The Grey Wolf algorithm optimization process conforms to the following constraints:

[0088] 1. The sum of the weights of all factors is 1;

[0089] 2. The weight of any factor is greater than 0;

[0090] 3. The number of similar furnace cycles ranges from 1 to 100.

[0091] The Grey Wolf optimization algorithm and error correction were implemented using Python programming. The fitness function in the optimization process was taken as the RMSE after model correction. Optimization determined that, under this weight allocation, when N is 5, the BPNN temperature prediction model using this method has the minimum RMSE.

[0092] (4) Prediction of molten steel temperature in new furnaces

[0093] Based on step (3), N historical heats are retrieved as similar heats for the new heat. Their smelting data are then input into the steel prediction model to obtain the predicted values ​​of their respective molten steel temperatures. The error of each similar heat on the prediction model is calculated, and the case is reused to calculate the error of the new heat on the prediction model.

[0094] The weighted average method is a widely used case reuse method in case-based reasoning. In this paper, the errors of similar historical furnaces are weighted and averaged based on the similarity between the retrieved similar furnaces and the new furnace to obtain the error of the new furnace on the prediction model. The final predicted value of molten steel temperature is obtained by subtracting the predicted result of the new furnace on the model (initial predicted value) from the obtained error.

[0095] The error ε of the prediction model for new furnaces i The steel temperature correction value Ti is shown in equations (4) and (5).

[0096]

[0097]

[0098] In the formula, f(X) i ) represents the predicted value of the molten steel temperature prediction model, N represents the number of similar historical heats, and q k ε represents the weight of furnace k. k This represents the error of furnace number k in the prediction model. q k The size of this value is related to the similarity between new furnace data and historical furnace data, and can be expressed by the following formula:

[0099]

[0100] Table 4 summarizes the accuracy of the original model and the model corrected using the BC-CBR algorithm with weights calculated based on the Grey Wolf algorithm. As shown in the table, compared to the original model, the RMSE of the model optimized by the method proposed in this invention is significantly reduced, and the prediction accuracy is significantly improved in the ranges of ±5℃, ±7℃, and ±10℃. The prediction accuracy is improved by approximately 6% in the ±5℃ range, approximately 5% in the ±7℃ range, and the prediction accuracy is greater than 95% in the ±10℃ range. This demonstrates that this method can effectively improve the accuracy of the model.

[0101] Table 4

[0102]

[0103] (5) Expanding the case library:

[0104] After a new batch of refining is completed, the corresponding data of each influencing factor and the actual temperature measurement results in the production data of that batch are added to the case library.

[0105] (6) Comparison with general case reasoning methods:

[0106] The method proposed in this invention combines case-based reasoning with other machine learning algorithms to correct errors in other machine learning models. Unlike general case-based reasoning, the method proposed in this invention reuses the errors of multiple similar cases in the prediction model when reusing cases, rather than directly using the molten steel temperature of similar cases as the molten steel temperature of the new furnace. To compare the effect of this method with general case-based reasoning, a model was built using the general case-based reasoning method to predict the LF molten steel temperature. The training set and test set defined above were used as the case library and test data for the case-based reasoning model. The Grey Wolf optimization algorithm was also used to calculate the influence weight of each factor on the molten steel temperature and the number of similar historical furnaces selected. The weighted average of the molten steel temperature of each similar historical furnace was also calculated based on the Euclidean distance similarity. When optimizing with the Grey Wolf algorithm, the constraints, fitness function, and required parameters were set in the same way as in step (3).

[0107] After iteration, the optimized weights of the influencing factors are shown in Table 5. After optimization, it was determined that when N is 8 under this weight allocation, the general case reasoning method has the highest prediction accuracy for LF molten steel temperature, with an RMSE of 4.63. Calculations show that the hit rates of the general case reasoning method within the ±5℃, ±7℃, and ±10℃ ranges are 74.25%, 86.62%, and 95.99%, respectively. As shown in Table 6, the general case reasoning method has better accuracy than the original BPNN model on the actual smelting data of this steel plant. However, the hit rates of the general case reasoning method within the ±5℃, ±7℃, and ±10℃ ranges are lower than the model modified by the method proposed in this invention, and the RMSE is also larger, proving the effectiveness of the method proposed in this invention.

[0108] Table 5

[0109]

[0110] Table 6

[0111]

[0112] This method combines the incremental learning advantages of case-based reasoning with the strengths of other artificial intelligence algorithms in fitting nonlinear relationships, thereby improving the accuracy of the prediction model. The invention includes the following points:

[0113] This invention proposes an error correction method based on case-based reasoning: The training process of a typical data-driven model involves fitting a function to the data space, minimizing the loss function of the fitted function on the training data. However, to avoid overfitting, the loss function is not zero, thus introducing error in the model's predictions. In the data space, the model's prediction error is the positional deviation between the fitted function and the true values ​​of the cases. The case-based reasoning method retrieves similar cases based on reasonable data feature weights. These similar cases are located close to the new case in the data space, and their relative positions to the fitted function are also similar, resulting in similar errors in the fitted function. Therefore, the case-based reasoning method can predict the error of a new heat in the LF molten steel temperature prediction model. Unlike other data-driven models, the case-based reasoning method does not rely on a specific dataset, does not require a complex training process, and its incremental learning allows the case library to gradually expand with practical applications. The predictive performance of the case-based reasoning method does not decrease with changes in data distribution; instead, it improves with continuous replenishment of the case library.

[0114] This invention uses machine learning algorithms to solve for the feature weights and number of reused cases in case-based reasoning. In the application of case-based reasoning, the weight allocation of data features (factors influencing the target value) largely determines the accuracy of case retrieval results. Currently, commonly used weight determination methods include the average weight method, correlation analysis, entropy weight method, analytic hierarchy process (AHP), and mutual information method. However, these methods often deviate from specific problems, rely on mathematical statistics, or are dominated by human experience, failing to guarantee objectivity. This invention selects a suitable fitness function and clear constraints, transforming the weight allocation and case selection into a parameter optimization problem, which is then solved using machine learning algorithms. This method can effectively calculate the optimal weight allocation and case number, thereby helping to accurately retrieve similar cases and improving the model's correction effect.

[0115] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0116] In the description of this invention, it should be understood that the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0117] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. An error correction method for predicting the temperature of molten steel in an LF refining furnace, characterized in that, Includes the following steps: Construct a backpropagation neural network model based on a single hidden layer; Historical smelting data of the LF refining furnace were collected, and the first smelting data corresponding to the factors affecting the temperature of LF refined molten steel were obtained as the dataset. The model was trained to construct an LF furnace refining molten steel temperature prediction model for predicting the temperature of LF refined molten steel. Based on the LF furnace refining steel temperature prediction model, the initial steel temperature prediction value for the current new furnace is obtained by collecting the second smelting data of the current new furnace. Based on the similarity between the second smelting data and the first smelting data, the predicted steel temperature corresponding to the first smelting data of the historical furnaces similar to the current new furnace is obtained. By comparing it with the actual steel temperature of the historical furnaces, the prediction error is obtained, the initial predicted steel temperature is corrected, and the predicted steel temperature of the current new furnace is generated. In the process of obtaining similarity, the similarity between the second smelting data and the first smelting data is measured according to Euclidean distance similarity. The similarity is expressed as: ； ； in, Represents Euclidean distance. Indicates similarity. The number of influencing factors. and These represent the data for new furnace runs and the data for historical furnace runs, respectively. One influencing factor, Indicates the first The weights of each influencing factor; In the process of obtaining similarity, the Grey Wolf optimization algorithm is used to obtain the first... Weights of each influencing factor And obtain the number of the historical furnace batches; Based on the number, the prediction errors of several historical heats are obtained. The error of the initial molten steel temperature prediction value is obtained by weighted average method. The initial molten steel temperature prediction value is corrected to generate the molten steel temperature prediction value of the current new heat. The dataset is updated based on the predicted steel temperature of the current new furnace corresponding to the second smelting data; When using the Grey Wolf optimization algorithm, the optimization process must comply with the following constraints: The sum of the weights of all factors is 1; The weight of any one factor is greater than 0; The number of similar furnace cycles ranges from 1 to 100; In the process of obtaining the predicted temperature of molten steel for the current heat cycle, the predicted temperature of molten steel for the current heat cycle is expressed as: ； in, This represents the predicted value of the molten steel temperature prediction model for the current new heat. The number of similar historical furnace batches. Indicates furnace number The weight, Indicates furnace number Errors in the prediction model; In the process of obtaining the predicted value of molten steel temperature, the weight qk of heat k is expressed as: 。 2. The error correction method for predicting molten steel temperature in an LF refining furnace according to claim 1, characterized in that: In the process of acquiring the first smelting data, based on the smelting operation flow of the LF process, the energy balance analysis of the refining process is used to obtain the first smelting data corresponding to the influencing factors of the temperature of the LF refined molten steel.

3. The error correction method for predicting molten steel temperature in an LF refining furnace according to claim 2, characterized in that: In the process of constructing the dataset, the first smelting data is preprocessed before the dataset is constructed. The preprocessing process includes: deleting blank values, deleting outliers, and data normalization.

4. The error correction method for predicting molten steel temperature in an LF refining furnace according to claim 1, characterized in that: In the process of obtaining the error of the initial predicted molten steel temperature, the error of the initial predicted molten steel temperature is expressed as: 。 5. An error correction system for predicting the temperature of molten steel in an LF refining furnace, characterized in that, include: The data acquisition module is used to collect historical smelting data of the LF refining furnace, obtain the first smelting data corresponding to the factors affecting the temperature of molten steel in the LF refining furnace as a dataset, and collect the second smelting data of the current new furnace. The temperature prediction module is used to train the backpropagation neural network model based on a single hidden layer using the first smelting data to construct an LF furnace refining molten steel temperature prediction model for predicting the temperature of LF refining molten steel. Based on the LF furnace refining molten steel temperature prediction model, the module obtains the initial molten steel temperature prediction value of the current new furnace batch using the second smelting data. The temperature correction module is used to obtain the predicted steel temperature corresponding to the first smelting data of a historical furnace similar to the current new furnace based on the similarity between the second smelting data and the first smelting data. By comparing this predicted value with the actual steel temperature of the historical furnace, the module obtains the prediction error, corrects the initial predicted steel temperature, and generates the predicted steel temperature for the current new furnace. In the process of obtaining the similarity, the similarity between the second smelting data and the first smelting data is measured using Euclidean distance similarity, where the similarity is expressed as: ； ； in, Represents Euclidean distance. Indicates similarity. The number of influencing factors. and These represent the data for new furnace runs and the data for historical furnace runs, respectively. One influencing factor, Indicates the first The weights of each influencing factor; In the process of obtaining similarity, the Grey Wolf optimization algorithm is used to obtain the first... Weights of each influencing factor And obtain the number of the historical furnace batches; When using the Grey Wolf optimization algorithm, the optimization process must comply with the following constraints: The sum of the weights of all factors is 1; The weight of any one factor is greater than 0; The number of similar furnace cycles ranges from 1 to 100; Based on the number, the prediction errors of several historical heats are obtained. The error of the initial molten steel temperature prediction value is obtained by weighted average method. The initial molten steel temperature prediction value is corrected to generate the molten steel temperature prediction value of the current new heat. In the process of obtaining the predicted temperature of molten steel for the current heat cycle, the predicted temperature of molten steel for the current heat cycle is expressed as: ； in, This represents the predicted value of the molten steel temperature prediction model for the current new heat. The number of similar historical furnace batches. Indicates furnace number The weight, Indicates furnace number Errors in the prediction model; In the process of obtaining the predicted value of molten steel temperature, the weight qk of heat k is expressed as: ； The dataset update module is used to update the dataset based on the predicted steel temperature of the current new furnace corresponding to the second smelting data.

Images