Component prediction method after molten iron pretreatment based on adaptive multi-output BP neural network

Abstract

The invention relates to a component prediction method after molten iron pretreatment based on an adaptive multi-output BP neural network, and the method comprises the steps: employing a data normalization processing and Pearson's correlation coefficient to carry out the processing of data, and determining the input parameters of a substituting model; a self-adaptive multi-output BP neural network forecasting model is established, and the Si content, the P content and the S content after molten iron pretreatment are forecasted at the same time; and the parameters of the self-adaptive multi-output BP neural network forecast model are obtained through simulation. According to the method, the complex reaction and the nonlinear relation in the data can be automatically learned, and simple assumptions do not need to be made on part of parameters of a mechanism model; the data is preprocessed and processed through data normalization processing and Pearson's correlation coefficient analysis, the robustness of the data is effectively improved, and data deviation and noise interference are reduced.

Description

Technical Field

[0001] This invention relates to a method for predicting the composition of molten iron after pretreatment based on an adaptive multi-output BP neural network, belonging to the field of automation technology in iron and steel smelting. Background Technology

[0002] Hot metal pretreatment is a key process for improving steel quality and shortening converter smelting time; therefore, research on hot metal pretreatment composition prediction has significant practical implications. For example, invention application CN202410180996.2 discloses a method for predicting silicon content in molten hot metal fed into a converter and controlling its charging. This technology analyzes the silicon content of different types of hot metal fed into the furnace, studies the variation law of flue gas composition during converter smelting, and establishes a hot metal silicon content prediction model, thereby achieving precise operation of converter smelting and improving converter dephosphorization effect and temperature control accuracy. Invention application CN201611200107.6 discloses a method and system for predicting blast furnace hot metal quality. This technology studies the nonlinear correlation factors affecting hot metal quality through the maximum information coefficient, selects input parameters based on the correlation magnitude, and uses a dynamic neural network method to train the prediction model, achieving rolling optimization of the model and thus predicting hot metal quality.

[0003] A careful analysis of the existing technologies revealed that none of them addressed the prediction of the composition of molten iron after converter pretreatment. Further research into traditional techniques for predicting the composition of molten iron after pretreatment primarily relies on manual experience and mechanistic models. Manual experience-based judgment, derived from historical data on blast furnace molten iron composition and desulfurization, is heavily influenced by subjectivity, resulting in significant errors. Mechanistic models, based on metallurgical physicochemical reaction mechanisms, derive predictions through material and heat balance. However, these models require numerous assumptions, limiting their accuracy and failing to meet production demands when significant process changes occur.

[0004] It is clear that existing methods for predicting the composition of molten iron after pretreatment are lacking in terms of data quality, algorithm selection, and parameter optimization, resulting in persistent errors. In order to more accurately control the final composition of molten iron after pretreatment and improve the efficiency of the steelmaking process, it is necessary to propose a new method for predicting the composition of molten iron after pretreatment, which is a prediction model based on neural networks. Summary of the Invention

[0005] This invention provides a method for predicting the composition of molten iron after pretreatment based on an adaptive multi-output BP neural network. This method can automatically learn complex reactions and nonlinear relationships in the data without making simple assumptions about some parameters of the mechanism model. By preprocessing and processing the data through data normalization and Pearson correlation coefficient analysis, the robustness of the data is effectively increased and the data bias and noise interference are reduced.

[0006] The technical solution adopted by this invention to solve its technical problem is:

[0007] A method for predicting the composition of molten iron after pretreatment based on an adaptive multi-output BP neural network, specifically including the following steps:

[0008] Step S1, data acquisition: process parameters are acquired from the production system. The process parameters include blast furnace hot metal condition data, desulfurizing agent data, equipment process parameters of the KR desulfurization station, and composition data of the hot metal after pretreatment.

[0009] Step S2: Data preprocessing, removing outliers and blank values ​​from the process parameters, and normalizing the filtered data;

[0010] Step S3: Pearson correlation coefficient analysis is used on the normalized process parameters to determine the correlation between different process parameters and the desulfurization composition of molten iron, and several process parameters with the highest correlation with the output parameters are selected as the input parameters of the model.

[0011] Step S4: Construct an adaptive BP neural network composition prediction model for molten iron after pretreatment, wherein the output parameters include the Si content, P content and S content of the molten iron after pretreatment.

[0012] Step S5: Substitute the input parameters determined in step S3 into the composition prediction model in step S4, continuously iterate the prediction results, optimize and adjust the output parameters of the composition prediction model, and obtain the optimal adaptive BP neural network composition prediction model after molten iron preprocessing; wherein, optimizing and adjusting the output parameters is achieved by adjusting the learning rate, the number of hidden layer nodes, and the number of iterations of the composition prediction model.

[0013] Step S6: Install the composition prediction model obtained in step S5 into the main control room of the KR desulfurization station, and input the production process parameters generated in the actual production process into the composition prediction model in real time to obtain the optimal prediction results of Si content, P content and S content after desulfurization of molten iron.

[0014] Step S7: Determine whether the preset desulfurization effect has been achieved based on the prediction result obtained in step S6. If the preset desulfurization effect has been achieved, the molten iron is tapped out of the furnace.

[0015] Furthermore, in step S1, the blast furnace molten iron condition data includes molten iron weight, Si content, P content, and S content.

[0016] The equipment process parameters of the KR desulfurization station include the lifespan of the agitator head, the lifespan of the slag removal head, the agitation time, the insertion depth of the agitator head, the desulfurization treatment time, and the total desulfurization treatment time.

[0017] The composition data of molten iron after pretreatment includes the Si content, P content, and S content in the molten iron.

[0018] Furthermore, the data preprocessing steps in step S2 are as follows:

[0019] Step S21: Remove blank values ​​from the process parameters;

[0020] Step S22: Check whether the process parameters for removing blank values ​​follow a normal distribution. If they do, proceed directly to step S23. If they do not, they are identified as outliers and removed before proceeding to step S23.

[0021] Step S23: Based on the actual working conditions and the preset process parameters production range, data whose fluctuation range exceeds the preset process parameters production range are identified as outliers and removed.

[0022] Step S24: Normalize the filtered data;

[0023] Furthermore, in step S23, the preset process parameters production range includes two parts. One part is before the molten iron pretreatment: the Si content in the molten iron fluctuates within the range of 0.15%-0.8%, the P content fluctuates within the range of 0.08%-0.15%, and the S content fluctuates within the range of 0.0005%-0.046%.

[0024] The second part is the molten iron after pretreatment: the Si content in the molten iron fluctuates between 0.1% and 0.8%, the P content fluctuates between 0.08% and 0.15%, and the S content fluctuates between 0.0005% and 0.0015%.

[0025] Furthermore, in step S24, the formula for normalizing the filtered data is:

[0026] y=((y max -y min )×(x i -x min )) / (x max -x min )+y min (1)

[0027] In formula (1), y represents the normalized data corresponding to the original process parameters, and x represents the normalized data. i For the original process parameters, y max y represents the maximum value after normalization of the original process parameters. min x is the minimum value after normalization of the original process parameters. max x represents the maximum value of the original process parameters. min This represents the minimum value of the original process parameters.

[0028] Furthermore, the formula for calculating the Pearson correlation coefficient in step S3 is as follows:

[0029]

[0030] In formula (2), r is the Pearson correlation coefficient, and r∈[-1,1], x and y are different process parameters, x i Here are the original process parameters, and n is the sample size. This represents the average value of different process parameters;

[0031] Furthermore, in step S4, the constructed adaptive BP neural network composition prediction model after molten iron pretreatment includes an input layer, a hidden layer, and an output layer. The composition prediction model iteratively outputs the difference between adjacent results through the BP neural network, which is defined as ΔE.

[0032] Furthermore, in step S5, the number of nodes in the hidden layer is the optimal number of hidden layers obtained through multiple simulations, and the formula for obtaining it is:

[0033]

[0034] In formula (3), h is the number of hidden layer nodes, a is a constant and a∈[0,10], m is the total number of input parameters in the input layer, and n is the total number of output parameters in the output layer;

[0035] The learning rate of the component prediction model is automatically adjusted as the error changes. The adjustment method is as follows:

[0036]

[0037] In formula (4), η is the learning rate, η0 is the initial learning rate of the component prediction model, a1 and a2 are constants, Y is the number of iterations, ΔE=E(Y)-E(Y-1), E=∑ n E n , where n is the number of samples.

[0038] By employing the above technical solutions, the present invention has the following beneficial effects compared to the prior art:

[0039] 1. The composition prediction method for molten iron after pretreatment based on adaptive multi-output BP neural network provided by the present invention, compared with the traditional single-model BP neural network, can predict multiple output parameters at the same time, without the need to establish multiple single-output prediction models, without the need to make simple assumptions about some parameters of the mechanism model, and the adaptive adjustment of the learning rate can avoid the prediction results from getting trapped in local optima, thus improving the prediction accuracy of the composition of molten iron after pretreatment.

[0040] 2. The composition prediction method for molten iron after pretreatment based on adaptive multi-output BP neural network provided by this invention uses data normalization to preprocess the data and Pearson correlation coefficient analysis to determine the correlation between other parameters and process parameters after molten iron pretreatment from the perspective of data. This allows for more intuitive screening of process parameters, effectively increasing the robustness of the data and reducing data deviation and noise interference. Attached Figure Description

[0041] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0042] Figure 1 This is a schematic diagram illustrating the correlation between process parameters and endpoint carbon content obtained by importing data into Pearson correlation coefficient analysis in a preferred embodiment of the present invention.

[0043] Figure 2 This is a structural diagram of the component prediction model constructed in a preferred embodiment of the present invention. Detailed Implementation

[0044] The present invention will now be described in further detail with reference to the accompanying drawings. The specific dimensions used in this embodiment are merely illustrative and do not limit the scope of protection of the present invention.

[0045] Traditional methods for predicting the composition of molten iron after pretreatment rely on models built upon mechanistic equilibrium. These models involve simplifying assumptions about complex parameters, resulting in cumbersome model construction and solution processes. The entire prediction process requires extensive calculations and adjustments. Furthermore, issues such as the difficulty in quantifying crucial parameters like stirring parameters in molten iron desulfurization stations and the lack of information on blast furnace molten iron temperature contribute to the inaccuracy of these mechanistic models compared to workshop requirements. Moreover, developing and maintaining such mechanistic models demands deep professional knowledge and extensive practical experience from researchers, making them unsuitable for later maintenance when significant process changes occur.

[0046] Considering that BP neural networks are good at handling complex nonlinear relationships and can fit complex functional relationships more accurately, this application is based on adaptive multi-output BP neural networks for modeling, replacing traditional mechanism-driven models with deep learning-driven models. Through data normalization and Pearson correlation coefficient analysis, the input data is optimized. At the same time, this application establishes a unified model framework by sharing features, adaptively adjusts the model's learning rate, and predicts multiple output parameters, effectively avoiding the repeated development costs and structural complexity of single-output models in multi-task scenarios.

[0047] The following section elaborates on the composition prediction method for molten iron pretreatment based on an adaptive multi-output BP neural network provided in this application, including the following steps:

[0048] Step S1, data acquisition, involves collecting process parameters from the production system. These parameters include blast furnace hot metal condition data, desulfurizing agent data, equipment and process parameters of the KR desulfurization station, and composition data of the pretreated hot metal. The key to this step is data acquisition. To ensure the comprehensiveness and relevance of the data, providing a high-quality raw data foundation for subsequent modeling, data is directly collected from the production system, covering the original initial state, processing parameters, and pretreated result parameters. This ensures that the data reflects the impact of each stage on the final composition and avoids model prediction bias due to missing key parameters.

[0049] Preferably, the blast furnace molten iron condition data includes molten iron weight, Si content, P content, and S content; the KR desulfurization station equipment process parameters include stirring head lifespan, slag removal head lifespan, stirring time, stirring head insertion depth, desulfurization treatment time, and total desulfurization treatment time; the molten iron pretreatment composition data includes Si content, P content, and S content in the molten iron. There are various methods to obtain the above data, including pressure sensors, weighing sensors, stirring head sensors, PLC data, and database composition information.

[0050] After obtaining the relevant data, it is necessary to ensure the rigor and standardization of the data in order to improve data quality and reduce model interference. Therefore, step S2 is to perform data preprocessing, remove outliers and blank values ​​in the process parameters, and normalize the filtered data.

[0051] Preferably, the data preprocessing steps in step S2 are as follows: Step S21, remove blank values ​​from the process parameters; Step S22, check whether the process parameters for which blank values ​​have been removed follow a normal distribution. If they do, proceed directly to step S23; otherwise, they are identified as outliers and removed before proceeding to step S23; Step S23, based on the actual operating conditions and the preset production range of the process parameters, identify data whose fluctuation range exceeds the preset production range of the process parameters as outliers and remove them; Step S24, normalize the filtered data.

[0052] After removing blank values, normal distribution verification was performed. Then, outliers exceeding the range were removed by combining the working conditions. This three-level screening eliminated extreme abnormal working conditions in production and improved data purity.

[0053] In step S23, the preset process parameter production range includes two sets of data, representing the values ​​before and after pretreatment. Before pretreatment, the Si content in the molten iron fluctuates between 0.15% and 0.8%, the P content between 0.08% and 0.15%, and the S content between 0.0005% and 0.046%. After pretreatment, the Si content in the molten iron fluctuates between 0.1% and 0.8%, the P content between 0.08% and 0.15%, and the S content between 0.0005% and 0.0015%. Both sets of data meet the actual process requirements of steel production. By double-screening the data before and after pretreatment, the accuracy of the data is effectively improved. Because different process parameters have vastly different dimensions, directly inputting these parameters into the model would amplify the influence of parameters with large dimensions. Therefore, in step S24, the filtered data is normalized to map all data to a uniform range. Preferably, the data is distributed between [-1, 1] to reduce errors caused by inconsistent data dimensions. The specific formula for this process is:

[0054] y=((y max -y min )×(x i -x min )) / (x max -x min )+y min (1)

[0055] In formula (1), y represents the normalized data corresponding to the original process parameters, and x represents the normalized data. i For the original process parameters, y max y represents the maximum value after normalization of the original process parameters. min x is the minimum value after normalization of the original process parameters. max x represents the maximum value of the original process parameters. min This represents the minimum value of the original process parameters.

[0056] After initial data processing, the input parameters are further reduced in dimensionality and optimized to improve model training efficiency and prediction accuracy. Specifically, in step S3, Pearson correlation coefficient analysis is used on the normalized process parameters to determine the correlation between different process parameters and the desulfurization composition of molten iron. Several process parameters with the highest correlation to the output parameters are then selected as the model's input parameters.

[0057] The formula for calculating the Pearson correlation coefficient is as follows:

[0058]

[0059] In formula (2), r is the Pearson correlation coefficient, and r∈[-1,1], x and y are different process parameters, and n is the sample size. This represents the average value of different process parameters.

[0060] When choosing to use Pearson correlation coefficient for assessment, the applicant also considered the random forest assessment method. Although the correlation of parameters obtained by these two assessment methods cannot be directly compared, the ranking of feature correlation can be used as a reference. Here, the metallurgical mechanisms such as heat balance and material balance in the KR desulfurization pretreatment process of molten iron are considered. The weight of molten iron is a key factor affecting the desulfurization reaction process, material consumption, and energy balance. The following parameters were selected as input parameters: molten iron weight, stirring head life, slag removal head life, stirring time, stirring head insertion depth, desulfurizing agent addition amount, slag removal time, desulfurization treatment time, total desulfurization cycle time, Si content, P content, and S content of molten iron. Table 1 is a table of the correlation importance of random forest features.

[0061] Table 1. Correlation of the Importance of Random Forest Features

[0062] DES_Si DES_P DES_S weight of molten iron 0.2615 1.2728 1.4343 Stirring head lifespan 0.7271 0.6601 1.4419 Slag head lifespan 0.5475 1.4502 1.2266 Stirring time 0.7786 0.9854 2.416 Stirring head insertion depth 0.9452 1.3452 0.8788 Desulfurizer addition amount 0.7813 1.1222 1.651 Time to dig up the scum 0.3039 1.214 1.5204 Desulfurization treatment time 0.7333 0.9426 1.5733 Full desulfurization cycle time 0.6657 0.8524 1.4556 Si content in molten iron 0.9619 1.9703 1.2996 P content in molten iron 1.3557 5.865 0.6717 S content in molten iron 3.9215 1.5993 0.8962

[0063] Figure 1 The image shows the Pearson correlation coefficient values. Clearly, the Pearson correlation coefficient results are more visually appealing, while the random forest modeling process is complex. Therefore, the Pearson correlation coefficient was chosen as the method for filtering feature correlation.

[0064] Step S4: Construct an adaptive BP neural network model for predicting the composition of molten iron after pretreatment, including an input layer, hidden layers, and an output layer. Figure 2 As shown, in the preferred embodiment, the number of input layer neurons is 9, the number of hidden layer neurons is 12, and the number of output layer neurons is 3, including the Si content, P content, and S content after molten iron pretreatment. The component prediction model iteratively outputs the difference between adjacent results through a BP neural network, which is defined as ΔE.

[0065] Step S5: Substitute the input parameters determined in step S3 into the composition prediction model in step S4, continuously iterate the prediction results, optimize and adjust the output parameters of the composition prediction model, and obtain the optimal adaptive BP neural network composition prediction model after molten iron preprocessing. In this step, the prediction results of Si content, P content and S content are obtained by continuously optimizing and adjusting parameters such as the learning rate of the composition prediction model, the number of hidden layer nodes, the initial weight of the BP model and the number of iterations.

[0066] The number of nodes in the hidden layer is the optimal number of hidden layers obtained through multiple simulations, and the formula for obtaining it is:

[0067]

[0068] In formula (3), h is the number of hidden layer nodes, a is a constant and a∈[0,10], m is the total number of input parameters in the input layer, and n is the total number of output parameters in the output layer;

[0069] The learning rate of the component prediction model is automatically adjusted as the error changes. The adjustment method is as follows:

[0070]

[0071] In formula (4), η is the learning rate, η0 is the initial learning rate of the component prediction model, a1 and a2 are constants, Y is the number of iterations, ΔE=E(Y)-E(Y-1), E=∑ n E n , where n is the number of samples.

[0072] When the overall error ΔE < 0, the convergence speed is accelerated and the learning rate is increased; when the overall error ΔE > 0, the convergence speed is reduced and the learning rate is decreased.

[0073] Step S6: Install the composition prediction model obtained in step S5 into the main control room of the KR desulfurization station, and input the production process parameters generated in the actual production process into the composition prediction model in real time to obtain the optimal prediction results of Si content, P content and S content after desulfurization of molten iron.

[0074] Step S7: Determine whether the preset desulfurization effect has been achieved based on the prediction result obtained in step S6. If the preset desulfurization effect has been achieved, the molten iron is tapped out of the furnace.

[0075] In summary, the key aspects of the composition prediction method for molten iron after pretreatment based on an adaptive multi-output BP neural network provided in this application include data normalization, Pearson correlation coefficient analysis, and other data processing methods to determine the input parameters for the model. To reduce the complexity of the prediction model, an adaptive multi-output BP neural network prediction model is established to simultaneously predict the Si, P, and S contents after molten iron pretreatment. The parameters of the adaptive multi-output BP neural network prediction model are obtained through simulation. This method can automatically learn complex reactions and nonlinear relationships in the data without making simple assumptions about some parameters of the mechanistic model, effectively increasing the robustness of the data and reducing data bias and noise interference.

[0076] Finally, this application provides a specific embodiment to verify the feasibility of the above prediction method.

[0077] Example:

[0078] The first step is data collection: Data from 7,000 heats in the converter workshop of a certain plant in 2025 was collected, including: heat number, molten iron weight, molten iron composition (Si content, P content, S content), stirring head life, slag removal head life, stirring head insertion depth, stirring time, slag removal time, desulfurizing agent dosage, desulfurization cycle time, and molten iron composition after pretreatment (Si content, P content, S content).

[0079] The second step is data preprocessing: removing blank values ​​from the process parameters; and removing outliers by testing for outliers using a normal distribution. After data preprocessing, 2770 heats of usable data remain.

[0080] The third step is Pearson correlation coefficient analysis: Figure 1 As shown, the data was imported into Pearson correlation coefficient analysis to obtain the correlation between process parameters and the endpoint carbon content (specifically, process parameters with absolute values ​​higher than 0.1). The data was then normalized to distribute it between [-1, 1], reducing errors caused by inconsistent data dimensions.

[0081] Step 4, Construction Figure 2 The adaptive BP neural network model for predicting the composition of molten iron after pretreatment is shown. It selects nine features as input parameters, namely molten iron weight, slag removal head life, stirring time, desulfurizing agent addition amount, slag removal time, desulfurization treatment time, molten iron Si content, P content, and S content as input parameters, and outputs Si content, P content, and S content.

[0082] The fifth step involves using the collected historical data from the desulfurization station as a training set, tracking the actual production of 100 furnaces, and inputting the process data of these 100 furnaces into an adaptive multi-output BP neural network prediction model. The optimal model parameters are obtained by continuously optimizing the learning rate, learning step size, and error parameters.

[0083] Step 6: After substituting the data into the adaptive BP neural network model for predicting the composition of molten iron after pretreatment, the prediction accuracy for P content within ±0.005% was 72%, and within ±0.008% was 90%. The prediction accuracy for Si content within ±0.05% was 73%, and within ±0.09% was 96%. The prediction accuracy for S content within ±0.0001% was 63%, and within ±0.0003% was 96%.

[0084] The data collected in the embodiment, after being further substituted into the wavelet neural network prediction model, yielded the following results: The prediction accuracy for P content within ±0.005% after molten iron pretreatment was 60%, and the prediction accuracy for within ±0.008% was 80%. The prediction accuracy for Si content within ±0.05% after molten iron pretreatment was 50%, and the prediction accuracy for within ±0.1% was 90%. The prediction accuracy for S content within ±0.0001% after molten iron pretreatment was 46%, and the prediction accuracy for within ±0.0003% was 87%.

[0085] Clearly, the prediction results obtained by substituting the adaptive multi-output BP neural network-based hot metal composition prediction method provided in this application are all more accurate than the results output by the wavelet neural network prediction model. This application is more in line with expectations and solves the problem of low composition prediction accuracy after KR desulfurization of hot metal.

[0086] Those skilled in the art will understand that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains. It should also be understood that terms such as those defined in general dictionaries should be understood to have the meaning consistent with their meaning in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless defined as herein.

[0087] The meaning of "and / or" as used in this application includes situations where each exists alone or both exist simultaneously.

[0088] The term "connection" as used in this application can mean a direct connection between components or an indirect connection between components through other components.

[0089] Based on the above-described preferred embodiments of the present invention, and through the foregoing description, those skilled in the art can make various changes and modifications without departing from the inventive concept. The technical scope of this invention is not limited to the contents of the specification, but must be determined according to the scope of the claims.

Claims

1. A method for predicting the composition of molten iron after pretreatment based on an adaptive multi-output BP neural network, characterized by: Specifically comprising the following steps: Step S1, data acquisition, collecting process parameters from the production system, the process parameters including blast furnace hot metal condition data, desulfurizer data, KR desulfurization station equipment process parameters and hot metal pretreatment composition data; Step S2, data preprocessing, eliminating abnormal values and blank values in the process parameters, and normalizing the screened data; Step S3, using Pearson correlation coefficient analysis on the normalized process parameters to determine the correlation size of different process parameters on hot metal desulfurization composition, and selecting a number of process parameters with the highest correlation with output parameters as input parameters of the model; Step S4, constructing an adaptive BP neural network hot metal pretreatment composition prediction model, wherein the output parameters include the Si content, P content and S content of the hot metal after pretreatment; Step S5, substituting the input parameters determined in step S3 into the composition prediction model of step S4, continuously iterating the prediction results, optimizing and adjusting the output parameters of the composition prediction model, and obtaining the optimal adaptive BP neural network hot metal pretreatment composition prediction model; wherein the optimization and adjustment of the output parameters are achieved by adjusting the learning rate, the number of hidden layer nodes and the number of iterations of the composition prediction model; Step S6, installing the composition prediction model obtained in step S5 into the main control room of the KR desulfurization station, inputting the production process parameters generated in the actual production process into the composition prediction model in real time, and obtaining the prediction results of the optimal Si content, P content and S content of the hot metal after desulfurization; Step S7, determining whether the preset desulfurization effect is achieved according to the prediction results obtained in step S6, and if so, tapping the hot metal.

2. The method according to claim 1, wherein the method is based on an adaptive multi-output BP neural network for predicting the composition of the molten iron after pretreatment. In step S1, the blast furnace hot metal condition data includes hot metal weight, Si content, P content and S content; The KR desulfurization station equipment process parameters include stirring head life, slagging head life, stirring time, stirring head insertion depth, desulfurization treatment time and total desulfurization treatment time; The hot metal pretreatment composition data includes Si content, P content and S content in the hot metal.

3. The method according to claim 1, wherein the method is characterized by: In step S2, the data preprocessing steps are as follows: Step S21, eliminating blank values in the process parameters; Step S22, checking whether the process parameters after eliminating blank values conform to normal distribution, if so, directly entering step S23, if not, determining as abnormal values and entering step S23 after elimination; Step S23, combining the actual working conditions to preset the production range of the process parameters, and determining the data with fluctuation range exceeding the preset production range of the process parameters as abnormal values for elimination; Step S24, normalizing the screened data.

4. The hot metal pretreatment composition prediction method based on adaptive multi-output BP neural network according to claim 3, wherein: In step S23, the preset production range of the process parameters includes two parts, one part before hot metal pretreatment: the Si content fluctuation range in the hot metal is 0.15%-0.8%, the P content fluctuation range is 0.08%-0.15%, and the S content fluctuation range is 0.0005%-0.046%; The second part is the pretreatment of molten iron: the Si content in the molten iron fluctuates in the range of 0.1%-0.8%, the P content fluctuates in the range of 0.08%-0.15%, and the S content fluctuates in the range of 0.0005%-0.0015%.

5. The method according to claim 3, wherein the method is characterized by: In step S24, the formula for normalizing the data after screening is: y = ((y max -y min ) × (x i -x min )) / (x max -x min ) + y min (1) In formula (1), y is normalized data corresponding to the original process parameter, x i is the original process parameter, y max is the maximum value of the normalized original process parameter, y min is the minimum value of the normalized original process parameter, x max is the maximum value of the original process parameter, x min is the minimum value of the data of the original process parameter.

6. The method of claim 1, wherein the method is based on an adaptive multi-output BP neural network for predicting the composition of the molten iron after the pretreatment. In step S3, the calculation formula of Pearson correlation coefficient analysis is: In formula (2), r is a Pearson correlation coefficient, and r ∈ [-1, 1], x and y are different process parameters, x i is an original process parameter, n is a sample number, is an average value of different process parameters.

7. The method according to claim 1, wherein the method is characterized by: In step S4, the component prediction model of the adaptive BP neural network after the pretreatment of molten iron comprises an input layer, a hidden layer and an output layer, and the difference between adjacent results output by the BP neural network is defined as ΔE.

8. The method of claim 1, wherein the method is based on an adaptive multi-output BP neural network for predicting the composition of the molten iron after the pretreatment. In step S5, the number of nodes in the hidden layer is the optimal number of hidden layers obtained through multiple simulation results, and the acquisition formula is: In formula (3), h is the number of hidden layer nodes, a is a constant, and a [0, 10], m is the total number of input parameters in the input layer, and n is the total number of output parameters in the output layer; The learning rate of the component prediction model is automatically adjusted according to the error change, and the adjustment method is: In formula (4), η is a learning rate, η0is an initial learning rate of the component prediction model, a1and a2are both constants, Y is the number of iterations, ΔE=E(Y)-E(Y-1), E=∑ n E n , and n is the number of samples.

Images