Method for predicting sintered ore quality and method for predicting sintered ore yield

Abstract

To predict quality and yield of sintered ore accurately more than ever.SOLUTION: A quality prediction method of sintered ore is provided, the method comprises: a variable selecting step for selecting input variables in which a coefficient imparted to each of input variables has a predetermined value or more by performing Lasso regression with at least the blending specifications for each raw material brand during sinter production, firing specifications, heat source blending ratio, quicklime blending ratio, and component values of sintered ore as input variables, and with the quality value of sintered ore as an output variable; a statistical model learning step for learning a statistical model with the variables selected in the variable selecting step as input variables and the quality value of the sintered ore as an output variable; and a quality prediction step for predicting the quality value of the sintered ore from the prediction target data of the variables selected in the variable selecting step using the learned statistical model.SELECTED DRAWING: Figure 2

Description

[Technical Field]

[0001] The present invention relates to a method for predicting the quality of sintered ore used as a raw material for a blast furnace, and a method for predicting the yield during sintered ore production. [Background technology]

[0002] Sintered ore is widely used as a raw material for blast furnaces. For stable blast furnace operation, it is important that sintered ore meets the required blast furnace feedstock quality (e.g., strength, reduction disintegration, reducibility), while maintaining a specified yield level to reduce production costs. For example, sintered ore disintegrates during reduction at the top of the blast furnace (low temperatures of 400–600°C). Severe disintegration impairs the permeability of the blast furnace, ultimately resulting in poor furnace performance and poor molten iron quality. Proper control of reduction disintegration is essential for stable blast furnace operation. Therefore, in the sintered ore production process, the quality and yield of sintered ore are estimated in advance, and operational actions are taken based on these estimates to appropriately control quality and yield. Accurate prediction of sintered ore quality and yield is important for quality control and production cost management.

[0003] As for sinter quality prediction methods, there are methods for predicting the Reduction Disintegration Index (RDI) from FeO in the sinter ore, and methods for predicting it from the Al2O3 / SiO2 ratio of the fine powder in the sinter raw material. For example, Patent Document 1 discloses a technique in which the RDI is predicted from the slag content and iron ore blending ratio using the slag component content of the sinter ore (SiO2% + CaO% + MgO% - Al2O3%) and an RDI set for each iron ore in the sinter raw material, and the iron ore blending ratio is adjusted based on the difference between the predicted RDI and the set RDI. Patent Document 2 discloses a technique in which image analysis of sinter ore samples for RDI measurement is performed to identify each structure and predict the RDI from the number of cracks exceeding the critical crack length that leads to structure destruction. Patent Document 3 discloses a technique for predicting the RDI based on the air flow rate through the sinter machine bed.

[0004] Regarding the strength of sintered ore, for example, Patent Document 4 discloses a technique for calculating the matrix strength of sintered ore from the calcium ferrite content and slag content of the sintered ore, calculating the tensile strength of sintered ore from the matrix strength and porosity of the sintered ore, and estimating the shatter strength index (SI) from the tensile strength. Regarding the reducibility of sintered ore, for example, Patent Document 5 discloses a technique for measuring the particle density and bulk density of sintered ore particles, quantifying the particle shape from these values, and predicting reducibility.

[0005] Regarding the yield of sintered ore, for example, Patent Document 6 discloses a technique for estimating the yield of sintered ore from the crushing energy when crushing the sintered ore. [Prior art documents] [Patent documents]

[0006] [Patent Document 1] Japanese Patent Application Publication No. 01-176041 [Patent Document 2] Special Publication No. 01-23734 [Patent Document 3] Special Publication No. 58-38496 [Patent Document 4] Japanese Patent Application Publication No. 07-11349 [Patent Document 5] Japanese Unexamined Patent Publication No. 61-217534 [Patent Document 6] Special Publication No. 06-102811 Summary of the Invention [Problem to be solved by the invention]

[0007] However, the quality and yield of sintered ore vary not only depending on the values ​​of sintered ore components such as FeO, the type of raw material used, the porosity and slag content of the sintered ore, but also on the specific conditions of each sintering machine, equipment renewal conditions, firing conditions, etc. For this reason, there remain issues with the accuracy of predicting sintered ore quality and yield using conventional methods.

[0008] Therefore, the present invention has been made in consideration of the above problems, and an object of the present invention is to provide a method for predicting the quality and yield of sintered ore, which are capable of predicting the quality and yield of sintered ore with greater accuracy. [Means for solving the problem]

[0009] In order to solve the above-mentioned problems, according to one aspect of the present invention, there is provided a method for predicting the quality of sintered ore, which includes a variable selection step of performing Lasso regression using at least the blending specifications for each raw material brand during sintered ore production, the firing specifications, the heat source blending ratio, the quicklime blending ratio, and the component values ​​of the sintered ore as input variables, and the quality value of the sintered ore as an output variable, and selecting input variables having coefficients equal to or greater than a predetermined value; a statistical model learning step of training a statistical model using the variables selected in the variable selection step as input variables and the quality value of the sintered ore as an output variable; and a quality prediction step of predicting the quality value of the sintered ore from the prediction target data of the variables selected in the variable selection step using the trained statistical model.

[0010] The method for predicting the quality of sintered ore includes a variable extraction step of extracting variables that are estimated to affect the quality of sintered ore based on past performance from among the variables not selected in the variable selection step, and in the statistical model learning step, a statistical model may be trained using the variables selected in the variable selection step and the variables extracted in the variable extraction step as input variables.

[0011] In addition, in order to solve the above-mentioned problems, according to another aspect of the present invention, there is provided a method for predicting the yield of sintered ore, including: a variable selection step of performing Lasso regression using at least the blending specifications, firing specifications, heat source blending ratio, quicklime blending ratio, and measured component values ​​of sintered ore for each raw material brand during sintered ore production as input variables, and the measured yield value during sintered ore production as output variables, and selecting input variables with constants assigned to each input variable that are equal to or greater than a predetermined value; a statistical model training step of training a statistical model using the variables selected in the variable selection step as input variables and the measured yield value during sintered ore production as output variables; and a yield prediction step of predicting the yield of sintered ore from the prediction target data of the variables selected in the variable selection step using the trained statistical model.

[0012] The sintered ore yield prediction method includes a variable extraction step of extracting variables that are estimated to affect the sintered ore yield based on past performance from among the variables not selected in the variable selection step, and in the statistical model training step, a statistical model may be trained using the variables selected in the variable selection step and the variables extracted in the variable extraction step as input variables.

[0013] The statistical model may be a Ridge regression model. [Effects of the Invention]

[0014] As described above, according to the present invention, it is possible to predict the quality and yield of sintered ore with higher accuracy. [Brief explanation of the drawings]

[0015] [Figure 1] 1 is a block diagram showing an example of the configuration of a prediction device according to a first embodiment of the present invention. [Figure 2] 3 is a flowchart showing a method for predicting the quality of sintered ore according to the embodiment. [Figure 3] FIG. 10 is an explanatory diagram illustrating an example of a data set. [Figure 4] 5 is a flowchart showing a method for predicting a yield of sintered ore according to a second embodiment of the present invention. [Figure 5] 10 is a table showing the absolute values ​​of the coefficients of the variables obtained by Lasso regression and the variable selection determination results in Examples A1 and A2 for Example A. [Figure 6] 10 is a graph showing the relationship between predicted and measured RDI values ​​for Example A. [Figure 7] 10 is a graph showing the relationship between predicted and measured values ​​of TI for Example B. [Figure 8] 10 is a graph showing the relationship between predicted and measured yield values ​​for Example C. DETAILED DESCRIPTION OF THE INVENTION

[0016] Hereinafter, preferred embodiments of the present invention will be described in detail with reference to the accompanying drawings. In this specification and drawings, components having substantially the same functional configurations are designated by the same reference numerals, and redundant explanations will be omitted.

[0017] In the present invention, Lasso regression is used to select only important variables from a large number of variables that affect the quality and yield of sintered ore, and a statistical model for predicting the quality and yield of sintered ore is constructed using the variables selected by Lasso regression. This allows for efficient selection of important variables from a large number of variables, such as sintered ore component values, raw material type, sintered ore porosity and slag content, conditions specific to each sintering machine, equipment renewal conditions, and firing conditions, to construct a statistical model with high prediction accuracy. Below, a method for predicting sintered ore quality and a method for predicting sintered ore yield will be described in detail as one embodiment of the present invention.

[0018] [1. First embodiment] As a first embodiment of the present invention, a quality prediction method for predicting the quality of sintered ore will be described.

[0019] [1-1. Prediction device] 1 is a block diagram showing an example of the configuration of a prediction device 100 that executes the quality prediction method according to this embodiment. As shown in FIG. 1, the prediction device 100 according to this embodiment includes a variable selection unit 110, a model construction unit 120, and a prediction unit 130.

[0020] (Variable selection section) The variable selection unit 110 selects variables to be used in constructing a statistical model for predicting the quality of sintered ore from a dataset of various variables obtained during the sintered ore production process. The dataset is stored in a database (not shown) that stores various variables obtained during the sintered ore production process for each acquisition date and time. The various variables in the dataset include, for example, blending specifications for each raw material brand during sintered ore production, firing specifications, heat source blending ratio, quicklime blending ratio, measured component values ​​of the finished sintered ore, and the RDI of the finished sintered ore. Here, the blending specifications for each raw material brand refer to information representing the blending state of the raw materials used and include at least the blending ratio. The firing specifications refer to information regarding the conditions during firing in the sintering machine and include at least the temperature of the wind box arranged in the longitudinal direction of the sintering machine, the main blower suction pressure, and the main blower air volume. The heat source blending ratio refers to the blending ratio of the heat source, such as coke breeze, used in sintered ore production.

[0021] The variable selection unit 110 performs Lasso regression using variables other than the quality value of sintered ore as input variables from among the various variables included in the dataset, and the quality value of sintered ore as the output variable. For example, when predicting the RDI as the quality value of sintered ore, Lasso regression is performed using variables other than the quality value of sintered ore as input variables and the RDI as the output variable. Lasso regression is a type of multiple regression analysis that is resistant to outliers and is a method for reducing multicollinearity by eliminating explanatory variables that have low correlation with the target variable. By using Lasso regression, it is possible to identify variables that have a high influence on the quality value of sintered ore (e.g., RDI) from the many variables included in the dataset.

[0022] Before performing Lasso regression, the variable selection unit 110 performs scaling processing on each variable in the dataset. For example, as scaling processing, the variable selection unit 110 performs centering (processing in which the mean is set to 0), scaling (processing in which the standard deviation is set to 1), and outlier removal. The scaling processing is performed to improve the prediction accuracy of the statistical model, which will be described later.

[0023] The variable selection unit 110 performs Lasso regression using the dataset after scaling processing, and selects input variables whose coefficients (partial regression coefficients) assigned to each input variable are equal to or greater than a predetermined value. In Lasso regression, the coefficient of an input variable with a low correlation with the quality value of sintered ore is 0, and the greater the correlation with the quality value of sintered ore, the greater the absolute value of the coefficient of the input variable. Based on these characteristics of Lasso regression, the variable selection unit 110 selects input variables whose absolute value of the coefficient is greater than a predetermined value. The predetermined value of the coefficient used as the selection criterion is usually 0, but may be set to a value greater than 0. The larger the predetermined value of the coefficient used as the selection criterion, the more the variables selected by the variable selection unit 110 can be narrowed down.

[0024] (Model Construction Department) The model construction unit 120 constructs a statistical model that uses the variables selected by the variable selection unit 110 as input variables and outputs the quality value of sinter ore. The statistical model can be constructed using, for example, Ridge regression, multiple regression analysis, support vector machines, decision trees, random forests, etc. Among these, Ridge regression is a type of multiple regression analysis that is resistant to outliers and is a method that reduces multicollinearity by setting limits on the magnitude of the influence coefficients. By using a Ridge regression model constructed by Ridge regression as the statistical model, the slope (constant) of each variable becomes clear, and the accuracy of the constructed statistical model can be further improved.

[0025] (Prediction Department) The prediction unit 130 predicts the quality value of sintered ore using the statistical model constructed by the model construction unit 120. For example, the prediction unit 130 predicts the quality value of sintered ore produced during a period for which the quality value of sintered ore is to be predicted using data acquired in a sintered ore production process during that period as the prediction target data. The prediction unit 130 can also predict the quality value of sintered ore to be produced in the future using data based on the planned values ​​of a scheduled operation plan as the prediction target data. The prediction unit 130 uses the variables selected by the variable selection unit 110 as input variables, performs scaling processing on the prediction target data of the input variables, and then inputs the scaling processing into the statistical model. The prediction unit 130 outputs the output value of the statistical model as the predicted quality value of sintered ore.

[0026] The prediction device 100 may be configured by an information processing device such as a computer equipped with a CPU, ROM, RAM, etc. An input device 10 and an output device 20 are also connected to the prediction device 100. The input device 10 is a device, such as a keyboard or a mouse, that allows an operator who performs quality prediction of sintered ore to input information to the prediction device 100. The output device 20 is a device, such as a display device, that presents the results of processing performed by the prediction device 100 to the operator.

[0027] [1-2. Method for predicting sinter quality] Next, a method for predicting the quality of sintered ore according to this embodiment will be described with reference to Fig. 2 and Fig. 3. Fig. 2 is a flowchart showing the method for predicting the quality of sintered ore according to this embodiment. Fig. 3 is an explanatory diagram showing an example of a data set.

[0028] (S100: Dataset acquisition) First, the variable selection unit 110 acquires a dataset of various variables obtained in the sinter ore production process (S100). The various variables in the dataset include, for example, blending specifications for each raw material brand during sinter ore production, firing specifications, heat source blending ratio, quicklime blending ratio, measured component values ​​of the finished sinter ore, and RDI of the finished sinter ore. The dataset is stored for each acquisition date and time in a database (not shown) that stores various variables obtained in the sinter ore production process. For example, as shown in FIG. 3, the database stores variables such as RDI, quicklime blending ratio, blending ratio of brand A, blending ratio of brand B, coke powder blending ratio, and average value of firing temperature for each acquisition date and time. The variable selection unit 110 acquires a predetermined period or a predetermined number of datasets from the database.

[0029] (S110: Scaling process) Next, the variable selection unit 110 performs scaling processing on each of the variables in the data set acquired in step S100 (S110). By performing scaling processing, it is possible to improve the prediction accuracy of the statistical model described below. For example, as scaling processing, the variable selection unit 110 performs centering (processing in which the mean is set to 0), scaling (processing in which the standard deviation is set to 1), and outlier removal. Centering and scaling for each variable X can be performed using the following formula (1). Note that μ is the mean value of variable X in the entire data set, σ is the standard deviation of variable X in the entire data set, and X' is the variable after processing.

[0030]

number

[0031] (S120: Variable selection process) The variable selection unit 110 performs Lasso regression using the dataset after scaling processing to select variables that have a high correlation with the quality value of sintered ore (S120). In Lasso regression, variables other than the quality value of sintered ore are used as input variables from the various variables included in the dataset, and the quality value of sintered ore is used as the output variable. For example, when predicting the RDI as the quality value of sintered ore, Lasso regression is performed using variables other than the quality value of sintered ore as input variables and the RDI as the output variable. Note that the input variables include at least the blending specifications for each raw material brand during sintered ore production, firing specifications, heat source blending ratio, quicklime blending ratio, and sintered ore component values.

[0032] In Lasso regression, the coefficient of an input variable with a low correlation with the sinter ore quality value is 0, and the coefficient of the input variable increases as the correlation with the sinter ore quality value increases. Therefore, the variable selection unit 110 selects input variables with coefficients whose absolute values ​​are greater than a predetermined value as variables with a high correlation with the sinter ore quality value. The predetermined value of the coefficient used as the selection criterion is set to a value greater than or equal to 0. Lasso regression by the variable selection unit 110 makes it possible to automatically select input variables to be used in building a statistical model.

[0033] Here, in order to improve the accuracy of variable selection, the variable selection unit 110 may perform Lasso regression multiple times and select as input variables variables for which the average value of the coefficient for each input variable obtained in each Lasso regression is equal to or greater than a predetermined value. In Lasso regression, in order to improve prediction accuracy, variables are selected by setting the coefficient of a variable with a low correlation with the quality value of sintered ore to 0, but the variables for which this coefficient becomes 0 may differ depending on the initial value of the coefficient. In order to reduce the influence of the initial value of such a coefficient, Lasso regression may be performed multiple times to select variables to be used as input variables. Lasso regression may be performed, for example, approximately 100 times.

[0034] Furthermore, the input variables to be input to the Lasso regression may be aggregated, and the aggregated variables may be used as input variables. For example, the variable selection unit 110 aggregates multiple variables into a new variable from among the various variables included in the dataset after scaling processing. The input variables may be aggregated using a method that uses principal component analysis or a method that aggregates the blending specifications of multiple brands with similar properties. While reflecting the characteristics of the aggregated multiple variables in the aggregated variable, the number of variables that are candidates for selection as input variables can be reduced, allowing for more accurate calculation of the coefficients of the input variables.

[0035] It is possible that the variables not selected by the Lasso regression performed by the variable selection unit 110 include variables that are known to have a significant impact on the quality value of sintered ore. In this case, the operator performing the quality prediction can review the variables not selected by the Lasso regression, extract variables that have a significant impact on the quality value of sintered ore based on prior knowledge based on past performance, and use these as input variables along with the variables selected by the Lasso regression. For example, depending on the acquisition period of the data set, there may be variables that are known to affect the quality value of sintered ore but are not selected by the Lasso regression because their fluctuations during that acquisition period were small. Extracting such variables and adding them to the input variables can further improve the prediction accuracy of the constructed statistical model.

[0036] The worker checks, via the output device 20, the variables that were not selected as input variables in the Lasso regression performed by the variable selection unit 110 of the prediction device 100. If there are variables that should be added as input variables, the worker uses the input device 10 to instruct the prediction device 100 on the variables to be added as input variables. As a result, the variables selected by the Lasso regression and the variables extracted by the worker are adopted as input variables for the statistical model.

[0037] (S130: Statistical model construction) When the input variables are selected in step S120, the model construction unit 120 constructs a statistical model in which the selected variables are used as input variables and the quality values ​​of sintered ore are used as output variables (S130). The model construction unit 120 constructs the statistical model by learning using a data set of the input variables and the output variables as learning data, for example, using a method such as Ridge regression, multiple regression analysis, support vector machine, decision tree, or random forest.

[0038] (S140: Quality Prediction) The prediction unit 130 then predicts the quality value of the sintered ore using the statistical model constructed in step S130 (S140). The prediction unit 130 can predict the quality value of the sintered ore produced during a period for which the quality value of the sintered ore is to be predicted using data acquired in the sintered ore production process during that period, or can predict the quality value of the sintered ore to be produced in the future using data based on the planned values ​​of a scheduled operation plan. The data to be predicted can be specified by the operator using the input device 10.

[0039] The prediction unit 130 first performs scaling processing on the prediction target data of the input variables, similar to step S110, using the variables used to construct the statistical model (i.e., the variables selected in step S120) as input variables. At this time, the μ and σ for each variable used in the scaling processing are the same values ​​as those used in step S110. Then, the prediction unit 130 inputs the input variables after the scaling processing into the statistical model to obtain a quality value of the sintered ore.

[0040] Since the output value of the statistical model is a value after scaling, the prediction unit 130 converts the output value of the statistical model back to the original value. For example, when scaling of the variable X is performed according to the above formula (1), the original value can be obtained according to the following formula (2).

[0041]

number

[0042] The prediction unit 130 may output the quality value of the sintered ore obtained using the statistical model to the output device 20 and present it to the worker.

[0043] The method for predicting the quality of sintered ore according to this embodiment has been described above. According to this embodiment, only important variables are selected from a large number of variables that affect the quality of sintered ore by Lasso regression, and a statistical model for predicting the quality of sintered ore is constructed using the variables selected by Lasso regression. This makes it possible to efficiently select important variables from a large number of variables and construct a statistical model with high prediction accuracy, and to use this statistical model to predict the quality of sintered ore with high accuracy.

[0044] [2. Second Embodiment] As a second embodiment of the present invention, a yield prediction method for predicting the yield of sintered ore during sintered ore production will be described. Prediction of the yield of sintered ore during sintered ore production can also be performed in the same manner by the prediction device 100 described in the first embodiment. Since the configuration and function of the prediction device 100 are the same as those in the first embodiment, detailed description will be omitted here. Below, the sintered ore yield prediction method according to this embodiment will be described with reference to FIG. 4. FIG. 4 is a flowchart showing the sintered ore yield prediction method according to this embodiment.

[0045] (S200: Dataset acquisition) First, the variable selection unit 110 acquires a dataset of various variables obtained in the sinter ore production process (S200). The processing of step S200 may be performed in the same manner as step S100 in FIG. 2. The various variables in the dataset include the blending specifications for each raw material brand during the above-mentioned sinter ore production, firing specifications, heat source blending ratio, quicklime blending ratio, measured component values ​​of the finished sinter ore, and measured yield values ​​during sinter ore production. The variable selection unit 110 acquires a predetermined period or a predetermined number of datasets from a database.

[0046] (S210: Scaling process) Next, the variable selection unit 110 performs scaling processing on each of the variables in the data set acquired in step S200 (S210). The processing in step S210 may be performed in the same manner as step S110 in Fig. 2. By performing scaling processing, it is possible to improve the prediction accuracy of the statistical model, which will be described later.

[0047] (S220: Variable selection process) The variable selection unit 110 performs Lasso regression using the dataset after scaling processing to select variables that have a high correlation with the sintered ore yield (S220). The processing of step S220 may be performed in the same manner as step S120 in FIG. 2. In this embodiment, Lasso regression may be performed using variables other than the actual sintered ore yield value to be predicted from the various variables included in the dataset as input variables, and the actual sintered ore yield value as the output variable. Note that the input variables include at least the blending specifications for each raw material brand during sintered ore production, firing specifications, heat source blending ratio, quicklime blending ratio, and the actual component values ​​of the sintered ore.

[0048] The variable selection unit 110 selects input variables whose absolute values ​​of coefficients are greater than a predetermined value as variables that have a high correlation with the measured yield of sintered ore. The predetermined value of the coefficient used as the selection criterion is set to a value greater than or equal to 0. The input variables used to build the statistical model can be automatically selected by Lasso regression performed by the variable selection unit 110.

[0049] As in the first embodiment, Lasso regression may be performed multiple times, and variables for which the average value of the coefficients for each input variable obtained in each Lasso regression is equal to or greater than a predetermined value may be selected as input variables. The input variables to be input to the Lasso regression may be aggregated using techniques such as principal component analysis, and the aggregated input variables may be used. Furthermore, the operator who performs the quality prediction may extract variables that have a significant impact on the sinter ore yield based on conventional knowledge based on past performance from among the variables not selected by the Lasso regression, and use these as input variables along with the variables selected by the Lasso regression.

[0050] (S230: Statistical Model Building) When the input variables are selected in step S220, the model construction unit 120 constructs a statistical model in which the selected variables are used as input variables and the predicted yield value of sintered ore is used as an output variable (S230). The processing in step S230 may be performed in the same manner as in step S130 in FIG. 2.

[0051] (S240: Quality Prediction) The prediction unit 130 then predicts the sinter ore yield using the statistical model constructed in step S230 (S240). The prediction unit 130 can predict the sinter ore yield for a period for which the sinter ore yield is to be predicted using data acquired in the sinter ore production process, or can predict the sinter ore yield during future sinter ore production using data based on planned values ​​in a planned operation plan. The prediction target data can be specified by the operator using the input device 10.

[0052] The processing of step S240 may be performed in the same manner as step S140 in FIG. 2. The prediction unit 130 performs scaling processing on the prediction target data of the input variables, using the variables used to construct the statistical model (i.e., the variables selected in step S220) as input variables. The prediction unit 130 then inputs the input variables after the scaling processing into the statistical model to obtain a predicted value of the sintered ore yield. At this time, when the prediction unit 130 performs scaling processing on the variable X according to the above formula (1) in step S210, it converts the output value of the statistical model back to the original value according to the above formula (2). The prediction unit 130 may output the predicted value of the sintered ore yield obtained using the statistical model to the output device 20 and present it to an operator.

[0053] The sinter ore yield prediction method according to this embodiment has been described above. According to this embodiment, only important variables are selected from a large number of variables that affect the sinter ore yield by Lasso regression, and a statistical model for predicting the sinter ore yield is constructed using the variables selected by Lasso regression. This makes it possible to efficiently select important variables from a large number of variables and construct a statistical model with high prediction accuracy, and to use this statistical model to predict the sinter ore yield with high accuracy. [Example]

[0054] In order to verify the effects of the present invention, prediction of the RDI of the product sintered ore (Example A), prediction of the TI (Tumbler Index) (Example B), and prediction of the yield of the sintered ore (Example C) were carried out.

[0055] [Example A] In Examples A1 and A2, the RDI of the product sintered ore was predicted based on the method for predicting the quality of sintered ore according to the first embodiment described above.

[0056] In Examples A1 and A2, a data set (365 records of daily aggregated data) was first acquired from the sinter ore production process for one year. The variables in the data set included the blending specifications (blending ratio, components) for each raw material brand during sinter ore production, firing specifications (temperature, pressure, air volume), heat source blending ratio, quicklime blending ratio, measured component values ​​of the finished sinter ore, and RDI of the finished sinter ore. Scaling processing was then performed on all variables in the acquired data set.

[0057] Next, using the dataset after scaling processing, Lasso regression was performed with variables other than RDI as input variables and RDI as output variable, and among the variables used as input variables, variables with absolute values ​​of coefficients greater than 0 were selected as input variables. Figure 5 shows the absolute values ​​of the coefficients of the variables obtained by Lasso regression and the variable selection judgment results in Examples A1 and A2.

[0058] In Example A1, as shown in Fig. 5, variables whose absolute values ​​of coefficients of Lasso regression are greater than 0 were used as input variables. That is, in Example A1, the average calcination temperature, the calcination temperature variance, the blending ratio of brand A, the blending ratio of brand B, the blending ratio of coke fines, and the blending ratio of quicklime were selected. In Example A2, in addition to the variables selected in Example A1, the blending ratio of brand C, which has a large influence on RDI based on conventional knowledge, was extracted by an operator from the input variables not selected, and selected as an input variable.

[0059] Thereafter, for each of Examples A1 and A2, the selected variables were used as input variables and the RDI was used as an output variable to perform learning by Ridge regression, and a Ridge regression model was constructed as a statistical model.The RDI was then predicted using the Ridge regression model.

[0060] In the comparative example, the RDI was predicted using the relational expression between the air flow rate passing through the sintering machine bed, RD and I described in Patent Document 3.

[0061] The relationship between the predicted and measured values ​​of RDI in Examples A1 and A2 and Comparative Example is shown in Figure 6. The coefficient of determination R for the regression equation in which the predicted and measured values ​​of RDI matched 2 was 0.93 for Example A1, 0.96 for Example A2, and 0.16 for the Comparative Example. That is, the prediction accuracy of Examples A1 and A2 was much higher than that of the Comparative Example, and Example A2 had a higher prediction accuracy than Example A1.

[0062] [Example B] In Example B, the TI of the sintered ore product was predicted based on the sintered ore quality prediction method according to the first embodiment. In Example B, a dataset in which the RDI of the sintered ore product in Example A was replaced with TI was used. After scaling, Lasso regression was performed with variables other than TI as input variables and TI as output variable, and variables with coefficients greater than 0 were selected as input variables. In Example B, the average firing temperature, firing temperature, firing temperature variance, blending ratio of brand C, blending ratio of brand D, and blending ratio of quicklime were selected. Ridge regression was then used to train the selected variables as input variables and TI as output variable, and a Ridge regression model was constructed as a statistical model. TI was then predicted using the Ridge regression model.

[0063] Figure 7 shows the relationship between the predicted and measured values ​​of TI. The coefficient of determination R for the regression equation that matches the predicted and measured values ​​of TI is 2 was a high value of 0.94, indicating that high prediction accuracy was also achieved in Example B.

[0064] [Example C] In Example C, the yield of sinter was predicted based on the sinter yield prediction method according to the first embodiment. In Example C, a dataset in which the RDI of the sintered ore in Example A was replaced with the yield was used. After scaling, Lasso regression was performed with variables other than the yield as the input variable and the yield as the output variable, and variables with coefficients greater than 0 were selected as input variables. In Example C, the average firing temperature, firing temperature, firing temperature variance, the blending ratio of brand C, the blending ratio of brand D, and the blending ratio of quicklime were selected. Ridge regression was then used to train the selected variables as input variables and the yield as the output variable, and a Ridge regression model was constructed as a statistical model. The yield was then predicted using the Ridge regression model.

[0065] Figure 8 shows the relationship between the predicted and measured yields. The coefficient of determination R for the regression equation that matches the predicted and measured yields is 2 was a high value of 0.76, indicating that high prediction accuracy was also achieved in Example C.

[0066] Although the preferred embodiments of the present invention have been described in detail above with reference to the accompanying drawings, the present invention is not limited to these examples. It is clear that a person skilled in the art to which the present invention pertains can conceive of various modifications and alterations within the scope of the technical ideas set forth in the claims, and it is understood that these also naturally fall within the technical scope of the present invention. [Explanation of symbols]

[0067] 10 Input Devices 20 Output Devices 100 Prediction Device 110 Variable Selection Section 120 Model Construction Department 130 Prediction Department

Claims

1. a variable selection step of performing Lasso regression using at least the blending specifications, firing specifications, heat source blending ratio, quicklime blending ratio, and sintered ore component values ​​for each raw material brand at the time of sintered ore production as input variables and the quality value of the sintered ore as an output variable, and selecting input variables whose coefficients assigned to each of the input variables are equal to or greater than a predetermined value; a statistical model learning step of learning a statistical model using the variables selected in the variable selection step as input variables and the quality value of the sintered ore as an output variable; a quality prediction step of predicting a quality value of sintered ore from prediction target data of the variables selected in the variable selection step using the trained statistical model; A method for predicting the quality of sintered ore, including:

2. a variable extraction step of extracting variables that are estimated to affect the quality of sintered ore based on past performance from among the variables not selected in the variable selection step, 2. The sinter ore quality prediction method according to claim 1, wherein in the statistical model learning step, the statistical model is trained using the variables selected in the variable selection step and the variables extracted in the variable extraction step as the input variables.

3. The method for predicting the quality of sintered ore according to claim 1 or 2, wherein the statistical model is a Ridge regression model.

4. a variable selection step of performing Lasso regression using at least the blending specifications, firing specifications, heat source blending ratio, quicklime blending ratio, and measured component values ​​of sintered ore for each raw material brand at the time of sintered ore production as input variables, and the measured yield value at the time of sintered ore production as output variables, and selecting input variables whose constants assigned to each of the input variables are equal to or greater than a predetermined value; a statistical model learning step of learning a statistical model using the variables selected in the variable selection step as input variables and the actual measured yield value during sinter production as output variables; a yield prediction step of predicting the yield of sintered ore from prediction target data of the variables selected in the variable selection step using the trained statistical model; A method for predicting the yield of sintered ore, comprising:

5. a variable extraction step of extracting variables that are estimated to affect the yield of sintered ore based on past performance from among the variables not selected in the variable selection step; 5. The sintered ore yield prediction method according to claim 4, wherein in the statistical model learning step, the statistical model is trained using the variables selected in the variable selection step and the variables extracted in the variable extraction step as the input variables.

6. The method for predicting a yield of sintered ore according to claim 4 or 5, wherein the statistical model is a Ridge regression model.

Images