Blast furnace coke ratio action guidance method, blast furnace coke ratio action guidance system, blast furnace control device, blast furnace coke ratio action guidance program, information output device, blast furnace operation method, and molten iron production method

Abstract

This blast furnace coke ratio action guidance method includes: a model learning processing step for performing machine learning using, as input data, image data obtained by imaging history data for observation amounts including air permeability, tapping speed, and molten iron temperature in a blast furnace process, and as output data, history data for a label indicating the direction of operation of the coke ratio applied within a predetermined time range including the timing of operation of the coke ratio by an operator, and thereby generating a machine learning model in which an input variable is image data for a guidance presentation period and an output variable is an action prediction value indicating the direction of operation of the coke ratio in the guidance presentation period; an action prediction value calculation step for inputting image data for a guidance presentation period into the machine learning model generated in the model learning processing step, and thereby calculating an action prediction value for the guidance presentation period; and an action guidance presentation step for presenting, on the basis of the action prediction value calculated in the action prediction value calculation step, guidance for an action of operating the coke ratio.

Description

Blast furnace coke rate action guidance method, blast furnace coke rate action guidance system, blast furnace control device, blast furnace coke rate action guidance program, information output device, blast furnace operation method, and molten iron manufacturing method

[0001] The present invention relates to a coke rate action guidance method for a blast furnace, a coke rate action guidance system for a blast furnace, a blast furnace control device, a coke rate action guidance program for a blast furnace, an information output device, a blast furnace operation method, and a molten iron manufacturing method.

[0002] In the steelmaking industry, blast furnaces (blast furnaces) are charged with iron ore and coke as raw materials from the top of the furnace, where they are reduced and melted, and then molten pig iron and slag are discharged from the bottom. In the blast furnace process, it is important to maintain the temperature of the molten pig iron product within a specified range. Furthermore, because the furnace is operated with solid raw materials filled inside, it is necessary to ensure good ventilation within the furnace to ensure the stable descent of the raw materials. Furthermore, it is necessary to maintain the molten pig iron production rate required for the next process, the steelmaking process.

[0003] In recent blast furnace operations, CO ２ To reduce emissions and production costs, it is necessary to reduce the coke ratio, which is the ratio of the iron content of the iron ore charged from the top of the furnace (equivalent to the weight of molten pig iron) to the weight of the coke. The coke ratio is defined as the weight of coke required to produce one ton of molten pig iron. Coke also serves as a reducing agent and a spacer to ensure gas permeability within the furnace. Furthermore, the molten pig iron production rate is determined by the product of the rate of coke consumption by oxygen and steam supplied into the furnace through the tuyere and the weight ratio of iron in the coke layer at the tuyere level and the iron in the iron ore layer.

[0004] In other words, the coke rate control affects all of the hot metal temperature, furnace permeability, and hot metal production rate. Therefore, the coke rate control is important for stable blast furnace operation by appropriately controlling these control variables. The coke rate is a control variable used to control the furnace permeability in particular. However, the coke rate control is highly dependent on the knowledge and experience of the operator. Therefore, a new method for predicting the overall coke rate control variable without being influenced by the operator's knowledge and experience is needed.

[0005] In light of this background, Patent Literature 1 proposes a method for using machine learning to construct a model that imitates the operational actions of an experienced blast furnace operator and for controlling the molten iron temperature using the constructed model. Specifically, the method described in Patent Literature 1 creates a model that determines the operational action to be taken in three stages, namely, increasing, decreasing, and waiting to see, in order to control the molten iron temperature of the blast furnace to a temperature close to a target temperature. The method described in Patent Literature 1 then controls the molten iron temperature of the blast furnace in accordance with the manipulated variable obtained by inputting trend data of items that are taken into consideration when the operator decides which operational action to take.

[0006] Japanese Patent Application Laid-Open No. 2021-18569

[0007] However, the method described in Patent Literature 1 has difficulty in suggesting operational actions that surpass the operational actions considered by the operator. In particular, the operational action of increasing the coke rate is often performed after an operational problem occurs, such as a sudden deterioration in furnace permeability or a sudden drop in molten iron temperature. Therefore, from the perspective of preventing operational problems, the timing at which the operator performs the operational action of increasing the coke rate is not necessarily appropriate. Therefore, rather than simply imitating the timing at which the operator increases the coke rate, it is necessary to detect signs of operational problems in advance and provide guidance on the operational action of increasing the coke rate ahead of the operator. However, the method described in Patent Literature 1 cannot provide guidance on such operational actions for the coke rate.

[0008] Furthermore, in general, coke rate adjustment is often performed after allowing time for the materials to descend into the blast furnace because it may result in excessive operational action. However, the method described in Patent Literature 1 does not perform machine learning taking into account such a time for the materials to descend. Therefore, according to the method described in Patent Literature 1, the learning data used in machine learning is unbalanced data with a small number of data items for coke rate adjustment, which may result in the construction of a model that makes it difficult to predict the appropriate timing for coke rate adjustment.

[0009] The present invention has been made to solve the above-mentioned problems, and an object of the present invention is to provide a blast furnace coke rate action guidance method, a blast furnace coke rate action guidance system, a blast furnace coke rate action guidance program, and an information output device that are capable of presenting optimal coke rate operation actions at optimal timing. Another object of the present invention is to provide a blast furnace control device and a blast furnace operation method that enable stable operation of a blast furnace. Another object of the present invention is to provide a molten iron production method that enables production of molten iron with a high yield.

[0010] The blast furnace coke rate action guidance method according to the present invention includes: a model learning processing step of performing machine learning using, as input data, image data obtained by visualizing historical data of observation quantities in a blast furnace process, including permeability, tapping rate, and molten iron temperature, and as output data, historical data of labels indicating a coke rate operation direction assigned within a predetermined time range including the timing of an operator's coke rate operation, to generate a machine learning model having, as an input variable, the image data for a guidance presentation period and an action predicted value indicating the coke rate operation direction for the guidance presentation period; an action predicted value calculation step of calculating, as an input variable, the image data for the guidance presentation period to the machine learning model generated in the model learning processing step; and an action guidance presentation step of presenting guidance for a coke rate operation action based on the action predicted value calculated in the action predicted value calculation step.

[0011] The label history data may be generated by assigning a label indicating that the operator decreased the coke rate within a time range of α hours before and after the timing at which the operator decreased the coke rate, and assigning a label indicating that the operator increased the coke rate within a time range of β hours before the timing at which the operator increased the coke rate, and by setting the time parameter β to be greater than the time parameter α.

[0012] The value of the time parameter β may be changed depending on the deviation of the molten iron temperature and the tapping rate from their target values ​​after a predetermined time has elapsed since the operator changed the coke rate.

[0013] If the coke rate is changed in the opposite direction to the previous coke rate change direction between the time when the operator changed the coke rate and the time when a predetermined time has elapsed, the previous coke rate change should be considered an incorrect operational action, and the historical data related to that change should be excluded from the learning data used during machine learning.

[0014] If the molten iron temperature or the tapping rate deviates from the target range or the permeability exceeds the upper control limit after a predetermined time has elapsed since the operator changed the coke rate, the coke rate change should be considered an incorrect operational action, and the historical data related to that change should be excluded from the learning data used during machine learning.

[0015] The blast furnace coke rate action guidance system according to the present invention may include: a model learning processing unit that performs machine learning using, as input data, image data that is an image of historical data of observation quantities in a blast furnace process, including permeability, tapping rate, and molten iron temperature, and as output data, historical data of labels that indicate a coke rate operation direction assigned within a predetermined time range including the timing of an operator's coke rate operation, to generate a machine learning model in which the image data for a guidance presentation period is used as an input variable and an action predicted value that indicates the coke rate operation direction for the guidance presentation period is used as an output variable; a coke rate action prediction unit that calculates the predicted action value for the guidance presentation period by inputting the image data for the guidance presentation period into the machine learning model generated by the model learning processing unit; and an action guidance presentation unit that presents guidance for a coke rate operation action based on the predicted action value calculated by the coke rate action prediction unit.

[0016] The action guidance presentation unit may present, as guidance for an operational action for the coke rate, an operational action for increasing the coke rate, an operational action for maintaining the coke rate, or an operational action for decreasing the coke rate, based on the action prediction value.

[0017] The action guidance presentation unit may present an operational action to increase the coke rate when the action prediction value exceeds an upper limit value, present an operational action to decrease the coke rate when the action prediction value falls below a lower limit value, and present an operational action to maintain the coke rate when the action prediction value is equal to or less than the upper limit value and equal to or greater than the lower limit value.

[0018] The action guidance presentation unit may include means for setting the upper limit value and the lower limit value.

[0019] The coke ratio action prediction unit may further include an output unit that outputs the action predicted value calculated by the coke ratio action prediction unit.

[0020] The blast furnace control device according to the present invention includes a means for controlling the blast furnace based on the action predicted value output from the output unit included in the blast furnace coke rate action guidance system according to the present invention.

[0021] The blast furnace coke rate action guidance program according to the present invention causes a computer to function as: a model learning processing unit that performs machine learning using, as input data, image data that is an image of historical data of observation quantities in a blast furnace process, including permeability, tapping rate, and molten iron temperature, and as output data, historical data of labels that indicate a coke rate operation direction assigned within a predetermined time range including the timing of an operator's coke rate operation, to generate a machine learning model in which the image data for a guidance presentation period is used as an input variable and an action predicted value that indicates the coke rate operation direction for the guidance presentation period is used as an output variable; a coke rate action prediction unit that calculates the action predicted value for the guidance presentation period by inputting the image data for the guidance presentation period into the machine learning model generated by the model learning processing unit; and an action guidance presentation unit that presents guidance for a coke rate operation action based on the action predicted value calculated by the coke rate action prediction unit.

[0022] The information output device according to the present invention is an information output device constituting the blast furnace coke rate action guidance system according to the present invention, and includes the action guidance presentation unit that presents guidance for an action to manipulate the coke rate based on the action predicted value output from the blast furnace coke rate action guidance system, and presents an operational action to increase the coke rate when the action predicted value exceeds an upper limit value, presents an operational action to decrease the coke rate when the action predicted value falls below a lower limit value, and presents an operational action to maintain the coke rate when the action predicted value is equal to or less than the upper limit value and equal to or greater than the lower limit value.

[0023] The action guidance presentation unit may include means for setting the upper limit value and the lower limit value.

[0024] The blast furnace operation method according to the present invention includes a step of manipulating the coke rate based on the guidance for the coke rate manipulation action presented by the blast furnace coke rate action guidance method according to the present invention.

[0025] The method for producing molten iron according to the present invention includes a step of producing molten iron by manipulating the coke rate based on the guidance for the coke rate manipulation action presented by the method for guiding coke rate action for a blast furnace according to the present invention.

[0026] The blast furnace coke rate action guidance method, blast furnace coke rate action guidance system, blast furnace coke rate action guidance program, and information output device according to the present invention make it possible to present an optimal coke rate operation action at an optimal timing. Furthermore, the blast furnace control device and blast furnace operation method according to the present invention make it possible to operate a blast furnace with high efficiency and stability. Furthermore, the molten iron production method according to the present invention makes it possible to produce molten iron with high efficiency, stability, and high yield.

[0027] FIG. 1 is a block diagram showing the configuration of a blast furnace coke rate action guidance device according to one embodiment of the present invention. FIG. 2 is a flowchart showing the flow of a model learning process according to one embodiment of the present invention. FIG. 3 is a diagram showing an example of time-series operation history data of the coke rate by an operator and labels assigned to the operation history data. FIG. 4 is a diagram showing the relationship between the deviation of a control index from a target value and a time parameter β. FIG. 5 is a diagram showing the time change in shaft pressure in a case where the operator's coke rate operation content was incorrect. FIG. 6 is a flowchart showing the flow of a guidance presentation process according to one embodiment of the present invention. FIG. 7 is a diagram showing a machine learning model of an example. FIG. 8 is a diagram showing the accuracy rate of an example. FIG. 9 is a diagram showing the time change in the aeration index and the blast pressure. FIG. 10 is a block diagram showing the configuration of a coke rate action guidance system according to one embodiment of the present invention.

[0028] Hereinafter, the configuration and operation of a coke rate action guidance device for a blast furnace according to one embodiment of the present invention will be described with reference to the drawings.

[0029] [Configuration] First, with reference to FIG. 1, the configuration of a coke rate action guidance device for a blast furnace according to one embodiment of the present invention will be described.

[0030] Fig. 1 is a block diagram showing the configuration of a blast furnace coke rate action guidance device according to one embodiment of the present invention. As shown in Fig. 1, a blast furnace coke rate action guidance device 1 (hereinafter abbreviated as guidance device 1) according to one embodiment of the present invention is configured by an information processing device such as a workstation or a personal computer. The guidance device 1 functions as a model learning processing unit 11, a coke rate action prediction unit 12, and an action guidance presentation unit 13 by an arithmetic processing unit within the information processing device, such as a CPU (Central Processing Unit), executing a computer program. The functions of each unit will be described later.

[0031] Furthermore, a blast furnace operation database (blast furnace operation DB) 2 in which operation data of the blast furnace is stored is connected to the guidance device 1 in a data-readable form. In this embodiment, the blast furnace operation DB 2 stores time-series history data of observation quantities including the permeability index (permeability), the tapping rate, and the molten iron temperature in the blast furnace, as well as time-series operation history data of the coke rate by the operator. An example of the permeability index in the blast furnace is the permeability resistance index ΔP / V expressed by the following mathematical formula (1). In mathematical formula (1), BP is the blast pressure [Pa], TP is the furnace top pressure [Pa], and BGV is the bosh gas volume [m ３ (standard conditions) / min].

[0032]

[0033] The guidance device 1 having such a configuration performs a model learning process and a guidance presentation process described below to guide the operator on the optimal coke ratio operation action. Hereinafter, the operation of the guidance device 1 when performing the model learning process and the guidance presentation process will be described with reference to the flowcharts shown in FIGS.

[0034] [Model Learning Process] First, the operation of the guidance device 1 when performing the model learning process will be described with reference to FIGS.

[0035] 2 is a flowchart showing the flow of the model learning process according to one embodiment of the present invention. The flowchart shown in FIG. 2 starts when an execution command for the model learning process is input to the guidance device 1, and the model learning process proceeds to step S1.

[0036] In step S1, the model learning processing unit 11 acquires, from the blast furnace operation DB 2, time-series historical data on observation quantities, including the permeability index, the tapping rate, and the molten iron temperature in the blast furnace, as well as time-series historical data on the coke rate operation by the operator. These observation quantities can be calculated from sensor information (e.g., the coke moisture content, the stock line, the shaft pressure, the pulverized coal ratio, the blast pressure, the permeability index, and the molten iron temperature) and the target value of the iron-making rate. This completes step S1, and the model learning process proceeds to step S2.

[0037] In the process of step S2, the model learning processing unit 11 assigns a label indicating the operation direction of the coke rate, which is the target of prediction (output data) by machine learning, to the time-series operation history data of the coke rate by the operator acquired in the process of step S1 for each unit time (e.g., every hour) of the operation history data. Hereinafter, a guideline for assigning the label will be described with reference to Fig. 3. Fig. 3 is a diagram showing an example of time-series operation history data of the coke rate (top CR) by the operator (Fig. 3(a)) and a label assigned to this operation history data (Fig. 3(b)).

[0038] Generally, an increase in the coke rate ２This leads to increased emissions and higher molten iron production costs. For this reason, the operational action of increasing the coke rate is not performed frequently, but is often performed after an operational problem (such as a drop in the tapping rate or molten iron temperature outside the control range) occurs or after the operator detects signs of an operational problem. In other words, the timing at which the operator actually increases the coke rate may be later than the appropriate timing. Therefore, in such cases, it is advisable to attach a label to increase the coke rate before the actual timing of the operator's action.

[0039] On the other hand, operational actions to reduce the coke rate are ２ From the perspective of reducing emissions and hot metal production costs, these actions are performed frequently and at optimal times, while taking into consideration concerns about deterioration of furnace permeability and a drop in hot metal temperature. Therefore, for operational actions to reduce the coke rate, it is sufficient to learn (imitate) the timing at which an operator actually reduces the coke rate. Therefore, in order to perform operational actions to reduce the coke rate at the appropriate timing, it is particularly important to ensure that operational actions to increase the coke rate are performed earlier than the timing at which the operator actually increases the coke rate.

[0040] Furthermore, for the reasons described above, the amount of operation per operation action that increases the coke rate (the amount of change in the coke rate) is greater than the amount of operation per operation action that decreases the coke rate. Furthermore, the frequency of occurrence of operation actions that decrease the coke rate is higher than the frequency of occurrence of operation actions that increase the coke rate. Therefore, when performing machine learning, it is desirable to eliminate the imbalance in the number of data items between operation actions that increase the coke rate and operation actions that decrease the coke rate.

[0041] 3A and 3B, the model learning processing unit 11 assigns a label of "coke rate decrease" to each unit time of the operation history data within an α-hour time range before and after the timing when the operator actually decreased the coke rate, indicating that the coke rate was decreased. The model learning processing unit 11 also assigns a label of "coke rate increase" to each unit time of the operation history data within a β-hour time range before the timing when the operator actually increased the coke rate, indicating that the coke rate was increased. The model learning processing unit 11 also assigns a label of "wait and see" to each unit time of the operation history data during other time periods, indicating that the coke rate was maintained.

[0042] The model learning processing unit 11 preferably sets the value of the time parameter β greater than the value of the time parameter α, thereby increasing the number of operation history data labeled "coke rate increase," so that the number of operation history data labeled "coke rate decrease" and the number of operation history data labeled "coke rate increase" are approximately equal. The time parameters α and β are preferably set between 0 and 8 hours, taking into account the elapsed time until the coke charged from the furnace top descends to the tuyere tip. The time parameter β is preferably varied depending on the deviation of the control indices (tapping rate and molten iron temperature) from their target values ​​after a predetermined time T has elapsed since the operator changed the coke rate. The predetermined time T is preferably set between 8 and 12 hours, taking into account the elapsed time until the coke charged from the furnace top descends to the tuyere tip.

[0043] Furthermore, if the raw material layer (charge) with an increased coke rate descends to the tuyere level after a predetermined time T, and the blast furnace control indicators (e.g., the tapping rate and molten iron temperature) do not increase and deviate from their target values ​​(e.g., the median of the control range of the control indicators), it is considered that the timing of the coke rate increase was inappropriate and that the timing of the coke rate increase should be brought forward. On the other hand, if the control indicators recover to near their target values ​​after a predetermined time T has elapsed since the operational action to increase the coke rate was taken, it is considered that the timing of the operational action to increase the coke rate was appropriate, and a label "coke rate increase" should be attached to coincide with the timing at which the operator increased the coke rate.

[0044] For this reason, as shown in FIG. 4 , the model learning processing unit 11 may gradually change the time parameter β so that the value of the time parameter β increases as the deviation of the control index from its target value (e.g., the median of each control range) increases after a predetermined time T has elapsed since the coke rate-increasing action was performed. The value d on the horizontal axis in FIG. 4 is a value obtained by normalizing the deviation δ of the control index from its target value R by dividing the deviation δ by the target value R. The control indexes may be the tapping rate and the molten iron temperature, which are control indexes for the iron ore reduction reaction. In this case, the value d may be calculated for each of the tapping rate and the molten iron temperature in each historical data, and the arithmetic mean or weighted mean of the values ​​d calculated for each historical data may be used as the value on the horizontal axis. This completes step S2, and the model learning process proceeds to step S3.

[0045] In the process of step S3, the model learning processing unit 11 generates learning data. The learning data is generated by linking input data and output data. The input data are observations acquired in the process of step S1, indicating the permeability index in the blast furnace, the tapping rate, and the molten iron temperature, which are the basis for the coke rate operational action performed by the operator. The model learning processing unit 11 then associates the input data with image data for a predetermined continuous time interval (from time t-γ to time t), and the output data with a label indicating the coke rate operation direction at time t. The input data may be normalized before being converted into image data to unify the range and dimensions. The image data is in the form of a two-dimensional image, with one axis representing the time axis and the other axis representing the sensor information and the target value of the ironmaking rate, in which the input data history data is aligned with the time axis and one or more input data are arranged in a direction different from the time axis. Furthermore, if the numbers of data items labeled with "reduce coke rate," "increase coke rate," and "wait and see" are uneven after the process of step S2, it is advisable to perform oversampling using SMOTE (Synthetic Minority Oversampling TEchnique) or the like on the training data whose output variables are labels with fewer data items so that the ratio of the numbers of data items approaches 1:1:1.

[0046] Furthermore, to suggest optimal coke rate actions, data on incorrect coke rate operation actions performed by the operator should be excluded from the learning data. As an example, Figures 5(a) and 5(b) show the change in shaft pressure over time in a case where the operator performed an incorrect coke rate operation action. In the case shown in Figures 5(a) and 5(b), the operator excessively reduced the coke rate at 16 hours, which deteriorated the permeability inside the blast furnace and caused the shaft pressure to rise. Therefore, the operator performed an action to increase the coke rate at 27 hours. In this way, if the operator changes the coke rate in the opposite direction to the previous operation within a predetermined time τ after changing the coke rate, it is preferable to consider the previous operation to be an incorrect coke rate operation action and exclude it from the learning data. N, W, S, and E in Figure 5 indicate the orientation of the blast furnace.

[0047] Furthermore, the predetermined time τ for determining the reverse operation action is preferably set between 8 and 12 hours, taking into consideration the elapsed time until the coke charged from the furnace top descends to the tuyere tip. Furthermore, if the molten iron temperature or the tapping rate deviates from the target range or the permeability inside the furnace exceeds the upper limit of control even without changing the coke rate in the opposite direction to the previous operation during the predetermined time τ after the operator changes the coke rate, these actions can be considered to be erroneous coke rate operation actions and excluded from the learning data. This completes the process of step S3, and the model learning process proceeds to step S4.

[0048] In step S4, the model learning processing unit 11 uses the learning data generated in step S3 to generate a machine learning model that predicts an optimal coke rate operational action from the time-series data of the sensor information and the target value of the ironmaking rate. The model generated by machine learning is preferably a convolutional neural network (CNN). Furthermore, if the image data is in the form of a two-dimensional image with a time axis on one axis and historical data on the other axis, the CNN convolution operation should be performed only in the time axis direction. This completes step S4, and the model learning process proceeds to step S5.

[0049] In step S5, the model learning processing unit 11 outputs a machine learning model that predicts the optimal coke rate operational action based on the sensor information and the time change in the target value of the ironmaking rate, using the image data generated in step S4, which visualizes the time change in the observed quantities during the guidance presentation period (time t - γ to time t), as input variables, and the coke rate operational action at time t as output variables. The time γ is preferably set between 8 and 120 hours, taking into account the time it takes for the materials to descend into the blast furnace and the time it takes for the coke to be replaced in the core. This completes step S5, and the model learning process ends.

[0050] When new historical data is added to the blast furnace operation DB2, it is preferable that the model learning processing unit 11 updates the model by additionally learning the coke rate action prediction model using the newly added sensor information and iron-making rate as additional learning data.

[0051] [Guidance Presentation Process] Next, with reference to FIG. 6, the operation of the guidance device 1 when performing the guidance presentation process will be described.

[0052] 6 is a flowchart showing the flow of the guidance presentation process according to one embodiment of the present invention. The flowchart shown in FIG. 6 starts when an execution command for the guidance presentation process is input to the guidance device 1, and the guidance presentation process proceeds to step S11.

[0053] In the process of step S11, the action guidance presentation unit 13 obtains the guidance presentation period (time t ０ -γ ~ time t ０ , where time t ０ The time-series historical data of the observation quantities including the permeability index, the tapping rate, and the molten iron temperature in the blast furnace at the time (is the current time) is acquired. This completes the processing of step S11, and the guidance presentation processing proceeds to the processing of step S12.

[0054] In the process of step S12, the action guidance presenter 13 inputs the time-series history data of the observation amount acquired in the process of step S11 into the machine learning model as an input variable, thereby obtaining the guidance presenting period (time t ０ -γ ~ time t ０ ) based on the historical data of the time series of observations at the current time t ０ Specifically, the machine learning model generated by the model learning process outputs the probabilities of the operational actions of "coke rate reduction," "coke rate increase," and "wait and see." １ , P ２ , P ３ Then (P １ +P ２ +P ３ = 1), the predicted value of the coke ratio action P is P = P ２ -P １ This completes the process of step S12, and the guidance presentation process proceeds to step S13.

[0055] In step S13, the action guidance presentation unit 13 presents the operator with a recommended coke rate operational action based on the predicted action value P calculated in step S12. Specifically, when the predicted action value P exceeds the upper limit value ε (P>ε), the action guidance presentation unit 13 presents an operational action to increase the coke rate. Furthermore, when the predicted action value P falls below the lower limit value (−η) (P<−η), the action guidance presentation unit 13 presents an operational action to decrease the coke rate. Furthermore, when the predicted action value P is equal to or less than the upper limit value ε and equal to or greater than the lower limit value (−η), the action guidance presentation unit 13 presents an operational action to maintain the coke rate. An operational action to increase the coke rate is urgently required to avoid a deterioration in gas permeability in the blast furnace and a decrease in the molten iron temperature. On the other hand, an operational action to decrease the coke rate entails the risk of a deterioration in gas permeability in the blast furnace and a decrease in the molten iron temperature, so it must be presented with caution. Therefore, it is preferable to set the upper limit value ε and the lower limit value (-η) in a range that satisfies 0<ε≦η<1. The action guidance presenter 13 may also include a setting means for the upper limit value ε and the lower limit value (-η). Examples of the setting means include a tuning screen or a control on an operator control panel that allows the upper limit value ε and the lower limit value (-η) to be changed. This completes the processing of step S13, and the series of guidance presentation processing ends.

[0056] As is clear from the above description, in the blast furnace coke rate action guidance device 1 according to one embodiment of the present invention, the model learning processing unit 11 performs machine learning using, as input data, image data that visualizes historical data of observation quantities in the blast furnace process, including permeability, tapping rate, and molten iron temperature, and as output data, historical data of labels indicating the coke rate operation direction assigned within a predetermined time range including the timing of the operator's coke rate operation, to generate a machine learning model having, as input variables, the image data for a guidance presentation period and, as output variables, a predicted action value indicating the coke rate operation direction for the guidance presentation period. The coke rate action prediction unit 12 then inputs the image data for the guidance presentation period into the generated machine learning model to calculate a predicted action value for the guidance presentation period, and the action guidance presentation unit 13 presents guidance for the coke rate operation action based on the calculated predicted action value.

[0057] According to this configuration, machine learning is performed using historical data of labels indicating the coke rate operation direction assigned within a predetermined time range including the timing of the operator's coke rate operation, so that an optimal coke rate operation action can be presented at the optimal timing. Furthermore, by operating the coke rate based on the presented guidance for the coke rate operation action, the blast furnace can be operated with high efficiency and stability. Furthermore, by producing molten iron by operating the coke rate based on the presented guidance for the coke rate operation action, molten iron can be produced with high efficiency, stability, and high yield.

[0058] In this example, the guidance presentation period γ was set to 24 hours, and features were extracted using a CNN based on sensor information and the target ironmaking rate acquired every hour to predict the optimal coke rate operational action. Using 22,500 pieces of sensor information (approximately 600 days' worth) and the target ironmaking rate, the input and output data were linked. The ratio of the data items for "coke rate reduction," "coke rate increase," and "wait and see" was unbalanced at 2:1:20. Therefore, SMOTE was applied to the remaining 22,000 pieces, excluding 500 pieces for model accuracy verification, to generate 57,500 pieces of training data, so that the ratio of the data items for "coke rate reduction," "coke rate increase," and "wait and see" approached 1:1:1. Next, machine learning of the model using the CNN was performed using the first 57,000 pieces of training data, and the accuracy of the model was verified using the remaining 500 pieces of training data.

[0059] The structure of the CNN used is shown in Figure 7. The blocks in Figure 7 represent a convolution layer (Convolution), a pooling layer (Pooling), a pooling layer (Pooling), a pooling layer (Pooling), a fully connected layer (Affine), a dropout layer (Dropout), a fully connected layer (Affine), a dropout layer (Dropout), and a fully connected layer (Affine). The numbers in parentheses in Figure 7 indicate the size of the data array. For example, (13, 24, 64) indicates a three-dimensional array of 13 vertical (variables) × 24 horizontal (time) × 64 channels. However, the CNN structure shown in Figure 7 is merely an example, and structures other than that shown in Figure 7 may also be used.

[0060] The verification results of this example are shown in Figure 8. In Figure 8, the total number of cases in which the models that were machine-learned using the same verification data and the following three methods predicted a "coke rate decrease" or a "coke rate increase" is N, and among those, the total number of cases in which the molten iron temperature and the tapping rate were within the target range and the permeability in the blast furnace was below the upper limit of control 8 hours after the coke rate was adjusted is N. １ In this case, N １The first method (Example 1) is a case where machine learning of a model is performed using learning data in which erroneous manipulation of the coke rate by the operator is also included in the learning data, and only the timings when the operator actually manipulated the coke rate are labeled, with both α and β set to 0 time.

[0061] The second method (Example 2) involved performing machine learning after excluding training data on incorrect coke rate control by the operator. The third method (Example 3) involved performing machine learning on the model using training data in which labels were added to extend the timing of the operator's coke rate control, with α set to 1 hour and β varied between 1 and 8 hours depending on the deviations from the target values ​​of the hot metal temperature and tapping rate at a predetermined time T = 8 hours. As shown in Figure 8 , the accuracy rate improved in the order of Examples 1, 2, and 3. This confirmed that the present invention can predict appropriate coke rate control.

[0062] 9(a) and 9(b), in the case where the model predicted a coke rate increase but the actual operation by the operator was a "wait and see" operation, the permeability index and blast pressure included in the input variables increased starting from two to three hours before the timing indicated by the dashed line when the model predicted the coke rate increase, and it was possible to present a coke rate action that predicted a deterioration in permeability inside the blast furnace. This confirmed that the present invention can predict and provide guidance on coke rate operational actions that surpass the ability of the operator.

[0063] Although the present invention has been described above as an embodiment applying the invention made by the present inventors, the present invention is not limited by the description and drawings that form part of the disclosure of the present invention according to the present embodiment. For example, the image data may be in the form of a two-dimensional image in which, for each input variable, one axis represents the time axis and the other axis represents the range of values ​​that the variable can take. Labels of multiple pieces of image data and coke rate manipulations may be associated, and CNN convolution operations may be performed along the time axis and the variable axis. Furthermore, instead of the three categories of "coke rate reduction," "coke rate increase," and "wait and see," the labels may be the coke rate manipulation amounts themselves, and a model may be used to predict the coke rate manipulation amounts.

[0064] Furthermore, a blast furnace coke rate action guidance system may be constructed using the blast furnace coke rate action guidance device according to one embodiment of the present invention. FIG. 10 is a block diagram showing an example of the configuration of a blast furnace coke rate action guidance system. The blast furnace coke rate action guidance system 10 shown in FIG. 10 includes the blast furnace operation DB 2 shown in FIG. 1 , an information processing device 20, and an information output device 30. The blast furnace operation DB 2, the information processing device 20, and the information output device 30 are configured to be able to communicate with each other via a network N such as the Internet, for example. In the coke rate action guidance system, the information output device 30 is located, for example, in a steelworks.

[0065] The information processing device 20 includes the model learning processing unit 11, the coke ratio action prediction unit 12, and the output unit 21 shown in FIG. 1 . The output unit 21 transmits the action prediction value P, which is the output of the coke ratio action prediction unit 12, to the information output device 30 or another information processing device via the network N. The other information processing device to which the output unit 23 transmits the action prediction value P may include, for example, a control computer for a blast furnace. This control computer may control the coke rate of the blast furnace based on the transmitted action prediction value P. Furthermore, the control computer may control the coke rate of the blast furnace by accepting a correction by an operator.

[0066] The information output device 30 includes an output information acquisition unit 31 and the action guidance presentation unit 13 shown in Fig. 1. The output information acquisition unit 31 acquires an action prediction value P from the information processing device 20, and the action guidance presentation unit 13 presents an operation action for the coke rate based on the action prediction value P. The action guidance presentation unit 13 may include a tuning screen that displays and enables setting of judgment thresholds (upper limit value ε, lower limit value -η) for the action prediction value when presenting an operation action for the coke rate based on the action prediction value.

[0067] The information processing device 20 and the information output device 30 may not include all of the components shown in FIG. 10 . Furthermore, the information processing device 20 and the information output device 30 may include components other than the components shown in FIG. 10 . Furthermore, the components included in each of the information processing device 20 and the information output device 30 are not limited to the example shown in FIG. 10 . For example, some of the components included in the information processing device 20 may be provided on the information output device 30 side. Furthermore, for example, some of the components included in the information output device 30 may be provided on the information processing device 20 side. Therefore, the coke ratio action guidance system including the information processing device 20 and the information output device 30 may include, as a whole, a model learning processing unit 11, a coke ratio action prediction unit 12, and an action guidance presentation unit 13.

[0068] In this way, all other embodiments, examples, operational techniques, etc. made by those skilled in the art based on this embodiment are included in the scope of the present invention.

[0069] According to the present invention, it is possible to provide a blast furnace coke rate action guidance method, a blast furnace coke rate action guidance system, a blast furnace coke rate action guidance program, and an information output device that are capable of presenting an optimal coke rate operation action at an optimal timing. Also, according to the present invention, it is possible to provide a blast furnace control device and a blast furnace operation method that enable stable operation of a blast furnace. Also, according to the present invention, it is possible to provide a molten iron production method that enables molten iron to be produced with a high yield.

[0070] REFERENCE SIGNS LIST 1 Blast furnace coke rate action guidance device 2 Blast furnace operation database (blast furnace operation DB) 10 Blast furnace coke rate action guidance system 11 Model learning processing unit 12 Coke rate action prediction unit 13 Action guidance presentation unit 20 Information processing device 21 Output unit 30 Information output device 31 Output information acquisition unit N Network

Claims

1. A coke rate action guidance method for a blast furnace, comprising: a model learning processing step of performing machine learning using, as input data, image data obtained by visualizing historical data of observation quantities in a blast furnace process, including permeability, tapping rate, and molten iron temperature, and as output data, historical data of labels indicating a coke rate operation direction assigned within a predetermined time range including the timing of an operator's coke rate operation, to generate a machine learning model having, as an input variable, the image data for a guidance presentation period and an action predicted value indicating the coke rate operation direction for the guidance presentation period; an action predicted value calculation step of calculating, as an input variable, the image data for the guidance presentation period to the machine learning model generated in the model learning processing step; and an action guidance presentation step of presenting guidance for a coke rate operation action based on the action predicted value calculated in the action predicted value calculation step.

2. The coke rate action guidance method for a blast furnace according to claim 1, wherein the historical label data is generated by assigning a label indicating that the operator decreased the coke rate within a time range of α hours before and after the timing at which the operator decreased the coke rate, and assigning a label indicating that the operator increased the coke rate within a time range of β hours before the timing at which the operator increased the coke rate, and setting the time parameter β to be greater than the time parameter α.

3. The coke rate action guidance method for a blast furnace according to claim 2, wherein the value of the time parameter β changes depending on the deviations of the molten iron temperature and the tapping rate from their target values ​​after a predetermined time has elapsed since the operator manipulated the coke rate.

4. A blast furnace coke rate action guidance method according to any one of claims 1 to 3, wherein if the coke rate is operated in a direction opposite to the previous coke rate operation direction between the time when the operator operated the coke rate and the time when a predetermined time has elapsed, the previous coke rate operation is considered to be an erroneous operational action, and historical data related to the operation is excluded from the learning data used in machine learning.

5. A blast furnace coke rate action guidance method according to any one of claims 1 to 4, wherein if the molten iron temperature or the tapping rate deviates from the target range or the permeability exceeds the upper control limit after a predetermined time has elapsed since the operator changed the coke rate, the coke rate change is considered to be an erroneous operational action, and historical data related to the change is excluded from the learning data used in machine learning.

6. A coke rate action guidance system for a blast furnace, comprising: a model learning processing unit that performs machine learning using, as input data, image data that is an image of historical data of observation quantities in a blast furnace process, including permeability, tapping rate, and molten iron temperature, and as output data, historical data of labels that indicate a coke rate operation direction assigned within a predetermined time range including the timing of an operator's coke rate operation, to generate a machine learning model in which the image data for a guidance presentation period is an input variable and an action predicted value that indicates the coke rate operation direction for the guidance presentation period is an output variable; a coke rate action prediction unit that calculates the action predicted value for the guidance presentation period by inputting the image data for the guidance presentation period into the machine learning model generated by the model learning processing unit; and an action guidance presentation unit that presents guidance for a coke rate operation action based on the action predicted value calculated by the coke rate action prediction unit.

7. The coke rate action guidance system for a blast furnace according to claim 6, wherein the action guidance presentation unit presents, as guidance for the coke rate operation action, an operation action to increase the coke rate, an operation action to maintain the coke rate, or an operation action to decrease the coke rate, based on the action prediction value.

8. The coke rate action guidance system for a blast furnace according to claim 7, wherein the action guidance presentation unit presents an operational action to increase the coke rate when the action prediction value exceeds an upper limit value, presents an operational action to decrease the coke rate when the action prediction value falls below a lower limit value, and presents an operational action to maintain the coke rate when the action prediction value is equal to or less than the upper limit value and equal to or greater than the lower limit value.

9. A blast furnace coke rate action guidance system according to claim 8, wherein the action guidance presentation unit includes means for setting the upper limit value and the lower limit value.

10. A blast furnace coke rate action guidance system according to any one of claims 6 to 9, further comprising an output unit that outputs the action prediction value calculated by the coke rate action prediction unit.

11. A blast furnace control device comprising means for controlling a blast furnace based on the action prediction value output from an output unit provided in the blast furnace coke rate action guidance system according to claim 10.

12. A coke ratio action guidance program for a blast furnace that causes a computer to function as: a model learning processing unit that performs machine learning using, as input data, image data that is an image of historical data of observation quantities in a blast furnace process, including permeability, tapping rate, and molten iron temperature, and as output data, historical data of labels that indicate a coke ratio operation direction assigned within a predetermined time range including the timing of an operator's coke ratio operation, to generate a machine learning model in which the image data for a guidance presentation period is an input variable and an action predicted value that indicates the coke ratio operation direction for the guidance presentation period is an output variable; a coke ratio action prediction unit that calculates the action predicted value for the guidance presentation period by inputting the image data for the guidance presentation period into the machine learning model generated by the model learning processing unit; and an action guidance presentation unit that presents guidance for a coke ratio operation action based on the action predicted value calculated by the coke ratio action prediction unit.

13. An information output device constituting the blast furnace coke rate action guidance system according to claim 6, comprising: an action guidance presentation unit that presents guidance for a coke rate manipulation action based on the action prediction value output from the blast furnace coke rate action guidance system, wherein the information output device presents an operational action to increase the coke rate when the action prediction value exceeds an upper limit value, presents an operational action to decrease the coke rate when the action prediction value falls below a lower limit value, and presents an operational action to maintain the coke rate when the action prediction value is equal to or less than the upper limit value and equal to or greater than the lower limit value.

14. The information output device according to claim 13, wherein the action guidance presentation unit includes means for setting the upper limit value and the lower limit value.

15. A method for operating a blast furnace, comprising a step of manipulating the coke rate based on guidance for a coke rate manipulation action presented by the method for guiding a coke rate action for a blast furnace according to any one of claims 1 to 5.

16. A method for producing molten iron, comprising the step of producing molten iron by manipulating the coke rate based on the guidance for the coke rate manipulation action presented by the method for guiding coke rate action in a blast furnace according to any one of claims 1 to 5.

Images