JP7644337B2 - Machine learning device for blast furnace operation, program used therein, method for generating blast furnace operation prediction model, and method for operating a blast furnace - Google Patents

Description

The present invention relates to a machine learning device for blast furnace operation.

Systems that use physical models (e.g., a system that predicts molten iron temperature using a non-steady-state model) are known as systems for predicting and controlling blast furnace operations. Prediction systems that use physical models can achieve high prediction accuracy when the boundary conditions are clear, but the prediction accuracy decreases when the boundary conditions are unknown (e.g., when the composition of the raw fuel is unknown, or when the filling status of the packed bed in the furnace is unknown).

Another type of prediction system is the expert model, which makes predictions based on a knowledge base expressed by rules. With expert models, if time has passed since the rules were created and operating conditions have changed, the rules need to be revised.

One method known to solve these problems is to search for similar examples from recent past data and generate a local statistical model for the similar examples to make future predictions. However, this model has problems in that it is difficult to make predictions when operating in operating conditions that are different from the conventional ones, and the prediction accuracy decreases when there are not enough similar past examples accumulated to guarantee the accuracy of the local statistical model.

JP 2017-095761 A JP 2018-193579 A

The inventors focused on the operational data accumulated daily by the blast furnace process computer and investigated a method of generating a blast furnace operation prediction model by machine learning using this operational data as training data. However, blast furnace operation data is not necessarily normally distributed, and much of it exhibits a biased distribution. Many models based on machine learning assume that the training data is normally distributed, and when non-normally distributed operational data is selected as training data, the prediction accuracy decreases. Therefore, the present invention aims to provide a blast furnace operation prediction model and the like with high prediction accuracy.

In order to solve the above problem, the machine learning device for blast furnace operation according to the present invention is characterized by having a data pre-processing unit that converts non-normally distributed data, of the blast furnace operation data, whose skewness exceeds a skewness threshold, into normally distributed data, thereby unifying the blast furnace operation data into normally distributed data; a data type selection unit that selects an explanatory variable corresponding to a predetermined objective variable based on the data type of each operation data included in the blast furnace operation data; and a model generation unit that generates a blast furnace operation prediction model by performing machine learning using the blast furnace operation data corresponding to the selected explanatory variable and the blast furnace operation data corresponding to the objective variable as training data.

(2) The machine learning device for blast furnace operation described in (1) above, characterized in that the data pre-processing unit performs a removal process to remove outliers contained in each operation data before converting non-normally distributed data into normally distributed data.

(3) The machine learning device for blast furnace operation described in (2) above, characterized in that the data pre-processing unit performs the removal process in accordance with a Hampel Identifier method based on the following equation (1).
median(xi)±cMAD(xi)...Formula (1)
Here, MAD(xi) = median (|xi - median(xi)|), median is the median value, and c is a parameter selected from the range of 3 to 5 (however, not limited to integers).

(4) The machine learning device for blast furnace operation described in any one of (1) to (3) above, characterized in that when the blast furnace operation data includes the same type of operation data that are acquired at different times, the data type selection unit selects explanatory variables by treating the data types of the operation data as different data types.

(5) A method for operating a blast furnace, comprising predicting the operating state of a blast furnace using any one of the machine learning devices for blast furnace operation described above in (1) to (4) and operating the blast furnace by executing an action based on the prediction result.

(6) A machine learning device for blast furnace operation, comprising: a data pre-processing unit that converts non-normally distributed data, the skewness of which exceeds a skewness threshold, into normally distributed data among the blast furnace operation data obtained after carrying out the action described in (5) above, to unify the blast furnace operation data into normally distributed data; a data type selection unit that selects explanatory variables corresponding to a predetermined objective variable based on the data type of each piece of operation data included in the blast furnace operation data; and a model generation unit that generates a blast furnace operation prediction model by performing machine learning using the blast furnace operation data corresponding to the selected explanatory variables and the blast furnace operation data corresponding to the objective variable as training data.

(7) A program used in a machine learning device for blast furnace operation, which causes a computer to execute the following steps: a conversion step of unifying blast furnace operation data into normal distribution data by converting non-normal distribution data, the skewness of which exceeds a skewness threshold, into normal distribution data; a data type selection step of selecting explanatory variables corresponding to a predetermined objective variable based on the data type of each operation data included in the blast furnace operation data; and a model generation step of generating a blast furnace operation prediction model by machine learning using the blast furnace operation data corresponding to the explanatory variable selected in the data type selection step and the blast furnace operation data corresponding to the objective variable as training data.

(8) A method for generating a blast furnace operation prediction model, comprising: a conversion step of unifying the blast furnace operation data into normal distribution data by converting non-normal distribution data, the skewness of which exceeds a skewness threshold, into normal distribution data; a data type selection step of selecting an explanatory variable corresponding to a predetermined objective variable based on the data type of each operation data included in the blast furnace operation data; and a model generation step of generating a blast furnace operation prediction model by machine learning using the blast furnace operation data corresponding to the explanatory variable selected in the data type selection step and the blast furnace operation data corresponding to the objective variable as training data.

The present invention provides a blast furnace operation prediction model with high prediction accuracy.

FIG. 1 is a block diagram of a machine learning device for blast furnace operation. 1 is a flowchart for explaining a method for constructing a blast furnace operation prediction model, etc. This is the frequency distribution of ηCO. This is the frequency distribution of molten iron temperature (℃). The relative RMSE when predicting the hot metal temperature with different parameters c. This is a frequency distribution of solution loss carbon amount. This is the frequency distribution of skewness of all operational data obtained over a certain period of time. This is the frequency distribution of the average furnace top temperature (℃). This is the frequency distribution of blast furnace heat load (kW). FIG. 13 is a diagram showing the relationship between the skewness threshold and the relative RMSE. FIG. 1 is a graph showing the relationship between the change in pulverized coal injection amount and the change in molten iron temperature. 13 shows the relative RMSE of the predicted values of the molten iron temperature for each of the comparative examples and the example. 4 shows the coefficient of variation of the coke ratio in each of the comparative examples and the examples.

FIG. 1 is a block diagram of a machine learning device for blast furnace operation according to the present embodiment. Referring to the figure, the machine learning device for blast furnace operation 1 includes a data pre-processing unit 10, a data type selection unit 20, a learning model generation unit 30, and a storage medium 40, and performs a predetermined data processing on operation data acquired from a blast furnace 2, selects explanatory variables corresponding to a predetermined objective variable based on the data type of each operation data included in the blast furnace operation data, and performs machine learning using the blast furnace operation data corresponding to the selected explanatory variables and the blast furnace operation data corresponding to the objective variable as training data, thereby constructing a blast furnace operation prediction model with high prediction accuracy. A method for constructing a blast furnace operation prediction model will be described in detail with reference to the flowchart in FIG. 2.

(Step S1: Data pre-processing)
In step S1, the data pre-processing unit 10 performs pre-processing of the operation data acquired from the blast furnace 2. The pre-processing of the operation data includes outlier removal, interpolation processing (step S1-1), and normal distribution processing (step S1-2). However, step S1-1 may be omitted. In this case, the processing flow of the acquired operation data starts from step S2.

(Step S1-1: Outlier Removal and Interpolation)
The operation data of a blast furnace includes information acquired from the blast furnace during operation (including blasting stoppage) and information calculated from the acquired information. The operation data of a blast furnace may contain missing data or outliers due to various reasons (e.g., system malfunctions). If the operation data contains outliers or is partially missing, it will affect the selection of training data and the prediction accuracy of a prediction model. Therefore, it is desirable to remove the outliers (corresponding to "outlier removal") or to interpolate a portion of the missing operation data (corresponding to "interpolation processing") according to a specified condition. The specified conditions will be described later.

Here, the operational data includes, for example, blast furnace operating conditions, charging raw fuel conditions, iron and slag data, sensor data, solution loss carbon, heat balance above the tuyere, and the like.
The operating conditions of the blast furnace include the amount of air blown from the tuyere, the oxygen enrichment rate of the oxygen-enriched air blown from the tuyere (oxygen enrichment rate: (amount of oxygen in blast + amount of enriched oxygen) / (amount of air blown) × 100 - oxygen concentration in blast (%)), the blast temperature of the gas blown from the tuyere, the blast pressure of the gas blown from the tuyere, the amount of pulverized coal blown, the furnace top pressure, etc.
The charging raw material fuel conditions include the amount of blast furnace raw materials charged, the components of the blast furnace raw materials, etc. The blast furnace raw materials include the ore layer charging raw materials used to form the ore layer, and the coke layer charging raw materials used to form the coke layer. The ore layer charging raw materials include sintered ore, pellets, lump ore, and non-calcined carbon-containing agglomerated ore. The ore layer charging raw materials may also include materials other than ore (e.g., reduction assistant materials such as small lump coke). The coke layer charging raw materials may also include ferro coke.
Tapping slag refers to the high-temperature molten material removed from the taphole of a blast furnace, and tapping slag data includes the amount of molten metal discharged, the molten metal temperature, the molten metal components, the amount of slag discharged, and the slag components.
Sensor data refers to data detected directly from sensors installed in the blast furnace, and includes furnace gas temperature data, analysis data of furnace gas components (CO, _CO2 , _N2 , _H2 , etc.), stave temperature data, furnace bottom side wall temperature data, furnace bottom temperature data, shaft pressure data, etc.
The heat balance above the tuyere is the difference between a specified input heat quantity (the sensible heat of the gas blown through the tuyere, the heat of combustion of carbon burning in the furnace) and a specified output heat quantity (the heat of decomposition of the moisture in the gas blown through the tuyere, the reaction heat due to the direct reduction reaction and the solution loss reaction), and is an index for successively evaluating the heat balance in the furnace.

The process of removing outliers can be performed using, for example, the "Hampel Identifier method " etc. The following formula (1) is a formula used to remove outliers in the "Hampel Identifier method ".

Here, MAD(xi) = median (|xi - median(xi)|), median is the median value, and c is a parameter selected from the range of 3 to 5 (however, it is not limited to an integer; the same applies below).
In the Hampel Identifier method , values outside the range of "median(xi) - cMAD(xi)" to "median(xi) + cMAD(xi)" are treated as "outliers."

As is clear from formula (1), parameter c is an important parameter related to determining outliers, and the inventors have found that setting this parameter c within a specified range improves prediction accuracy. The specified range is preferably 3 to 5 (hereinafter also referred to as the "preferred range"). If the main purpose is to predict an operating state that occurs relatively infrequently, parameter c needs to be set relatively large, and if the main purpose is to predict a relatively stable operating state, parameter c needs to be set relatively small.

If some of the blast furnace operation data is missing due to the removal of outliers, it is desirable to perform an interpolation process. There are no particular limitations on the method for interpolating the missing operation data, but for example, a method of interpolating using the previous value or the average value of the previous and subsequent values, linear interpolation based on the previous and subsequent values, etc. can be used.

As described above, the Hampel Identifier method is a method for determining whether or not a value is an outlier based on the degree of deviation from the median, not the average value, so it can identify outliers even in data that does not necessarily follow a normal distribution. In this respect, the Hampel Identifier method is an excellent means for processing blast furnace operation data.

The method of determining the parameter c will be described in more detail with a specific example. FIG. 3 shows the frequency distribution of the furnace top CO gas utilization rate (%) (hereinafter referred to as "ηCO") obtained from a blast furnace (furnace volume: 5000 ^m3 class) operated under predetermined operating conditions. The frequency is expressed as a relative frequency. The relative frequency is obtained by dividing the frequency of each section by the sum of the frequencies of all sections. The definition of the relative frequency is the same in FIG. 4 and FIG. 6 to FIG. 9. The predetermined operating conditions are a reducing agent ratio of 500 kg/pig-t, a coke ratio of 330 kg/pig-t, and a productivity of 2.2. The horizontal axis is ηCO, and the vertical axis is the relative frequency. In blast furnace operation, the frequency of ηCO becoming an outlier is low, so the vertical axis is displayed in logarithm. The threshold value of the outlier when the parameter c is changed in the range of 2 to 6 can be seen from the figure. That is, the intersection of the dotted line extending parallel to the vertical axis and the horizontal axis is the outlier threshold value corresponding to each parameter c. For example, when the parameter c is set to 2, the numerical values outside the area between the pair of dotted lines corresponding to c=2 (indicated by the arrows) are determined to be outliers.

However, if parameter c is set to 2, approximately 20% of the data will be classified as outliers, and the outlier threshold will be excessively large. Furthermore, if parameter c is set to 6, only about 0.3% of the data will be classified as outliers, and there is a possibility that significant outliers will be missed. Therefore, as mentioned above, parameter c is preferably between 3 and 5, and basically parameter c should be set to 4, which is the middle value of the "preferred range".

However, depending on the region to be predicted, parameter c can be changed within the "preferred range" based on the following ideas. Specifically, when making predictions near the median, parameter c can be set to a value of around 3, allowing data with a frequency of approximately 90% or more to be treated as normal values. When wishing to improve the prediction accuracy of relatively infrequent operating conditions, it is desirable to set parameter c to around 5. This makes it possible to use operational data with a relatively low frequency of around 0.3% to 1%. In this way, prediction accuracy can be improved by appropriately setting parameter c within the "preferred range" depending on the purpose of prediction.

FIG. 4 shows the frequency distribution of molten iron temperature (°C) obtained from a blast furnace (furnace volume: 5000 ^m3 class), with the same operating conditions as in FIG. 3. The horizontal axis is molten iron temperature, and the vertical axis is relative frequency. Note that in blast furnace operation, the molten iron temperature rarely becomes an outlier, so the vertical axis is plotted in logarithm. The outlier thresholds when the parameter c is changed from 2 to 6 can be seen from the figure. This is the same as in the case of ηCO in FIG. 3.

If parameter c is set to 2, approximately 20% of the operational data will be classified as outliers, and the outlier threshold will be excessively large. Furthermore, if parameter c is set to 6, only about 0.3% of the operational data will be classified as outliers, and there is a possibility that significant outliers will be missed. Therefore, as in the case of ηCO, parameter c is preferably between 3 and 5, and it is basically considered to set parameter c to 4, which is the intermediate value of the "preferred range." As in the case of ηCO, parameter c can be changed within the "preferred range" depending on the purpose of the prediction, but this point will not be explained again.

Figure 5 shows the relative RMSE when predicting the molten iron temperature while changing the value of parameter c. In Figure 5, the horizontal axis is parameter c, and the vertical axis is relative RMSE. "RMSE" is an index of prediction error, and is also called the root mean square error. Relative RMSE is normalized with the reference RMSE (RMSE when c = 2 in Figure 5, and RMSE when the frequency threshold is 0 in Figure 10) set to 1. The general formula for RMSE will be described later. As the relative RMSE decreases, the prediction accuracy improves. It can be seen from this figure that the relative RMSE is smallest and the prediction accuracy is highest when parameter c is set to the neighborhood of 4.

FIG. 6 shows the frequency distribution of the solution loss carbon amount (kg/pt) obtained from a blast furnace (furnace volume: 5000 ^m3 class), and the operating conditions are the same as those in FIG. 3. The horizontal axis shows the solution loss carbon amount (kg/pt), and the vertical axis shows the relative frequency. In blast furnace operation, the solution loss carbon amount is rarely an outlier, so the vertical axis is displayed in logarithm. In a blast furnace, gas blown in from the tuyere burns coke and pulverized coal to generate CO gas, and this generated CO gas reduces the iron raw material while rising in the furnace, generating CO ₂ gas. Then, a phenomenon occurs in which this generated CO ₂ gas reacts with the coke in the furnace and returns to CO gas again. This regenerated CO gas may be discharged outside the furnace without being used to reduce the iron raw material, and in this specification, the amount of carbon contained in this CO gas is defined as the solution loss carbon amount. The unit of solution loss carbon amount is "kg/tp" as described above, and it can be expressed as the amount of carbon per ton of pig iron production.

The outlier threshold can be seen from the figure when parameter c is changed from 2 to 6. In other words, the intersection of the dotted line extending parallel to the vertical axis and the horizontal axis is the outlier threshold for each parameter c.

If parameter c is set to 2, approximately 20% of the data will be classified as outliers, and the outlier threshold will be excessively large. Furthermore, if parameter c is set to 6, only about 0.3% of the data will be classified as outliers, and there is a possibility that significant outliers will be missed. Therefore, as in the case of ηCO, parameter c is preferably between 3 and 5, and it is basically considered to set parameter c to 4, which is the intermediate value of the "preferred range." As in the case of ηCO, parameter c can be changed within the "preferred range" depending on the purpose of the prediction, but this point will not be explained again.

In this way, the "preferred range" of parameter c is 3 to 5 even for different types of operation data, so the same value of parameter c can be set for all operation data. However, it is also possible to set different parameters c depending on the type of operation data, although this will complicate the process.

(Step S1-2: normal distribution processing)
As described above, many blast furnace operation data are not necessarily normally distributed but have a biased distribution. Many blast furnace operation prediction models assume that operation data used as training data for machine learning is normally distributed, and if non-normally distributed data is used, an appropriate model (a model that satisfies the desired prediction accuracy) cannot be constructed. Therefore, in this step S1-2, a process (corresponding to a conversion process and a conversion step) is performed to convert the non-normal distribution to a normal distribution. It is technically common knowledge that the prediction accuracy is improved by constructing a learning model using a normal distribution. It goes without saying that normal distribution processing is not required for operation data that follows a normal distribution.

Specifically, the data pre-processing unit 10 determines whether the absolute value of skewness of the acquired blast furnace operation data is greater than a threshold value (hereinafter also referred to as the skewness threshold value). If the absolute value of skewness exceeds the skewness threshold value, a process is performed to bring the operation data closer to a normal distribution (hereinafter also referred to as a normal distribution process). For the normal distribution process, for example, the "Box-Cox transformation" and the "Yeo-Johson transformation" can be used. The general formulas of the "Box-Cox transformation" and the "Yeo-Johson transformation" are shown in Equation (2) and Equation (3), respectively.

Fig. 7 shows the frequency distribution of the skewness of all operational data acquired over a certain period of time. As is clear from this frequency distribution, some operational data has a skewness close to 0 (i.e., the distribution is close to a normal distribution), while other operational data has an absolute value of skewness far from 0 (i.e., a non-normal distribution). Fig. 8 shows the frequency distribution of the average furnace top temperature (°C), with the horizontal axis being the average furnace top temperature (°C) and the vertical axis being the frequency. In the case of the average furnace top temperature (°C), the absolute value of the skewness is 0.01, which roughly follows a normal distribution, so there is no need to perform normal distribution processing.
9 shows the frequency distribution of the blast furnace heat load (kW), where the horizontal axis is the blast furnace heat load (kW) and the vertical axis is the frequency. In the case of the blast furnace heat load (kW), the absolute value of the skewness is 0.53, and the distribution is highly biased, so it is desirable to perform normal distribution processing.

The method for setting the "skewness threshold", which is the criterion for determining whether or not normal distribution processing should be performed, will be explained in detail with a concrete example. The relationship between the skewness threshold and the relative RMSE for the prediction results of the molten iron temperature is shown in Figure 10 while changing the skewness threshold. The general formula for RMSE is shown below.
Here, n is the total number of data points for the molten iron temperature, yi is the measured value of the molten iron temperature, and fi is the predicted value of the molten iron temperature.

As shown in the figure, when the skewness threshold reaches approximately 0.35 to 0.4, the relative RMSE increases rapidly. Therefore, for this data, the skewness threshold can be set appropriately in the range of 0.35 to 0.4 (for example, 0.38). The skewness threshold differs depending on the type of operational data, and is not uniquely determined.

(Step S2: Selection of explanatory variables)
By carrying out step S1, all the acquired operational data are standardized into normal distribution data, and a normal distribution data group consisting of a large number of normal distribution data can be acquired. As described above, this normal distribution data group also includes operational data that is determined not to require normal distribution processing (i.e., operational data that originally follows normal distribution).

In step S2, the data type selection unit 20 selects explanatory variables corresponding to the predetermined objective variables based on the data types of the individual operation data included in the blast furnace operation data (corresponding to a selection step and a selection process). Here, the data type refers to the category of each operation data, and is a concept different from the numerical value representing the operation data itself. Therefore, in this specification, both explanatory variables and objective variables refer to categories of operation data. For example, if the user sets a data type (category) of "pulverized coal ratio" as the objective variable, the data type selection unit 20 can select a data type (category) of "pulverized coal injection amount" as the explanatory variable corresponding to this "pulverized coal ratio". Here, when selecting an explanatory variable, a time-dependent factor can also be taken into consideration. In other words, operation data acquired at different times can be treated as different data types. For example, "pulverized coal injection amount from 0 o'clock to △ o'clock" and "pulverized coal injection amount from △ o'clock to x o'clock" are different data types and can be treated as independent explanatory variables, so they may be selected as explanatory variables at the same time.

When selecting explanatory variables, explanatory variables that are significant and independent in explaining the objective variable can be automatically selected (automatic selection method). For the automatic selection method, for example, the well-known Random Forest (RF), LASSO method, EN method, etc. can be used.
Random Forest (RF): A method for selecting variables based on the importance of variables in Random Forest (RF), and selects variables with high importance after running RF.
LASSO method: LASSO is an abbreviation for "Least Absolute Shrinkage and Selection Operator." With this method, since the standardized regression coefficients of variables that do not contribute to the model tend to be zero, only variables whose standardized regression coefficients are not zero are selected.
EN method: EN is an abbreviation for "Elastic Net", and like the LASSO method, the standard regression coefficients tend to be 0. Therefore, we select variables that are not 0.

(Step S3: Construction of blast furnace operation prediction model)
The learning model generation unit 30 generates a blast furnace operation prediction model (in other words, a neural network) by supervised machine learning using the operation data corresponding to the explanatory variables selected in step S2 and the operation data corresponding to the predetermined objective variables as training data. It goes without saying that these operation data are selected from the above-mentioned normal distribution data group. Machine learning is one of the artificial intelligence (AI) technologies that realizes prediction based on a predetermined data analysis method by inputting predetermined data. The data analysis method refers to an analysis algorithm for analyzing data. Therefore, a blast furnace operation prediction model is generated for each objective variable predetermined by the user, and a blast furnace operation prediction model using machine learning technology is generated according to an analysis algorithm corresponding to each objective variable. The generated blast furnace operation prediction model can be tuned as appropriate. The tuning method will be described below. The same type of operation data in a different time series from the operation data selected as the training data is defined as an input value and an actual value. Then, a predicted value based on the blast furnace operation prediction model is obtained from this input value, and if the difference between this predicted value and the actual value is equal to or greater than a predetermined error, the blast furnace prediction model is tuned. Tuning means, for example, adjusting a hyperparameter. The hyperparameters can be tuned using various methods published as open source, for example.

There are two possible methods for taking this time series into account. The first is to treat data from the past few hours (e.g., 8 hours) as a separate variable and select the optimal training data. The second is to introduce past time series information (e.g., information from 1 to 8 hours ago) as training data when building a neural network, and build the neural network as an RNN (Recurrent Neural Network) structure. Time series information refers to the time when the operational data was obtained. Time series can be taken into account using either of these two methods.

When constructing a neural network, overfitting may be avoided by the following method (known method). That is, if the weight coefficients of the neural network become more than the explanatory variables, overfitting occurs, so the weight coefficient net is deleted as appropriate. This avoids overfitting and reduces the calculation of the weight coefficients of the neural network. For example, overfitting can be avoided by applying dropout and structural learning with forgetting.

The coefficient net to be eliminated can be selected, for example, by using the bootstrap-LASSO method to select variables whose coefficients are frequently 0 and set their coupling coefficients to 0. This makes it possible to avoid overlearning and build a learning model more efficiently. It is also possible to adopt a method in which the coefficient net to be eliminated is selected randomly.

(Step S4: Building an action decision model)
By carrying out step S3, a blast furnace operation prediction model is constructed. By using this blast furnace operation prediction model, the operation state of the blast furnace (for example, change in molten pig iron temperature, etc.) can be predicted. By predicting the change in molten pig iron temperature, it is possible to create relationship information between the change in molten pig iron temperature and the change in pulverized coal injection amount, as shown in FIG. 11. The vertical axis of FIG. 11 represents the change in molten pig iron temperature, and the horizontal axis represents the change in pulverized coal injection amount. The change in pulverized coal injection amount can be derived from past operation results. An operator monitoring the operation of the blast furnace can determine an appropriate injection amount based on the relationship in FIG. 11 and operate the blast furnace.

As shown in FIG. 11, there is a generally positive correlation between the change in molten iron temperature and the change in the amount of pulverized coal injection, but it is difficult to uniquely determine an operational action for a change in molten iron temperature. Therefore, an optimal action determination model for optimizing the action of pulverized coal injection by variously changing the action for the change in molten iron temperature may be constructed by reinforcement learning, and the optimal action may be determined using the constructed optimal action determination model. The construction of the optimal action determination model may be performed by the learning model generation unit 30.

Here, reinforcement learning means executing actions A(s) according to a strategy n(s) at the time of state S, and determining a strategy so that reward R(s) is maximized. In blast furnace operation, state S is the current operating state, and includes, for example, an operating state in which the molten iron temperature is decreasing. It goes without saying that the molten iron temperature is constantly monitored in order to monitor state S. Action A(s) is an action taken in response to a change in the operating state, and corresponds to, for example, determining the amount of pulverized coal injection to be increased when the molten iron temperature is decreasing. The reward can be determined, for example, from the difference between the predicted molten iron temperature and the molten iron temperature that is the result after the actual action.
For example, when the molten iron temperature is predetermined as the objective variable, when the difference between the predicted value of the molten iron temperature and the result after the action is taken is set to 0, and when it is equivalent to 1σ of the distribution of the target variable, the reward is set to 0, for example, by linear interpolation, and if it is less than that, it can be set to a negative value.

After that, when the molten iron temperature deviates from the target value for a certain period of time, variously adjusted pulverized coal injection actions are tried, and then an action with a high expected reward (determination of the pulverized coal injection amount) is implemented. When constructing an optimal action decision model using reinforcement learning, it is better to update the model parameters more frequently, and it takes about six months. After that, by sequentially updating the parameters, the accuracy of the correspondence between the objective variable and the action can be improved, thereby improving the controllability of the objective variable. In addition, when there are multiple control factors and control objectives, a model can be constructed by treating the strategy (Policy), action (Actions) A(s), and reward (Reward) R(s) as tensors. It is desirable to store the operation data when executing step S4.

(Step S5: Reconstruction of prediction model)
The operation data acquired by operating the blast furnace using the optimal action determined based on the action decision model (step S4) has fewer outliers than the past operation data (operation data used in step S1). Therefore, steps S1 to S3 may be executed using such operation data to reconstruct the prediction model. In addition to reconstructing the prediction model, the constructed prediction model may be tuned by executing supervised learning again for the prediction model constructed in step S3 using the new operation data recorded in step S4 as teacher data.

The processes described in steps S1 to S5 can be realized by a program. That is, programs prepared in advance to realize the various processes are stored in the storage medium 40, and a computer such as a CPU executes the programs stored in the storage medium 40 to realize the processes.

The present invention will be specifically described with reference to examples. Comparative Examples 1 to 3 and Example 1 below were carried out to evaluate the relative RMSE of the predicted value of the molten iron temperature and the coefficient of variation of the coke ratio. The evaluation results are shown in Figures 12 and 13.

(Comparative Example 1)
Based on the conventional knowledge-based prediction of molten iron temperature, the operator estimated the optimum action (injection amount) and decided on the result. The comparison was based on a blast furnace with an internal volume of 5,000 ^m3 , with a reducing agent rate of approximately 500 kg/t, a coke rate of approximately 340 kg/t, and a productivity rate of approximately 2.0. In this operation, the blast furnace operator adjusted the amount of pulverized coal injection based on the work standard in accordance with the predicted value of the molten iron temperature indicated by the knowledge-based guidance system, thereby operating to bring the molten iron temperature closer to the target value.

(Comparative Example 2)
Step S1 was skipped (in other words, no outliers were removed and no normal distribution processing was performed), and the molten iron temperature was estimated using the prediction model constructed only in steps S2 and S3. The operator decided the optimal action (injection amount). Based on the conventional operation, the molten iron temperature was predicted using a prediction model based on a neural network, and the pulverized coal injection amount was adjusted based on the work standard according to the predicted value, thereby controlling the molten iron temperature to the target value.

(Comparative Example 3)
A prediction model was constructed by excluding the "normal distribution processing" of step S1 among steps S1 to S3. Based on the operation of Comparative Example 2 using the neural network prediction model, a model in which an outlier processing function was added to the neural network prediction model was used to predict the molten iron temperature, and the amount of pulverized coal injection was adjusted based on the work standard according to the predicted value, thereby controlling the molten iron temperature to the target value. However, when this model was implemented, the function of performing normal distribution processing of the distribution was not incorporated.

Example 1
A normal distribution processing function was added to the prediction model constructed in step S3. Based on the operation of Comparative Example 3 using the neural network prediction model, a model in which an outlier processing function and a normal distribution processing function were added to the neural network prediction model was used to predict the molten iron temperature, and the amount of pulverized coal injection was adjusted based on the work standard in accordance with the predicted value, thereby controlling the molten iron temperature to the target value. As shown in Figures 12 and 13, it was possible to reduce the relative RMSE of the predicted value of the molten iron temperature and to further reduce the coefficient of variation of the coke rate.

1 Machine learning device for blast furnace operation 2 Blast furnace 10 Data pre-processing unit 20 Data type selection unit 30 Learning model generation unit 40 Storage medium

Claims (6)

a data pre-processing unit that converts non-normally distributed data, the skewness of which exceeds a skewness threshold, into normally distributed data, thereby unifying the blast furnace operation data into normally distributed data;
a data type selection unit that selects explanatory variables corresponding to a predetermined objective variable based on a data type of each piece of operation data included in the operation data of the blast furnace;
a model generation unit that generates a blast furnace operation prediction model by performing machine learning using the blast furnace operation data corresponding to the selected explanatory variables and the blast furnace operation data corresponding to the objective variables as training data;
having
The data pre-processing unit performs a removal process to remove outliers contained in each operational data before converting the non-normally distributed data into normally distributed data;
The data pre-processing unit performs the removal process in accordance with a Hampel Identifier method based on the following equation (1) .
median(xi)±cMAD(xi)...Formula (1)
Here, MAD(xi) = median (|xi - median(xi)|), median is the median value, and c is a parameter selected from the range of 3 to 5 (however, not limited to integers). 2. The machine learning device for blast furnace operation according to claim 1, wherein, when the blast furnace operation data includes the same type of operation data that are acquired at different timings, the data type selection unit selects the explanatory variables by treating the data types of the operation data as different data types. Predicting the operational state of a blast furnace using the machine learning device for blast furnace operation according to claim 1 or 2 ;
A method for operating a blast furnace, comprising the steps of: executing an action based on the prediction result to operate the blast furnace. a data pre-processing unit that converts non-normally distributed data having a skewness exceeding a skewness threshold into normally distributed data among post-action operation data of a blast furnace obtained after the action described in claim 3 is performed, and unifies the post-action operation data of the blast furnace into normally distributed data;
A data type selection unit that selects an explanatory variable corresponding to a predetermined objective variable based on a data type of each post-action operation data included in the post -action operation data of the blast furnace;
a model generation unit that generates a blast furnace operation prediction model by performing machine learning using as training data the post- action operation data of the blast furnace corresponding to the selected explanatory variables and the post-action operation data of the blast furnace corresponding to the objective variable;
A machine learning device for blast furnace operation comprising: A conversion step of unifying the blast furnace operation data into normal distribution data by converting non-normal distribution data, the skewness of which exceeds a skewness threshold, into normal distribution data among the blast furnace operation data;
A data type selection step of selecting an explanatory variable corresponding to a predetermined objective variable based on a data type of each piece of operation data included in the operation data of the blast furnace;
A model generation step of generating a blast furnace operation prediction model by machine learning using the blast furnace operation data corresponding to the explanatory variables selected in the data type selection step and the blast furnace operation data corresponding to the objective variables as training data;
Run the following on your computer :
In the conversion step, before converting the non-normally distributed data into normally distributed data, a removal step is performed to remove outliers contained in each operational data;
A program used in a machine learning device for blast furnace operation, characterized in that in the conversion step, the removal step is performed according to the Hampel Identifier method based on the following equation (1) .
median(xi)±cMAD(xi)...Formula (1)
Here, MAD(xi) = median (|xi - median(xi)|), median is the median value, and c is a parameter selected from the range of 3 to 5 (however, not limited to integers). A conversion step of converting non-normally distributed data of the blast furnace operation data, the skewness of which exceeds a skewness threshold, into normally distributed data, thereby unifying the blast furnace operation data into normally distributed data;
A data type selection step of selecting an explanatory variable corresponding to a predetermined objective variable based on a data type of each piece of operation data included in the operation data of the blast furnace;
A model generation process of generating a blast furnace operation prediction model by machine learning using the blast furnace operation data corresponding to the explanatory variables selected in the data type selection process and the blast furnace operation data corresponding to the objective variables as teacher data;
having
In the conversion step, before converting the non-normally distributed data into normally distributed data, a removal step is performed to remove outliers contained in each operational data;
A method for generating a blast furnace operation prediction model, characterized in that in the conversion step, the removal step is performed according to a Hampel Identifier method based on the following equation (1) .
median(xi)±cMAD(xi)...Formula (1)
Here, MAD(xi) = median (|xi - median(xi)|), median is the median value, and c is a parameter selected from the range of 3 to 5 (however, not limited to integers).

Images