JPH0673414A - Method for controlling quality of molten iron in blast furnace - Google Patents

Abstract

PURPOSE:To provide a quality control method of molten iron in a blast furnace so as to obtain the molten iron having a prescribed quality in the blast furnace by quantitatively estimating the molten iron temp. and component values of Si and S, etc., in the iron as indexes of the quality of the molten iron in the blast furnace and by adjusting operational quantity and various kinds of elemental values. CONSTITUTION:In this controlling device, an inputting layer for inputting various kinds of data values, sensor values and operational quantities in the blast furnace operation and an outputting layer for outputting the component values of Si and S in the iron are provided. The various kinds of elemental values, sensor values and operational quantities in the blast furnace operation are inputted to the inputting layer of the prelearnt two or more layers of molten iron quality neural net work. By changing the operating quantity so as to come in the permissible range when the eliminated value in the molten iron quality obtd. from the outputting layer is out of the permissible range of the aimed control value, the estimated value of the molten iron quality is obtd. When the molten iron quality does not come in the permissible range even in the case of changing the operating quantity, the various kinds of elemental values of the blast furnace operation is changed. When the estimated value in the molten iron quality comes in the permissible range, based on the various kinds of data values and operational quantity in the blast furnace operation, the operation is executed.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a quality control method for blast furnace hot metal, and in particular, it utilizes a neural network method to obtain past blast furnace operation data such as specifications, sensor values, manipulated variables, and hot metal quality values. Study and quality control method of blast furnace hot metal using it.

[0002]

2. Description of the Related Art Blast furnace operation is intended for complex reactions in which a large number of operation factors are alternately related to each other, and it is difficult to directly visually monitor from equipment conditions, etc. In order to achieve this, it is necessary to comprehensively judge the information from the sensors attached to the blast furnace and operate it appropriately. Therefore, the experience and knowledge of operators are still important for the daily operation management of blast furnaces. However, there is a problem in that it is difficult to standardize the operation method and quantify the evaluation due to individual differences in the operation based on human judgment, and it is difficult to always continue appropriate actions.

Attempts have been made to apply statistical processing and mathematical models to quantitatively make human judgments. As a method for predicting furnace heat using ARMA, which is an autoregressive model using various sensor information, Japanese Patent Laid-Open No. 63-210213 discloses.
JP-A-63-210221 and JP-A-63-213
There are methods proposed in Japanese Patent Laid-Open No. 608 to 63-213611 and Japanese Patent Laid-Open No. 63-266010.

Since a knowledge engineering system can process the experience and knowledge of a skilled person by working on a computer, it is disclosed in JP-A-62-27.
No. 0708, No. 62-270712, No. 1-205010, No. 1-205011.
Japanese Patent Laid-Open No. 2-77508, Japanese Patent Laid-Open No. 2-30
By systematizing blast furnace operation management as disclosed in Japanese Patent No. 1504, problems such as oversight of information and erroneous judgment are eliminated, and the operation management is optimized and standardized.

There is also a proposal to which a neural network imitating human nerve cells is applied. JP-A-2-170
In 904, a method is proposed in which a hierarchical neural network method is applied to predict the transition of the blast furnace furnace heat level from the present to the future and utilize it for furnace heat control. This furnace heat control system has an input layer that allows blast furnace operation data to be input to each unit, and from the output layer, how much probability the furnace heat will rise in the future and whether it will be maintained or not will decrease. It predicts and outputs the transition of the furnace heat level.

[0006]

Problems to be Solved by the Invention JP-A-63-2102
No. 13-Japanese Unexamined Patent Publication No. 63-210221, No. 63-213608 No. 63-213611.
In the blast furnace heat prediction method disclosed in Japanese Patent Laid-Open No. 63-266010 and Japanese Patent Laid-Open No. 63-266010, various sensor information and the like are predicted by ARMA which is a mathematical / statistical model method.
In order to continue to maintain the function of the mathematical model, it is necessary to keep the internal parameters optimally controlled, which imposes a heavy load on system maintenance and requires a great deal of manpower to make full use of it. Moreover, only the function of predicting the furnace heat of the blast furnace is taken into consideration and the quality of hot metal is not taken into consideration.

Further, in Japanese Patent Laid-Open No. 2-170904, a hierarchical neural network method is used to predict the transition of the blast furnace heat level from the present to the future and utilize it for furnace heat control. Although the transition of the furnace heat level can be predicted, the hot metal temperature value that is an index of the furnace heat cannot be quantitatively determined, and the component values such as Si and S in the pig iron that are other indexes of the hot metal quality cannot be predicted.

The present invention was developed in order to solve the following problems of the conventional furnace heat control method based on the furnace heat level transition prediction by the statistical processing, expert system and neural network. . (1) Although the trend of the furnace heat level can be predicted from various observation data, it is not possible to quantitatively predict the hot metal temperature and the component values such as Si and S in the pig iron, which are indicators of product quality. (2) Comprehensively judge the production situation or reaction situation in the furnace such as the operation amount at the top of the furnace such as coke ratio and basicity, the operation amount at the lower part of blower specifications and the amount of tapping, which have a significant effect on the quality of hot metal. The hot metal quality has not been decided.

[0009]

The blast furnace hot metal quality control method of the present invention quantitatively predicts and manages the hot metal temperature, which is an index of the quality of the blast furnace hot metal, and the component values such as Si and S in the hot metal. However, the following processing is performed for that purpose. (1) A hierarchy of two or more layers of pre-learned hot metal quality, which has an input layer that allows input of factors that have a significant effect on hot metal quality to each unit, and an output layer that outputs hot metal quality. Have a network ready. (2) In the system made based on the above-mentioned hot metal quality neural network, at least operating condition parameters such as coke ratio, basicity and blast parameters that affect hot metal quality, and tapping amount. And, the sensor value indicating the inside of the furnace is read at a predetermined timing. (3) Hot metal temperature and Si or S in the hot metal, which is an index of hot metal quality obtained from the output layer by inputting operation information to the input layer
The following operations are performed using the component values such as. a) If the hot metal quality prediction value is within the allowable range of the target control value, the current operation amount is maintained. b) If the predicted hot metal quality value is outside the allowable range of the target control value, the optimum manipulated variable is determined by changing the manipulated variable of the input layer so that the predicted hot metal quality value falls within the allowable range of the target controlled value. ,
The operation is changed according to the obtained optimum operation amount. Also,
If the hot metal quality prediction does not fall within the allowable range due to the change in the operation amount, change the operation specifications. (4) Due to changes in operating conditions such as changes in the raw material brands used and the amount of blast furnace production, the hot metal quality neural network cannot always be used as it was initially trained. In addition, since the condition of bricks in the furnace changes as the blast furnace equipment continues to operate, differences in the quality of hot metal appear even if the same operation is performed. Therefore, if the deviation of the predicted value of hot metal quality from the actual value obtained in operation is outside the allowable range, the actual value of hot metal quality is input into the input layer with the parameter values, sensor values and manipulated variables of the blast furnace operation at that time. It is used for the output layer, and re-learning by self-organization is appropriately performed to continue maintaining the prediction accuracy of the hot metal quality neural network. This re-learning is also performed in a fixed cycle.

[0010]

In the present invention, various factors affecting the quality of various types of blast furnace hot metal are input to the input layer of the hot metal quality neural network, and the quality of blast furnace hot metal is output to the output layer qualitatively and quantitatively for prediction. If the hot metal quality predicted value exceeds the allowable range of the target control value, the hot metal quality is controlled to be constant by changing the operation amount. If the target control value does not fall within the permissible range due to the change in the operation amount, change the operation specifications. Further, the mutual coupling coefficient of the hot metal quality neural network is automatically generated or modified in order to maintain and improve the prediction accuracy.

[0011]

1 is a block diagram of a blast furnace hot metal quality control system for carrying out a blast furnace hot metal quality control method according to an embodiment of the present invention. The hot metal quality influence factor 1 is operation data based on the operation operation amount, the sensor value, the operation amount, and the like. This is taken into the system through the data input means 2, and the delay time correction means 3 performs correction corresponding to the delay time at which the effect on the hot metal quality appears. For example, changing the coke ratio or basicity, which is the operation amount of the furnace top, requires a delay correction of 6 to 10 hours from the raw material descent time, and changing the blast specifications of the operation amount of the furnace bottom requires changing the tuyere and taphole. It is necessary to correct the delay of 1 to 4 hours due to the positional relationship. The data normalizing means 4 normalizes 0 to 1 so as to suit the input value of the neural network. These data are passed to the hot metal quality neural network 5, and the neural network 5 outputs a hot metal quality predicted value 6.

The hot metal quality prediction value 6 is (a) hot metal temperature,
It is composed of three types of information: (b) Si component value in pig iron, and (c) S component value in pig iron. Since each item of (a) to (c) is output as a normalized number between 0 and 1 so that it can be handled quantitatively, it is converted to a concrete hot metal quality prediction value and instructed to the operator. . The comparator 8 confirms whether the hot metal quality predicted value 6 is within the allowable range of the target control value 7. If the hot metal quality predicted value 6 is within the allowable range, the current operation is continued, but if the hot metal quality predicted value 6 is outside the allowable range, the change amount selecting means is set so that the hot metal quality predicted value 6 falls within the allowable range of the target control value 7. The operation amount and the like among the quality influencing factors are pseudo-changed by 9 to determine the optimum change operation amount and the like.

In addition to the operation using the above-mentioned hot metal quality prediction value, reference input data 10 and teacher output 11 are given to the system for learning of the hot metal quality neural network. Reference input data 10 is blast furnace operation data, and quality influence factor 1
And the teacher output 11 needs to be accurate information as an output for the reference input data 10. The daily operation result becomes a teacher, and the data of the teacher output 11 is given. Learning pattern creation processing means 1
The reference numeral 2 simply converts the reference input data 10 and the teacher output 11 into a pattern that the hot metal quality neural network 5 can easily learn. The learning pattern creation processing means 12 is designed to create the hot metal quality predictive value when the predicted hot metal quality does not match the actual value in the daily operation results.

The blast furnace hot metal quality control system having the above construction will be described in more detail. Blast furnace is 1000-500
Since the furnace volume is as wide as 0 m ³ and the characteristics are different in each furnace, the quality influence factor 1 that affects the data input means 2 in FIG.
You can also consider more than one. Hereinafter, as an example, a static model that does not include the time concept and a dynamic model that includes the time concept will be described. Similar expectations can be achieved by using items other than the quality affecting factors described below. First, the static model will be described. Items of quality influence factor 1 given to the data input means 2 of FIG. 1 are shown below. (1) Target amount of tapping (2) Coke ratio (3) Basicity of charging raw material (4) Air temperature (5) Air flow rate (6) Air pressure (7) Air humidity (8) Oxygenation rate (9) Feather Mouth embedding temperature average value (10) Tuyere embedding temperature deviation value

If the volume of the blast furnace increases and the number of tap holes and tuyere increases, the accuracy can be improved by adding the following items. (11) Top gas utilization rate (12) Raw material unloading rate (13) Solution loss (14) Tap hole number Next, the dynamic model will be described. Items of quality influence factor 1 given to the data input means 2 of FIG. 1 are shown below. (1) Target amount of tapping (2) Coke ratio (3) Basicity of charging raw material (4) Air temperature (5) Air flow rate (6) Air pressure (7) Air humidity (8) Oxygenation rate (9) Feather Mouth filling temperature average value (10) Top gas utilization rate (11) Raw material unloading rate (12) Solution loss (13) Tap hole number (14) Tap time after tap hole opening (15) Previous hot metal Temperature (16) Previous Si value (17) Previous S value

FIG. 2 shows a static example of the structure of the hot metal quality neural network 5 of the hot metal quality learning model.
It is the figure shown by the model, and FIG. 3 is the figure which showed the structural example by the dynamic (dynamic) model. The hot metal quality neural network 5 has a three-layer structure in which one hidden layer is provided between the input layer for receiving the processed data of FIG. 1 and the output layer for the change operation amount. The hot metal quality neural network 5 is linear in a two-layer structure, but is non-linear in three or more layers, and the information processing means is remarkably improved. However, if the number of layers increases, it will take a long time for learning, and if the number of layers is 5 or more, the accuracy of the solution cannot be expected to improve. Therefore, in general, a neural network of 3 layers with an intermediate layer (hidden layer) is used. . In this embodiment, the output layer has three units, and one unit is associated with each of the above-mentioned three information of "molten pig temperature", "pigment Si component value", and "pigment S component value".

FIG. 4 is a diagram showing an example in which the output of the output layer is viewed as time series data, and the output of the hot metal temperature unit is plotted. 0 as described above
It has a value of ˜1, and the closer the value is to 1, the larger the hot metal temperature value. By controlling the operation with the hot metal quality predicted value converted from this value, the distance from the target control value of the current blast furnace operating state can be grasped, and if the allowable range is exceeded, the operation amount etc. are changed so that it falls.

FIG. 5 is a flow chart showing the operation of the entire system. Normal hot metal quality prediction is started at a fixed period (for example, 5 minutes), various data of the quality influence factor 1 is input via the data input means 2 as described above, and the delay time correction means 3 delays the data for a predetermined time and then the data. The normalizing means 4 processes and normalizes the data. Next, the data is input to the input layer of the hot metal quality neural network 5. The input data propagates in the network in the forward direction, and the hot metal quality prediction value 6 is output from the output layer. Comparator 8
Thus, it is determined whether or not the hot metal quality predicted value 6 is within the allowable range of the target control value 7, and if it is outside the allowable range, the manipulated variable is changed. Then, the above-described processing is repeated, and if the deviation is still not within the allowable range, the operating specification values are changed. In this way, by repeating the above-mentioned processing, the changed operation amount or operation specification value at the time when the hot metal quality predicted value 6 falls within the allowable range is adopted, and the operation operation amount or operation specification is adopted. Operate based on the value.

When changing the operation amount, the priorities are set in advance as follows, for example, and are changed according to the order. The basicity and the oxygen addition rate are fixed. (1) Blast humidity (2) Charge coke ratio (3) Blast temperature (4) Target amount of tapping (5) Blast flow rate (6) Blast pressure

The learning process is also started at a fixed period (for example, every day), teacher data is created, and the reference input data 7 and the teacher output 8 are input to the input layer of the hot metal quality neural network 5 via the learning pattern creating means 9. It is read, propagated in the forward direction, corrects the mutual coupling coefficient of the hot metal quality neural network 5, and is repeated until the deviation becomes equal to or less than the allowable value. Table 1 shows the data normalization means 4 of the static model in the blast furnace with the furnace volume of 4000 m ³ and the range of the operation amount. Hereinafter, the embodiments of FIGS. 6 and 7 have the same volume.

[0021]

[Table 1]

FIGS. 6 and 7 are graphs showing the predicted hot metal quality predictive value trend and the hot metal quality data trend in the actual operation according to the above embodiment. When the predicted value of the hot metal temperature after the time T1 + 3 hours reaches the lower limit value of the allowable range by the above-described calculation processing at time T1 in FIG. 6, the blast humidity action (decrease blast humidity) is taken here as shown in FIG. The predicted value of the hot metal temperature is output as, and if it is within the allowable range, the action is taken. Then, when viewed at time T2 (= T1 + 6 hours), time T2-3
It can be seen that the actual value of the hot metal temperature at time (= T1 + 3 hours) is within the allowable range. This is because the above blast humidity action is effective after about 3 hours. The reaction time (delay time) differs depending on the hot metal volume depending on the distance between the tuyere and tap hole and the diameter of the furnace bottom, and it is about 3 hours in the above-mentioned embodiment. When taking an operation action in this way, the reaction time is taken into consideration. This also applies when changing the specifications of the operation. In this case, the reaction time is about 6 to 10 hours.

[0023]

As described above, according to the present invention, if the hot metal quality predicted value is outside the allowable range of the target control value, the hot metal quality predicted value is within the allowable range of the target control value. Since the optimum operation amount is determined by inputting the operation amount and the specifications of operation in a pseudo manner, and the predicted time is determined by various delay times introduced in the preprocessing of the input value, Appropriate quality prediction values were obtained when the influencing factors changed, and the quality control of blast furnace hot metal became appropriate.
Further, since the re-learning is appropriately performed by using the learning method of the hot metal quality neural network and the function of the system is continuously maintained, it is possible to flexibly respond to changes in the state of equipment and the operating state.

[Brief description of drawings]

FIG. 1 is a block diagram of a system for implementing the method of one embodiment of the present invention.

FIG. 2 is a diagram showing a configuration example of a hot metal quality neural network of a hot metal quality learning model as a static model.

FIG. 3 is a diagram showing a configuration example of a hot metal quality neural network 5 of a hot metal quality learning model by a dynamic model.

FIG. 4 is a diagram showing an example of output of an output layer of a hot metal quality neural network.

FIG. 5 is a flowchart showing the operation of the embodiment.

FIG. 6 is a diagram showing a predicted value trend of hot metal quality and a trend of hot metal quality data in actual operation according to the embodiment.

FIG. 7 is a diagram showing a predicted value trend of hot metal quality and a trend of hot metal quality data in actual operation according to the embodiment.

 ─────────────────────────────────────────────────── ─── Continued Front Page (72) Inventor Yoshiro Nishi Marunouchi 1-2-2, Chiyoda-ku, Tokyo Nihon Steel Pipe Co., Ltd. (72) Inventor Masaki Takenaka 1-2-1, Marunouchi, Chiyoda-ku, Tokyo Nippon Steel Tube Co., Ltd. (72) Inventor Katsushi Okuda 1-2 1-2 Marunouchi, Chiyoda-ku, Tokyo Nihon Steel Tube Co., Ltd.

Claims (4)

[Claims] 1. An input layer for inputting specifications, sensor values and manipulated variables of blast furnace operation, and hot metal temperature as an index of hot metal quality,
A layered neural network having two or more layers, which has an output layer for outputting the component values of Si and S in the pig iron and is trained in advance (hereinafter referred to as a hot metal quality neural network), is prepared. Input the specifications, sensor values and manipulated variables of the blast furnace operation currently in operation to the input layer of the above, and change the manipulated variables so that the hot metal quality prediction value obtained from the output layer is within the allowable range. The hot metal quality neural network is used to calculate the hot metal quality predicted value, and when the hot metal quality predicted value is within the allowable range, it is based on the blast furnace operation specifications and the manipulated value input to the input layer of the hot metal quality neural network. Blast furnace hot metal quality control method for operating and controlling hot metal quality. 2. When the predicted hot metal quality value is outside the allowable range even if the manipulated variable is changed, the specifications of the blast furnace operation are changed so that the predicted hot metal quality value falls within the allowable range. The quality control method for blast furnace hot metal according to claim 1, wherein the hot metal quality is determined by a neural network. 3. The quality control method for hot metal of blast furnace according to claim 1, wherein the self-organization of the hot metal quality neural network is re-learned at a predetermined cycle. 4. If the hot metal quality prediction value output from the output layer of the hot metal quality neural network and the actual quality value obtained in actual operation differ by more than a permissible deviation, the self-organization of the hot metal quality neural network is performed. The quality control method for blast furnace hot metal according to claim 1, 2 or 3, wherein the method is relearned.