JPH06240318A - Blast furnace charge distribution control method - Google Patents

Abstract

(57) [Summary] [Purpose] A method for controlling the distribution of blast furnace charges that enables the adverse effects on operations due to erroneous determination of manipulated variables to be minimized by monitoring the changes over time in the blast furnace gas flow state. [Configuration] An input layer for inputting operational specification values, sensor values and manipulated variables, and an output layer for outputting furnace port sonde temperature and gas utilization rate, and two or more layers learned in advance Type neural network is prepared in advance, and the operating specification values, sensor values and manipulated variables during operation are input to the input layer of this neural network, and the furnace port sonde temperature and gas utilization rate obtained from the output layer are input. The manipulated variable in which the predicted value falls within the allowable range with respect to the target control value is pseudo-changed, and the distribution of blast furnace charge is controlled based on the manipulated variable when the predicted value falls within the allowable range.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a blast furnace charging for recognizing a gas flow in a furnace by using a temperature of a furnace opening sonde at the top of a blast furnace and a gas utilization rate and thereby determining a distribution control manipulated variable of the charge. The present invention relates to a method for controlling distribution of objects, particularly a control method using a neural network.

[0002]

2. Description of the Related Art A control method of this kind is disclosed in, for example, Japanese Unexamined Patent Publication No.
There is one disclosed in Japanese Patent No. 90504. This publication discloses, as a method for controlling the distribution of the blast furnace charge using a neural network, a method of determining a bellless point which is a distributor based on the surface profile of the furnace top raw material charge in the bellless blast furnace. .

[0003]

[Patent Document 1] Japanese Patent Application Laid-Open No. 3-905
In the blast furnace charging control device disclosed in Japanese Patent Publication No. 04, the optimum raw material charging method is automatically determined by measuring the raw material profile before adjustment and providing the adjusted raw material profile. However, in the blast furnace operation, even if the raw material profile at the top of the furnace descends inside the furnace, it may interfere with uniform unloading due to the inside of the furnace, etc. Is changing. For this reason, even if an optimal furnace top profile is obtained, it is not enough by itself, and it is necessary to capture the changes accompanying unloading. By the way, the purpose of controlling the distribution of the raw material charge is to control the gas flow in the furnace.
The control method proposed in Japanese Patent No. 4 does not describe the function of gas flow monitoring.
Further, in the blast furnace process for producing pig iron, since rationalization effect due to the use of inferior raw material is expected, the use of small grain sintering and the like is carried out by trial and error, but the measures are also described in the above publication. Absent.

The present invention has been made in order to solve such problems, and by monitoring the change over time in the gas flow state in the blast furnace, the adverse effect on the operation due to the erroneous determination of the operation amount can be minimized. It is an object of the present invention to provide a method for controlling the distribution of blast furnace charges that can be stopped.

[0005]

In the method for controlling the distribution of blast furnace charge according to the present invention, the manipulated variable is determined as follows. (1) Operating specification values, distribution device state that is the distribution control operation amount of the charge, specification value before change of the charge raw material mixture ratio and its change amount, furnace port sonde temperature that recognizes the gas flow, inside the furnace Two or more layered neural network in which the gas utilization rate at which the reaction result appears and the elapsed time after changing the manipulated variable are taken into the input layer, and the changed amounts of the furnace port sonde temperature and gas utilization rate after the change are output layers To build. (2) A neural network that is pre-learned to have a squared error ε <0.01 by back propagation using the result of operation performed by an operator in the actual operation on the neural network is a learned neuron. (3) When the furnace sonde temperature or the gas utilization rate deviates from the target state in the actual operation, for example, the value obtained from the output layer as the predicted result after 3 hours is (output value) = (target value)
The change operation item of the input layer is changed in a predetermined order so as to be − (current value), and the optimum operation amount is determined. (4) When the operation amount obtained by the above method is applied to the actual operation, the change with time every 30 minutes is compared with the simulation result of the learning neuro, and the effect of the change result is confirmed. (5) In the above comparison, when the simulation result and the actual operation change amount deviate from the permissible range, a different operation amount change is immediately performed to prevent the deterioration of the blast furnace state due to the determination of the wrong operation change amount and the application thereof. Further, the erroneous operation amount is determined to be a change in the input / output value used for learning, and relearning is performed.

[0006]

In the present invention, the state of the gas flow in the furnace is recognized from the furnace port sonde temperature and the gas utilization rate, and if they are out of the target values, the optimum change amount of the distributor and the raw material mixture rate is set to the past. It is calculated from the learned neural network. Further, when the change operation amount is determined and applied, the change state of the furnace port sonde temperature pattern and the gas utilization rate every 30 minutes, for example, is compared with a neuro output simulation that has been learned in the past. As a result, the influence of incorrect determination of the manipulated variable can be minimized.

[0007]

FIG. 1 is a block diagram showing the configuration of a system for carrying out the method for controlling the distribution of blast furnace charge according to an embodiment of the present invention. The operation influence factor 1 is operation data based on various sensor values of the blast furnace, operation specification values, operation amounts, and the like.
The sensor at the top of the furnace used here detects the temperature pattern of the furnace port sonde and the gas utilization rate. In addition, when the furnace mouth sonde is used in the circumferential direction, 4 points in the north, south, east, west, and 5 points from the center to the wall side are used, and when used in the radial direction, the center in the north, south, east, and west is averaged. 5 points from to the wall side shall be used. Further, since the gas temperature at the top of the blast furnace fluctuates at the time of charging the raw materials, it is necessary to average the temperatures of one cycle of coke and ore before use. There are two kinds of manipulated variables when controlling the distribution of the charged raw materials: the state quantity of the distributor and the mixing ratio of the charged raw materials. The distributors are the bell type and the bellless type.
There is a method. The mixing ratio of the charging raw materials is determined by the mixing ratio of coke ratio, sintering, yard sintering, B lump, small grain sinter, MIX coke, central coke, and medium lump coke.

The operation affecting factor 1 is taken into the system through the data input means 2 and the preprocessing means 3 carries out a predetermined preprocessing described later. The data subjected to the pre-processing is normalized by the normalizing means 3. The normalized data is passed to the neural network 5, which predicts and outputs the furnace port sonde temperature change amount, the gas utilization rate, and the like. The comparator 8 confirms whether these predicted values 6 are within the allowable range of the target control value 7. If the predicted value 6 is within the permissible range, the current operation is continued to be maintained, but if the predicted value 6 is outside the permissible range, the operating amount changing means 7 operates so that the predicted value 6 falls within the permissible range of the target control value 7. Is pseudo-changed to determine the optimum change operation amount and the like.

By the way, there are two kinds of manipulated variables for the distribution control of the blast furnace charge distribution, that is, the distribution device and the raw material mixture ratio. Therefore, if the distribution device and the raw material mixture ratio are put in the input layer of the neural network 5, the number of input layers becomes large. Therefore, in the present embodiment, each is separated and a separate neural network is constructed in the neural network 5. FIG. 2 is an explanatory diagram showing a neural network for changing the distribution device, and FIG. 3 is an explanatory diagram showing a neural network for changing the raw material mixture ratio. The air flow parameters such as temperature, flow rate, and pressure are applied to each neural network in the input layer. In addition, there are three methods of changing the manipulated variable, namely, a case where the two manipulated variables of the distribution device and the blending ratio are simultaneously changed, and a case where each of them is independently changed. For this reason, the above two types of neural networks are provided with a discrimination flag for determining whether another neural network is changing the operation amount.

The pretreatment means 3 corrects the time of the effect of air blowing characteristics and treats the disturbance at the time of charging the raw material in the furnace with respect to the temperature of the furnace port sonde. As a process for time correction of blast specification effects, time delay is taken into consideration because the effect of blast blown from the tuyere at the bottom of the furnace does not immediately appear in the gas utilization rate at the top of the furnace or the temperature pattern of the sonde sonde. , For example, the data after 3 hours is used. In addition, as a measure against the disturbance at the time of charging the raw material in the furnace with respect to the temperature of the furnace port sonde, when the raw material on the large bell is charged into the furnace, the heat at the furnace top is temporarily removed and the temperature drops. The temperature is determined by taking the average of one cycle of charging the raw materials.

In the data normalizing means 3, since each item of the input / output layer of the neural network has a different dimension for each engineering value, the normalization of 0 to 1 is performed. Tables 1 and 2 show the normalized ranges of the input layers of the neural networks of FIGS.

[0012]

[Table 1]

[0013]

[Table 2]

In the system of FIG. 1, the reference input data 10 and the teacher output 11 are given to the neural network 5 for learning the neural network 5 in addition to the operation using the above-mentioned predicted value. The reference input data 10 is formally the same as the operation influencing factor 1 which is the blast furnace operation data, and the teacher output 11 needs to be accurate information as an output to the reference input data 10. Therefore, the daily operation result becomes a teacher, and the data of the teacher output 11 is given. That is, the reference input data 10
Is the manipulated variable of the charging material distribution control that is actually changed when the furnace port sonde temperature and the gas utilization rate in the furnace deviate from the target values. In addition, when this learning is performed, the furnace top gas flow has no immediate effect even if the manipulated variable is changed, and it takes, for example, 3 hours to be clearly confirmed. Therefore, 3 from the time of change
The time-lapse state after the change is learned at intervals of 30 minutes until the time. The learning pattern processing means 12 uses the reference input data 10
And the teacher output 11 is simply converted into a pattern that the neural network 5 can easily learn. It should be noted that this learning pattern processing means 12 does not only operate periodically, but also the predicted value 6 out of the daily operation results.
Is no longer in agreement with the actual value, it is always operated to correct the mutual coupling coefficient of the neural network 5.

In the learning of the neural networks shown in FIGS. 2 and 3, among the operation results in the actual operation, the changed operation amount is used as an input layer, the changed furnace port sonde temperature pattern and the gas utilization rate are used as teacher data 11, and output. The output data obtained from the layer is learned so that the square error ε with the teacher data is ε <0.01.

In the system of FIG. 1 configured as above, the change operation amount is determined as follows.
The neural network 5 is preliminarily learned by the above-described method, and the operation influencing factor 1 including the operation specification value, the furnace port sonde temperature pattern, the sensor value of the gas utilization rate, and the operation amount is used as the data input means 2 and the preprocessing. It is taken into the neural network 5 through the means 3 and the data normalizing means 4 to monitor the gas flow state of the blast furnace throat. When the comparator 8 compares the predicted value 6 (furnace port temperature and gas utilization rate) of the neural network 5 with the target control value 7 and determines that the value is not within the allowable range, the manipulated variable changing means 9 changes the manipulated variable. Is pseudo-applied via the data input means 2 and is repeated until the predicted value 6 falls within the allowable range of the target control value 7. In this way, a simulation in which the manipulated variable is changed is performed, and when the simulation result 6 falls within the allowable range with respect to the target control value 7, the simulation manipulated variable is determined as the actual manipulated variable. To do.

When the change manipulated variable obtained by the above method is used, the effect does not immediately appear, but the change over time after the change is simulated every 30 minutes until 0 to 3 hours, and the simulation result and the actual value are obtained. When the change process with time is in agreement with the learned result, the correct change operation amount is obtained. FIG. 5 is an explanatory diagram of a comparative example of the current value of the sensor information and the target management value. Here, it is assumed that there is a difference between the current value of the sensor information and the target management value, and the above-mentioned simulation starts, and FIG. 6 is a characteristic diagram showing the result of the simulation from the state of FIG. If the change history over time after changing the manipulated variable does not match the simulation result, it is determined that the changed manipulated variable has been erroneously determined, and the changed manipulated variable is re-determined and re-changed to minimize the adverse effect on operation due to the erroneous determination. Stop at.

[0018]

As described above, according to the present invention, when the gas flow state in the blast furnace is quantitatively measured using the furnace port sonde temperature pattern and the gas utilization rate, and when there is a deviation from the target state, Since the optimum manipulated variable that matches the target value is determined, optimum control is possible. Further, after applying the determined optimum operation amount to the operation, it is possible to minimize the adverse effect on the operation due to an erroneous decision by monitoring the change over time in the gas flow state in the blast furnace.

[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 an explanatory diagram of an embodiment of a distribution device changing neural network.

FIG. 3 is an explanatory diagram of an embodiment of a charging raw material changing neural network.

FIG. 4 is an explanatory diagram of a comparative example of the current value and target value of sensor information.

FIG. 5 is a timing chart showing a change over time when the determined change operation amount is performed.

Claims (1)

[Claims] 1. A hierarchical type having two or more layers learned in advance, having an input layer for inputting operational specifications, sensor values and manipulated variables, and an output layer for outputting furnace port sonde temperature and gas utilization rate. This neural network is prepared in advance, and the operational specification values, sensor values and manipulated variables during operation are input to the input layer of this neural network, and the furnace port sonde temperature and gas utilization rate obtained from the output layer are predicted. The manipulated variable is pseudo-changed so that the value falls within the allowable range with respect to the target control value, and the distribution of the blast furnace charge is controlled based on the manipulated variable when the predicted value falls within the allowable range. A method for controlling the distribution of blast furnace charges, which is characterized by the above.