KR950014631B1 - Apparatus of temperature pre-estinate and action control guide with molten metal - Google Patents

Abstract

The predicting equipment for the molten iron temperature improve the blast furnace stability and the standardization of operation control by using error back propagation of the multi-layer perceptron in artificial neural networks. The equipment for predicting molten iron temperature comprises: a process computer for supporting blast furnace operation; an artificial intelligent computer for predicting the temperature; and an instrumentation controller for managing digital signals and electric controller to manage valve on-off signal and switch on-off signal.

Description

Prediction and action control amount guide device for blast furnace chart using artificial neural network

Figure 1 is a graph showing the transition type of the conventional molten iron temperature and the operation factor.

2 is a graph of a drop, stability, rise, and boundary on a conventional trend determination reference value.

3 is a block diagram of a molten iron temperature prediction and action control amount guide system according to the present invention.

4 is a structural diagram of a neural network applied to the present invention.

5 is a graph showing the type of the molten iron temperature and the time series data of the operating factors according to the present invention.

6 is a control flowchart of the apparatus according to the present invention.

* Explanation of symbols for main parts of the drawings

1: blast furnace 2: SCC (sub computer)

3: process computer 4: artificial intelligence computer

5 sensor data collection unit 6 calculation unit

7: data storage unit 8: blast furnace control unit

9: Interface Buffer 10: Inference Engine

11: Knowledge Base 14: Instrument Controller

15: electric controller 17: hot stove

18: Cocos supply tank 21: pulverized coal supply tank

The present invention relates to a blast furnace molten iron temperature prediction and action control amount guide device using an artificial neural network to improve the predicted hit ratio of the molten iron temperature that is outgoing during blast furnace operation.

In the blast furnace operation, the charter temperature information is an important operation result that represents the internal situation of the furnace. In daily operation, the charter temperature is used as the information for the operator to directly estimate the situation in the furnace.

In order to maintain the molten iron temperature at a constant level (1510 ± 10 ℃), the operator changes the molten iron temperature and the molten iron temperature related to the molten iron temperature such as the top temperature, gas utilization rate, lower furnace body heat dissipation index, lower air permeability fluctuation index and charge drop rate. In order to maintain the future molten iron temperature level properly, the action control amount such as moisture, coke ratio and pulverized coal is determined and blown into the blast furnace.

Conventionally, the method of estimating the molten iron temperature in the blast furnace is generally performed by the blast furnace operator qualitatively determining the data provided at regular intervals from various sensors installed in the blast furnace, and evaluating the condition of the blast furnace. This was accomplished by quantitatively blowing molten iron temperature control actions such as coke bee into the blast furnace. However, the situation evaluation of the blast furnace involved individual differences according to the abilities and experiences of the operators, and the standardization of the control action was complicated. For this reason, a charter temperature prediction and control amount guide system was proposed and constructed to standardize and standardize the knowledge of human knowledge (Know-How), and to infer and operate like an expert. However, conventional linear coefficients in calculating short-term trends and charter temperature trends of furnace temperature, gas utilization rate, lower furnace body heat dissipation index, lower notification fluctuation index, and reinforcement rate (hereinafter abbreviated as operation factors) are essential for blast furnace prediction. The use of equations has led to problems in the accurate determination of these trends. Inaccurate trends are not only constraints for future charter temperature predictions, but also significant limitations in charter temperature prediction and action control quantity guide systems, where the charter temperature trends quantitatively determine the thermal action at which future charter temperature levels change. It acted as a main factor. This will be described in more detail as follows.

FIG. 1 shows four cases of trends in operating factors and charter temperatures.

A: D1> D2> D36 → Degradation B: D1> D2 <D32 → Stable

C: D1 <D2> D35 → Stable D: D1 <D2 <D31 → Up

Indicates.

This trend can be expressed as follows.

ΔDi: Change in operating factors over a period of time (D31-D1)

Δti: time difference (t3-t1)

Again, this trend variable is determined by a constant reference value of each operating factor and the charter temperature. It is composed of the following syntax.

If Ti> αi, TCi = 1 ........................................... .............................(2)

If Di ≤ Ti≤α then TCi = 1 ... (3)

TCi = -1 if Ti <βi ...........................(4)

αi, βi: integers that determine the trend

TCi: Integer indicating trend (2: much up, 1: up, 0: stable, -1: down, -2: down much)

However, as shown in equation (1) above, the expression represented by the transition variable is

Ti = f (△ Di, ΔTi) ....................... (5)

Since it is only a linearized two-variable function expressed by, it is judged to be stable (B: D1-D2-D32) or decrease (B: D1-D2-D33) even though the current trend is increasing in the case of B and C. Although the current trend is decreasing, there is a contradiction that is determined to be stable (C: D-D2-D35) or increase (C: D1-D2-D34). In addition, the criteria for determining the trend are divided into integers of αi and βi, as shown in Equations (2) (3) (4) above, so that the transition categories in the transition boundary (αi, βi) are βi1 and βi2 as shown in FIG. Even when the difference between αi1 and αi2 is minute, it is classified as falling, stable, stable, and rising, and thus there is a problem in the transition on the boundary. This results in a significant reduction in the temperature of the blast furnace, which in turn results in a lowering of the blast furnace temperature prediction and control quantity guide system that quantitatively presents the thermal action to the operator.

By using the error back propagation learning function of the multi-layer perceptron structure among various structures of the neural network, which is a new field of artificial intelligence of the present invention, each operator and The purpose of the present invention is to provide a blast furnace melting temperature prediction and action control amount guide device using an artificial neural network which can derive the accurate change of the melting temperature.

Hereinafter, the present invention will be described based on the accompanying drawings.

3 shows an embodiment of the present invention, a process computer 3 having a blast furnace operation support function, an artificial intelligence computer 4 dedicated to inference, an instrument controller 14 and a valve for processing digital signals. And an electric controller 15 for processing the open / close signal and the switch on / off signal.

The process computer 3 is installed in the blast furnace 1, the sensor data collection unit 5 for collecting the sensing data from the sensors (2a-2d) for sensing the blast furnace data, and the charge based on the collected sensor data Computation unit 6 for calculating various kinds of data (processing data) indicating blast furnace blasting, such as descent speed, gas utilization rate, lower furnace radiant heat, association of lower breathability fluctuation index, and measured molten iron temperature, and these sensor data and processing A blast furnace control unit 8 and an artificial intelligence computer 4 for controlling the blast furnace 1 through a data storage unit 7 for storing data in a storage location and an instrumentation and electric controller 14 and 15 as lower-level computers; An interface buffer 9 temporarily stores data for data transmission and reception.

The processor computer 3 is configured to receive data from the SCC 2, which is a lower level computer that collects and processes the molten iron temperature sensing signal of the molten iron temperature sensor 2a. The artificial intelligence computer 4 inferred by the artificial intelligence technique has an interface buffer 9 for receiving and transmitting data to and from the process computer 3, and various knowledge based on the experience and performance of the blast furnace operation for molten iron temperature management. The knowledge base 11 stored therein and the inference engine 10 which infers from the time series data stored in the interface buffer and transmits the result to the process computer 3.

When the control amount by the operator 13 is input through the operating terminal 12, the control amount data is configured to be provided to the process computer 3, and the control amount data is passed through the blast furnace control section 8 to the instrumentation controller 14. And to the electrical controller 15.

The instrumentation controller 14 is configured to control the moisture in the input air of the hot stove 17 for heating and supplying air to the blast furnace 1. The electric controller 1 is configured to control the weighter 19 of the coke supply tank 18 and also to control the pulverized coal supply tank 21.

20 is a conveyor belt which carries coke to the blast furnace 1, 23 is a nitrogen compressor for nitrogen pressure supply of a pulverized coal tank, and 22 is a pulverized coal control valve | bulb.

The operation of the present invention configured as described above will be described with reference to FIGS. 3 to 6.

The molten iron temperature information measured by using a shop-type thermocouple is transmitted to the processor computer 3 via the lower level computer 2 collecting various sensor data.

The process computer 3 corrects the delivered molten iron temperature information by the molten iron temperature correction method using the characteristic curve of the outlet opening, and collects various data of the sensor data collection unit 5 collected at a predetermined interval from the blast furnace sensors 2a-2d. The data is transmitted to the artificial intelligence computer 4 together with the processing data calculated from these sensor data. In the transmitted data, the top temperature, gas utilization rate, lower body heat dissipation, lower air permeability fluctuation index, load drop rate, and charter temperature by the blast furnace sensor (2a-2d) were respectively fixed period (1 hour 36 minutes, 4 hours). The trend is then determined using neural network with time series data.

Figure 4 shows the structure of the multilayer perceptron neural network constructed in the present invention in order to determine the short-term trend and the molten iron temperature trend. The neural network applied to the present invention is an input layer, and a plurality of nodes (20 nodes for a molten iron temperature neural network, an operation factor, and 8 nodes for a neural network) and one biased connection node receive data at intervals of 12 minutes. The data output from the plurality of nodes constituting the input layer includes the input layer and the middle layer at the plurality of nodes (6 nodes for the molten iron temperature neural network and 4 nodes for the operation factor neural network). Weights vary depending on the strength of the connection between them and are input to the intermediate layer nodes. Data output from the middle layer node is weighted by the connection strength between the middle layer and the output layer, and is input to a plurality of nodes constituting the output layer (two for the molten iron temperature neural network, one for the operation factor and one for the neural network). . The data output to the output layer indicates the trend of the data input to the input layer and is compared with the learning objective to calculate the error. The calculated error is applied to change the connection strength between the input layer and the middle layer, the middle layer and the output layer. As the weight changes, the input data is learned and the trend is output.

The neural network as described above is prepared for the molten iron temperature and five operating factors (expansion temperature, gas utilization rate, lower furnace body heat dissipation, lower aeration, charge rate), and six neural networks are provided.

As shown in Fig. 5, not only the data at time t1 and t3 but also the time series data between t1 and t3 and the additional data of bias connection node are included. Gas utilization rate, bottom furnace heat dissipation heat, bottom ventilation, charging speed) Short-term trend and charter temperature trend. In other words, for example, the input data of the molten iron temperature trend determination neural network includes 21 pieces of molten iron temperature data and 12 bias temperature nodes for 12 minutes in the past 4 hours from the present, and the output data are input and intermediate layers in the input data. And the weight is determined by the connection strength of the intermediate layer and the output layer is a molten iron temperature trend based on the molten iron temperature trend determination criteria determined by the operator in accordance with previously stored learning.

Short-term trend of decision factor The input data of the neural network is 9 including 8 factor data and 1 bias connector node at 12-minute intervals from the present for the past 1 hour and 36 minutes. The output data is determined by the operator according to the input data. Short-term trend of operation factor This is the trend of each operation factor based on the criterion.

The goal of the neural network is to reflect the influence of additional time series data for a certain period of time in the calculation of each operator factor and the charter temperature trend that is performed in the charter temperature forecast and the action guide system's forecasting cycle (every 12 minutes).

When the input data 31 representing the current situation up to now is applied to the neural network, the output 34 is obtained based on the current weight 32 and 33, and the output and the molten iron temperature and the trend of each operator are calculated. (35) is compared to detect an error, and then the weights 32 and 33 between the layers are corrected in the direction of reducing the error. This process is performed repeatedly until the error falls below a small enough value to learn the nonlinear correlation that exists between all time series input data and the trend output.

The process of blast furnace melting temperature and operation factor data by neural network is as follows. For example, in the case of a molten iron temperature of blast furnace temperature from 1470 ℃ in the past 228 minutes to the current temperature of 1520 ℃, the opinion of the blast furnace expert judges that the molten iron temperature is rising.

In the case of neural network, it is used as input data of the molten iron temperature from the past (228 minutes ago) to the present, and as the learning target (output data) of the neural network, it is used to change the input molten iron temperature data. We study nonlinear correlation of the rise (stable, decrease) of output according to the above. Charter temperature trend The learning neural network is shown in FIG. In describing the learning process, if the neural network output value does not coincide with the learning target value with respect to the input melting temperature of the neural network, the neural network output value is adjusted in the direction of the learning target by modifying the weighting values 32 and 33.

The learning ends when the output value of the neural network falls within a certain error (error that may be ignored) from the learning target value. At the end of the study, the nonlinear relationship between the molten iron temperature trend (rising, stable, and declining) is studied as the molten iron temperature changes. The operation factor data is processed by the neural network. The operation factor is the five factors of the top temperature, gas utilization rate, bottom furnace heat dissipation, bottom ventilation, charging rate, and the short term trend of the operation factor is the same as the above chart. It is carried out based on the principle that the temperature from the past 96 minutes ago to the present is the input of the neural network. Learning is conducted to learn the nonlinear relationship between the top and bottom temperature trends. Operational Factor Trend Learning Neural Network consists of 8 nodes that receive input data and 1 bias connect code, the middle layer consists of 4 nodes, and the output layer consists of 10 nodes. Same as

The learned knowledge 32, 33 represents the correlation between time series input data and output data trends. When calculating the temperature of molten iron and the trend of each operator from the completed neural network, determine the trend on the basis of the input data by receiving the molten iron temperature of the current inference cycle and time series data of each operator from the field sensor and applying them as inputs. The neural network calculates the molten iron temperature and each operating factor, thereby guiding the heat prediction and action control.

The characteristic of this neural network is that the data hunting due to sensor error has an error correction function, so that accurate trend determination is possible, and it shows high accuracy (99%) compared to the trend determination by numerical calculation and the trend determination result (TCi) is chartered. As the temperature ranges from 2 (high increase) to -2 (low decrease) and 1 (increase) to -1 (decrease) for each operator, it is a transition boundary. It is possible to prevent the error of dividing the trend into even small data values in.

As shown in FIG. 4, the reason why the neural network for the derivation of trends is configured for each factor is that the data characteristics of each factor are different and thus the degree of transition must be changed accordingly. The neural network inputs the determination criteria of the molten iron temperature and the operating factor at the time of operation at the fishing terminal and inputs the neural network learning whether or not the neural network is learned in order to obtain the change of the molten iron temperature and the operating factor according to the aging of the blast furnace. Learns again in the above-described way to derive new learning results and to use them in trend determination.

On the other hand, the short-term and periodic trends of each operation factor determined using the artificial neural network predict the change in the molten iron temperature (△ To) after 2 hours, and the past action control amount such as the coke ratio moisture and pulverized coal. By predicting the change in molten iron temperature (ΔTa) in the next two hours according to the change, it is predicted the total change in the molten iron temperature expected in the future, and the total change in the molten iron temperature in the next two hours.

This is expressed as follows.

Tp = Tc + △ To + △ Ta

Tp: Melting temperature after 2 hours

Tc: Current molten iron temperature

△ To: Estimated charter temperature change by operating factor trend combination

ΔTa: The amount of change in predicted molten iron temperature caused by the change of the action control amount.

The forecasted change in the molten iron temperature (△ To + △ Ta), together with the molten iron temperature determined by the neural network, calculates the molten iron temperature from the current molten iron temperature level, the current fuel cost deviation from the previous day, and the future latent heat level. The total amount of action control to be calculated) is determined, and this is again determined as the action amount of the moisture, pulverized coal, and coke bar, and the estimated molten iron temperature level and the action control amount obtained above are guided to the operator through the operation terminal. Accordingly, when the operator views the operation state of the action facility and sets the action control amount in the operation terminal, the instrument controller 14 and the electric controller 15 inject the appropriate amount of coke and heating air into the blast furnace 1 according to the set value. do.

As described above, according to the present invention, the machining data is calculated by receiving data from various sensors installed in the blast furnace, and the trend is calculated from the machining data (including the operating factors) and the molten iron temperature data. It is fully utilized to predict future molten iron temperature and guide the action control amount, so that accurate molten iron temperature prediction, standardization of operation management, human misjudgment prevention and aging stabilization can be achieved.

Claims (4)

A plurality of sensors for sensing the operating factor data in the blast furnace, a process computer for receiving and processing the operating factor data detected from the sensor, and receiving the processing data from the process computer to infer the trend of the operating factors An artificial intelligence computer for outputting the results to the process computer, an instrumentation controller for controlling moisture in the hot blast furnace receiving the control signal from the process computer and supplying the hot air to the blast furnace, and receiving the control signal from the process computer in the blast furnace In the blast furnace control device comprising a coke supply tank for supplying coke and pulverized coal and an electric controller for controlling the pulverized coal supply tank, the artificial intelligence computer is a plurality of nodes that sequentially receive time series molten iron temperature data input at predetermined time intervals With bias connection nodes An intermediate layer consisting of a plurality of nodes receiving data output by applying a predetermined weight from the input layer node and an output layer consisting of a plurality of nodes receiving data output by applying a predetermined weight from the intermediate layer node. And a neural network for correcting the weight by an error between the output data of the output layer and the learning target for each of the molten iron temperature and the operation factor. Device. The blast furnace molten iron temperature prediction and action control amount guide device according to claim 1, wherein the predetermined time interval is 12 minutes. The blast furnace melting temperature prediction and action control amount guide according to claim 1, wherein the molten iron temperature neural network comprises an input layer consisting of 20 input nodes and a bias connection node, six intermediate layers, and two output layers. Device. 2. The blast furnace melting temperature prediction and action control amount guide according to claim 1, wherein the operator factor neural network is composed of an input layer consisting of eight input nodes and a bias connect node, four intermediate layers, and one output layer. Device.