JPH05247467A - Control unit for combustion in coke oven - Google Patents

Abstract

PURPOSE:To provide the title unit whereby the fluctuation of coke grade can be reduced by learning the past actual results of operation by using a neural network, forecasting the coke grade from them and optimally operating the coke oven. CONSTITUTION:The title unit comprises a means for preparing a prelearned at least bilayered neural network 5 having an input layer for inputting data 1 before the change of operation of a coke oven and data 2 after this change and an output layer for outputting a controlled variable 6 in the forward direction of the coking process, a means for inputting the data before the change and the data after the change into the input layer and a means for controlling the coking process by using the controlled variable obtained from the output layer of the neural network as a result of inputting and a means for learning the self-learning structure of the neural network.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to combustion control of a coke oven for the purpose of stabilizing the operation by learning the operation result and predicting the action means when the operation is changed by using a neural network. It relates to the device.

[0002]

2. Description of the Related Art Coke quality represented by strength and grain size of coke is qualitatively evaluated by sensor information from various sensors installed by an operator in a coke oven to estimate and evaluate the carbonization state of coal and its transition. However, it has been managed and controlled by optimal adjustment of operating factors such as the amount of heat supplied (calorie and flow rate of fuel gas) and the amount of combustion air. However, there are individual differences in combustion control based on human judgment, it is difficult to standardize control methods and quantify evaluations, and it is difficult to always maintain appropriate actions.

Also, some control devices applying mathematical models and knowledge engineering have been reported in patent publications and "Materials and Processes" issued by the Iron and Steel Institute. For example, Japanese Patent Publication 57-5383
No. 2 gazette discloses "a combustion control method in a coke oven". This publication describes the total GDR (Gross Coking Time) of the target GCT of a furnace group (collection of individual kilns is referred to as a furnace group) Net Coking Tim is the time from the charging of the coal into the coke oven to the end of the carbonization of the coal.
e), the target flue temperature of the furnace and individual kilns is determined and measured from the actual charging amount and water content of each kiln (combination of carbonization chamber and combustion chamber (called flue) is called kiln) The actual flue temperatures of individual kilns based on the values, the average value of the furnace group and the actual NCT are compared periodically, and based on the difference, the heat quantity supplied to the furnace group (fuel gas (Calorie and flow rate) and combustion air amount are calculated and automatically controlled, and the adjustment amount of the valve opening of the gas valve and air valve necessary for adjusting the heat amount and air amount for each kiln is calculated and the operator guidance is given. ing. However, this method attempts to automatically control the "originally ambiguous information" regarding coke oven operation by forcibly expressing it in a mathematical formula and automatically controlling it.
Dividing may occur at some stages, which may cause problems in control accuracy.

Further, Japanese Unexamined Patent Publication No. 3-121190 discloses a "combustion control method for a coke oven" in which fuzzy control, which is one of the artificial intelligence application techniques, is introduced into the combustion control for the coke oven. This publication is intended to improve the control accuracy by solving the problem that the conventional combustion control by the mathematical model involves division at the stage of "expressing ambiguous information precisely by mathematical expression". "Ambiguity"
Is represented by a membership function. However, it is difficult to accurately represent knowledge about coke oven operation with a membership function. Moreover, the control accuracy cannot be maintained unless the membership function is optimized in accordance with changes in the operation and the state of equipment. Furthermore, there is a problem that it is difficult to optimize the membership function and learning for the optimization.

[0005]

As described above, the combustion control device for the coke oven using the conventional mathematical model and fuzzy control has the following problems. (1) A combustion control device based on a mathematical model attempts to automatically control the "originally ambiguous information" relating to coke oven operation by precisely expressing it in a mathematical formula. It is divided at the stage of "expressing", which becomes a problem in control accuracy. (2) It is difficult for a combustion control device using fuzzy control to accurately represent knowledge about coke oven operation with a membership function. Moreover, the control accuracy cannot be maintained unless the membership function is optimized in accordance with changes in the operation and the state of equipment. Furthermore, it is difficult to optimize the membership function and learning for optimization.

The present invention has been made in order to solve the above problems and optimize the combustion control of a coke oven, and learns the past operation results by using a neural network and the learning results thereof. The purpose is to develop a control device that predicts the transition of coke quality by using, performs optimal combustion control, and leads to stable and improved coke quality.

[0007]

[Means for Solving the Problems] A combustion control device for a coke oven according to the present invention is intended to predict the transition of coke quality, perform optimal combustion control, and stabilize and improve coke quality. 1) It has an input layer that can input the data before the operation change and the data after the operation change to each unit, and also has an output layer that outputs the control amount, and learn it based on the operation data and the sensor data in advance. Means for preparing a multilayer neural network of two or more layers in the forward direction of the coal carbonization process of the coke process (hereinafter referred to as forward device), and (2) operation change to the input layer of the "forward device" A means for inputting the previous data (current value) and the data after operation change (target value), and (3) the calorific value of the fuel gas by a mathematical model using the controlled variable obtained from the output layer of the "forward device". And flow rate The adjustment amount of the combustion air amount is calculated and means for automatically controlling the combustion control status of the coke oven and is formed by three or more means. Furthermore, the "forward device"
Self-organization of the neural network when the actual operation data obtained as a result of automatically controlling the control amount and the combustion control situation output from the output layer of Is learned to improve the control accuracy of the apparatus.

[0008]

According to the present invention, the coke quality is stabilized and improved by automatically controlling the combustion of the coke oven by a neural network and a mathematical model constructed in multiple layers on the basis of data and control amounts before and after various operation changes. Try to. Further, the learning accuracy of the neural network can be maintained by automatically generating or modifying the mutual coefficient of the neural network.

[0009]

EXAMPLES Next, examples will be described. FIG. 1 is a block diagram illustrating an input / output and processing procedure of a coke oven combustion control device according to an embodiment of the present invention. In FIG. 1, an oval frame indicates an input and an output of the control device, and a rectangular block indicates an internal process of the control device. 1 is a data value before operation change, 2 is a data value after operation change, 3 is a data input means, 4 is a data processing means, 5 is a neural network (forward device), 6 is a control amount (heat supply amount), 7 Is a mathematical model, 8 is a combustion control device for calorie and flow rate of fuel gas, and combustion air amount, 9 is reference input data, 10 is teacher input data, and 11 is a learning pattern creation process.

In FIG. 1, data 1 before operation change is coke operation data based on data measured by various sensors, for example, GCT and NCT as operation targets, and coal charging determined when charging coal into a kiln. Specifications (charge amount, water content, volatile matter), elapsed time after charging as operating conditions, flue temperature, carbonization time of coal, calorie and flow rate of fuel gas as control amount and amount of combustion air is there. The data 2 before the operation change is the target value of the coke operation data after the control amount change, and is the coke strength and grain size. These are taken into the forward direction device through the data input means 3, and the data necessary for judging the control amount of the combustion control after removing noise and the like by the smoothing processing or the linear regression processing by the data processing means 4, for example, flue. It is processed into information that includes temporal factors such as the difference between the target value and the actual value of the temperature, the amount of coal charged and the water content, and the amount of change over a certain period from the past to the present.

These data are passed to the neural network 5, and the neural network 5 outputs the controlled variable 6 by the forward device learned in the past. The control amount 6 is passed to the mathematical model 7, which determines the calorie and flow rate of the fuel gas and the combustion air amount, and outputs it to the combustion control device 8 to automatically control the calorie and flow rate of the fuel gas and the combustion air amount.

Separately from this, reference input data 9 and teacher input data 10 for learning the neural network are given to the control device. In this case, the reference input data 9 is formally the same as the coke oven operation data 1, and the teacher input data 10 is formally the same as the control amount 6. However, the teacher input data 10 corresponds to the reference input data 9. Exact information is required as output. Therefore, the daily operation result in which the operation condition is stable becomes a teacher, and the teacher input data 10 is given.

In the learning pattern creating process 11, the reference input data 9 and the teacher input data 10 are input to the neural network 5.
It is simply converted into a pattern that is easy to learn in.

The combustion control device having the above-mentioned structure will be described in more detail. First, the coke oven operation data 1 given to the data input means 3 of FIG. (1) Operation target GCT (2) Operation target NCT (3) Coal charge (4) Coal water content (5) Coal volatiles (6) Flue temperature (7) Fuel gas calories (8) Fuel Gas flow rate (9) Combustion air amount (10) Coal carbonization time (11) Elapsed time after charging (12) Coke quality

FIG. 2 shows an example of the configuration of the neural network 5 (forward device) in this embodiment.
The configuration of the neural network 5 (forward device) in this embodiment is an input layer for fetching various data from the above-mentioned (1) operation target GCT to (12) coke quality processed by the data processing means 4 of FIG. This is a three-layer structure in which an intermediate layer is placed between the output layers of the amount of heat supplied as the control amount 6. Generally, the neural network 5 is linear in the two-layer structure, but becomes non-linear in the three or more layers as in the present embodiment, and the information processing means is remarkably improved. Therefore, in many cases, a network having three or more layers with an intermediate layer is used. In this embodiment, the output layer is one unit of the heat supply amount. Next, the output value of the output layer of the neural network 5, and the result of calculating the calorie of fuel gas and the adjustment amount of the combustion air amount by a mathematical model based on this output value are displayed as a time series transition diagram. The thing is shown in FIG.
In FIG. 3, at 0, the supplied heat amount, which is the output value of the neural network, was in a normal (standard) state, but after that, the output value of decreasing was obtained. Heat supply (gas calorie)
It is proportional to x (gas flow rate). In this device, the gas flow rate is controlled first, and when it reaches the upper and lower limit values, the gas calorie is changed. Is output, and at the same time, the control amount of the air flow rate for combustion is output according to the control amount of the gas flow rate. At 3 o'clock, the amount of heat supplied was considerably higher than the reference value, and the control amount of the gas flow rate was caught at the upper and lower limit values, so the gas calorie was changed in steps and the gas flow rate was lowered, and thereafter, the gas flow rate again. It shows a state in which control by means of is possible.

In the present apparatus, the control amount during normal operation was changed at a constant period of 30 minutes because the time constant of the coke oven was long at 4 to 5 hours. Further, the learning of the learning structure of the neural network may be a one-day cycle, a one-month cycle, or the like, but the one-day cycle is used for more accurate learning control.

FIG. 4 shows changes in coke quality and changes in energy saving. As shown in Fig. 4, when this system is operated based on the operation result data for the past four months, the fluctuation of coke strength is the same as the conventional one (the conventional variation is 1.0).
About 10% reduction. As a result of the decrease in the fluctuation of coke strength, the amount of steam coal used increased by about 3% and the energy saving effect was about 5% kcal / kg.

[0018]

As described above, according to the present invention, it is possible to secure a controlled variable for optimal coke oven operation, and to perform re-learning using a neural network learning method at any time, thereby facilitating equipment and operation. Even if the situation changes, the accuracy of this controller can be maintained. As a result, the quality of the coke was stabilized and improved, and the coke required by the blast furnace could be produced stably at low cost.

[Brief description of drawings]

FIG. 1 is a block diagram of a combustion control device for a coke oven according to an embodiment of the present invention.

FIG. 2 is an explanatory diagram showing a configuration of a neural network in the present embodiment.

FIG. 3 is an explanatory diagram showing an output example of an output layer of a neural network as time series data.

FIG. 4 is a transition diagram of coke quality and energy saving in this example.

[Explanation of symbols]

1 data value before operation change 2 data value after operation change 3 data input means 4 data processing means 5 neural network (forward device) 6 control amount (heat supply amount) 7 mathematical model 8 for fuel gas calorie and flow rate and combustion Air amount combustion control device 9 Reference input data 10 Teacher input data 11 Learning pattern creation processing

Claims (3)

[Claims] 1. An apparatus for controlling the state of coal carbonization in a coke oven, which has an input layer for inputting data before and after operation change and an output layer for outputting a control amount, and has an operation data and a sensor in advance. A means for preparing a hierarchical neural network of two or more layers learned based on data in the forward direction of the coal carbonization process of a coke oven, and controlling the coal carbonization condition by using a control amount obtained from the output layer of the neural network. A combustion control device for a coke oven, comprising: 2. The coke oven according to claim 1, wherein the calorie and flow rate of the fuel gas and the combustion air amount are determined by a mathematical model based on the control amount output from the output layer of the neural network. Combustion control device. 3. The actual operation data obtained as a result of automatically controlling the carbonization condition of coal by the calorie and flow rate of the fuel gas and the amount of combustion air determined by the mathematical model before the operation change input into the input layer and The coke oven combustion control device according to claim 2, wherein when the data is different from the data after the operation change, the self-organization of the neural network is learned to improve control accuracy.