JPH0637648B2 - Blast furnace furnace heat prediction method - Google Patents

Description

TECHNICAL FIELD The present invention relates to a blast furnace heat prediction method, and more particularly to a blast furnace heat control and a blast furnace operation method including the same.

[Conventional technology]

The method of estimating the temperature inside the blast furnace, which is represented by the hot metal temperature, and managing and controlling it is the conventional method in which a blast furnace operator qualitatively determines information from various sensors installed in the blast furnace to determine the blast furnace temperature. The method is to estimate and evaluate the level and its transition, and to adjust the operation factor optimally.

However, there are individual differences in the evaluation and estimation results due to the ability and experience of operators, and it is difficult to standardize the operation method, and since the evaluation is not quantitative, appropriate actions were not always taken. It was Of these, the prediction of future changes in the furnace heat level is due in large part to the experience and judgment of operators, beyond the recognition of the current furnace heat level, and there is the problem that it is more difficult to accurately predict and take action. It was

In order to solve such a problem, JP-A-62-270
In 708, a furnace heat control system is proposed which applies a knowledge engineering technique to infer the furnace heat level and the furnace heat transition to determine the action amount for the blast furnace.

The transition of the furnace heat here is the level of the judgment rule based on the transition of the molten iron temperature from the past to the present, the tuyere embedded temperature, the unloading pressure, the blast pressure, the gas utilization rate, the solution loss amount, the furnace top temperature, etc. Judgment rules are based on the five-stage transition evaluation level, which is obtained comprehensively from the judgment rules based on the above, and the judgment rules based on the transition of reactor conditions, and the respective confidence factors.

Changes in the furnace heat level are caused and observed by many factors as described above, and the problem is solved by expressing the rule of thumb of the conventional operator in the production rule.

However, the causal relationship between the above-mentioned various data and the furnace heat transition is not sufficiently understood even by a skilled operator. Therefore, the accuracy of the above-mentioned judgment rules and certainty factors set for each is not sufficient, and there is a risk that the action of the furnace heat control is taken at an inappropriate timing or the action amount is not appropriate. It was Further, in order to maintain the accuracy of predicting the furnace heat transition, it is necessary to constantly revise the judgment rules and certainty factors and tune them, and the maintenance load on the system was high.

Also, JP-A-60-248804 and JP-B-60-187.
21, JP-A-60-204813, JP-A-61-254.
853 and the like are methods of predicting furnace heat by a mathematical model such as statistical estimation, exponential averaging, and frequency analysis.
These methods depend on the precision of the mathematical model, and it is difficult to assemble the model formula accurately. The maintenance load for maintaining the accuracy also increases.

[Problems to be Solved by the Invention] The present invention was developed in order to solve the following problems with the method for predicting the transition of the furnace heat by the production rule and the mathematical model in the conventional expert system as described above. Is.

(1) The causal relationship between various observation data and changes in reactor heat level is unclear, and it is difficult to assemble a formulated logic. Therefore, the prediction accuracy is insufficient.

(2) Maintenance load is high for maintaining system performance. Parameter tuning takes time and the parameters may change.

[Means for Solving the Problems]

In the present invention, the method of predicting the transition of the blast furnace heat level from the present time to the future is as follows: (a) It has an input layer that allows the blast furnace operation data to be input to each unit, and the probability of the furnace heat Prepare a pre-learned layered neural network of two or more layers, which has an output layer that outputs the predicted result of the transition of the furnace heat level whether it rises in the future, maintains the current state, or decreases. deep.

(B) Made based on various sensors installed in the blast furnace,
It is necessary to predict the furnace heat, and various data including at least the unloading status of the blast furnace, the gas composition status, the current furnace heat level, the furnace body temperature and pressure loss status, and the past action history data are provided at a predetermined timing. Read with.

(C) The data is processed into various data representing what kind of change each of these data has made in the last few hours and what kind of history has passed. For example, the amount of change in η _{CO or} the amount of change in N ₂ is calculated as a data indicating the state of the gaseous component by a calculation process from the measured gas analysis value.

(D) Input the processed data to the input layer of the neural network. Inside the network, each unit in each layer inputs the links from all units, sums them, multiplies the result by an internal function to obtain one output, outputs it to all units in the next layer, and finally outputs At least one unit provided in the output layer quantitatively outputs whether the furnace heat level is going to increase in the future, is maintained as it is, or is going down.

(E) Further, in learning the neural network, the transition record of the furnace heat level is quantified by the recent change amount of the measured value representative of the furnace heat such as the hot metal temperature, and when the change amount is large in the positive direction. Determines that the furnace heat has increased, and when it is large in the negative direction, it is determined that the furnace heat has decreased,
By creating teaching data of the network prediction result (output of the neural network) and learning the network together with the reference input data and the actual prediction result output at that time, the mutual coupling coefficient of the neural network is automatically generated or modified.

Here, the neural network will be briefly described with reference to pages 115 to 124 of Nikkei Electronics (1987.8.10).

A neural network is a network that imitates the human brain. Multiple units corresponding to brain neurons are connected intricately. The pattern recognition function and the knowledge processing function can be embedded by properly determining the operation of each unit and the connection form between the units.

The structure of the unit is a model of a neuron, which is shown schematically in FIG. The structure is simple and consists of a part that receives input from other units, a part that converts the input according to a certain rule, and a part that outputs the result.

A variable weight W is provided for each of the other units and the connecting portion.
_{Add ij} . This is to show the strength of the bond. Changing this value changes the network structure. Learning the network means changing this value. W takes positive, zero, and negative values. Zero means no bond.

When a unit receives input from multiple units, the total sum is used as the input value. The sum of input net is Is.

Each unit applies the total sum net of inputs to the function f and converts it. This input / output characteristic may differ for each unit. However, generally, the function shown in FIG. 4 (b) or (c) is often used. FIG. 4 (c) is called a sigmoid function. Is a differentiable, pseudo-linear function, Can be expressed as The area is 0 to 1, and approaches 1 as the input value increases and approaches 0 as the input value decreases. When the input is 0, it becomes 0.5. Add the value θ to the threshold, In some cases.

The network has a hierarchical structure as shown in FIG. An intermediate layer is provided between the input layer and the output layer. This intermediate layer is called a relaxation layer. There are neural nets with other structures, but here we limit ourselves to hierarchical networks that can use the pack propagation algorithm.

The networks are connected in the direction of the input layer, the middle layer and the output layer. There is no account within each layer. There is also no coupling from the output layer to the input layer. It is a feedforward or positive network.

Although it has a structure similar to that of the Perceptron, the capacity is significantly improved by providing an intermediate layer and making the input / output characteristics of each unit non-linear.

The network is trained as follows. Input data is given to each unit of the input layer. This signal is converted in each unit, propagates to the intermediate layer, and finally emerges from the output layer.
The output value is compared with the desired output value, and the coupling strength is changed so as to reduce the difference.

Learning is difficult with the middle class. This is because it is not known which W (coupling strength) causes the error. Backpropagation is a learning algorithm for such multi-layer networks. It was proposed by Ramel Hart et al. Of the University of California in 1986.

Actual output value (O _pj ) when given pattern p
And the desired output value (t _pj ) It is defined as This represents an error in output unit j. t
_pj is teacher data given by a human.

To train, you can change the strength of all bonds to reduce this error. Here, the change amount of W _ij when the pattern p is given is Decide.

By modifying this formula (5), ΔW can be formulated. The result is ΔpW _ij = ηδ _pj O _pj (6). Here, O _pi is an input value from the unit i to the unit j, and δ _pj is different depending on whether the unit j is the output unit or the intermediate unit. In the case of the output unit, δ _pj = (t _pj −σ _pj ) fi ′ (net _pj ) (7) In the case of the intermediate unit, Is. Expression (8) is a recursive function of δ.

The calculation of ΔW starts with units in the output layer and moves to units in the intermediate layer. In the intermediate unit, ΔW cannot be calculated unless the preceding ΔW is determined. Therefore, the calculation can be performed only by going back to the last input layer. In this way, learning proceeds in the opposite direction to the processing of the input data, that is, backward (backward).

Therefore, learning by backpropagation proceeds as follows. Input learning data and output the results (forward). Change the strength of the bond (backward) to reduce the resulting error. Input the learning data again.
This is repeated until it converges.

A general formula of ΔW is ΔW _ij (n + 1) = ηδ _pj O _pj + αΔW _ij (n). n represents the number of times of learning. The first term on the right side is ΔW just obtained, and the second term is added to reduce the error vibration and accelerate the convergence.

[Action]

The operation of the present invention is to predict the transition of the blast furnace heat level qualitatively and quantitatively from the various blast furnace operation data by the above-mentioned means, and also to use the learning means to improve the prediction accuracy of the network. The automatic generation or modification of the mutual coupling coefficient.

〔Example〕

 Next, a more detailed embodiment will be described with reference to the drawings.

FIG. 1 is a block diagram for explaining the process of input / output and processing of a thermal prediction system for a blast furnace for carrying out the method of the present invention.

In the figure, a circular frame shows the input and output of the system, and a rectangular block shows the processing inside the system.

This system is, of course, embodied by a blast furnace process computer, but the size, number, configuration, etc. of the computer are not restricted by the present invention. However, the input of blast furnace operation data and the output when the prediction result is used for furnace heat control must have sufficient real-time property.

Now, in FIG. 1, the blast furnace operation data 1 is operation data based on the measurement data by various sensors, and these are taken in the system through the data input means 2 and necessary for judging the furnace heat by the data processing means 3. It is processed into various observation events. Conventionally, the operator did not predict only the current value of the blast furnace operation and predicted the transition of the furnace heat, but rather predicted by looking at the trend of measurement data from the past several hours to the present, It is processed into information that includes a temporal factor such as the amount of change over a certain period from the past to the present.

These data are passed to the neural network 4,
The neural network 4 outputs the prediction result 5.

Prediction result 5 has three levels of information: (a) it will become hot, (b) it will not change, and (c) it will become younger. Each of the items (a) to (c) may be digital information of 0 or 1, but the certainty is represented by a number between 0 and 1 so that it can be handled more quantitatively.

Separately, reference input data 6 and teacher output 7 are given to the system for network learning. The reference input data 6 is formally the same as the blast furnace operation data 1, and the teacher output 7 is the same form as the prediction result 5, but the teacher output 7 is accurate information as an output for the reference input data 6. Need to be Although the skilled operator will be the teacher and will provide the data of the teacher output 7, the present invention proposes a method of automatically creating this teacher data as described later. The learning pattern creation process 8 simply converts the reference input data 6 and the teacher output 7 into a pattern that is easy for the neural network to learn.

The furnace heat prediction method having the above configuration will be described in more detail.

First, the blast furnace operation data 1 given to the data input means 2 in FIG. 1 is as follows.

One-sided index This index is an index that indicates whether the index finger interval is faster or slower.

Number of charges Number indicating how many times the current charge is on the day Gas utilization rate (η _CO ) CO, CO ₂ : Top gas component N ₂ component (N ₂ ) [N ₂ ] analysis value of top gas Furnace heat index (H ₀ ) Represents the amount of heat used to heat the pig iron slag, and is used as a so-called furnace heat index. It is used. It distinguishes the heat loss due to the solution loss reaction, the reduction reaction by H ₂ , and the decomposition reaction of CaCO ₃ from the amount of heat generated before the tuyere.

Solution loss Carbon amount (CSOL) Carbon amount gasified by direct reduction Stave heat load (STL): Shows the amount of heat removed from the stave and is represented by the following formula.

STL = F × (T ₀ −T _i ) × k where k: constant F: stave cooling water amount T ₀ : outlet water temperature T _i : inlet water temperature lower pressure loss (DPL) DPL = (hot air pressure) − (shaft pressure lower stage) ) If there are multiple shaft pressure measurement points, there are some DPLs.

Of the above data, is data regarding unloading,
, Are data on gas components, are data on furnace heat itself, are data on temperature and pressure of the furnace body, which are conventionally monitored by operators as clues for predicting furnace heat.

Next, an embodiment of the data processing process 3 in FIG. 1 will be described. The following data processing is performed corresponding to the above items.

′ “One-way index” is processed into the amount of change. The difference between the average of the latest 5 charges and the average of the preceding 5 charges is defined as the change amount of the unilateral index.

′ “Number of charging” is processed to the difference of charging number. The difference between the charging number in the latest 8 hours and the same value used in the previous estimation is the difference in the charging number.

′ “Η _CO ” is processed into the amount of change in η _CO . Last 30
The difference between eta _CO average value of minute eta _CO mean and last 2 hours and the amount of change in eta _CO.

′ “(N ₂ )” is processed into the amount of change in N ₂ . The processing method is the same as ′.

′ “H ₀ ” is processed into the change amount of H ₀ . H ₀ every 30 minutes
The difference between the moving average of the latest 3 hours and the moving average of 5 hours to 7 hours before is the change amount of H ₀ .

'The amount of solution loss C is the amount of change in the amount of solution loss C. The processing method is the same as ′.

'"STL" is processed into the amount of change in STL. Last 30
The difference between the STL average value of minutes and the STL average value of 1 to 2 hours ago is defined as the STL change amount.

'"DPL" is processed into the DPL change amount. The processing method is the same as ′.

Not only this example, but various methods of data processing can be considered. It only needs to be information that can predict changes in furnace heat.

FIG. 2 shows a configuration example of the neural network. A three-layer structure having one hidden layer between the input layer that receives the data processed by the data processing means 3 in FIG. 1 and the output layer of the prediction result is adopted. The two-layer structure of the neural network is linear, whereas the three-layer structure or more is non-linear and the information processing capability is significantly improved. Therefore, in general, a network having three or more layers with an intermediate layer (hidden layer) is used. In the present embodiment, the output layer has three units, and one unit is associated with each of the above-mentioned three pieces of information "will be hot", "maintain the current state", and "will be young".

For the hidden layer, the input layer data can be classified into those related to unloading, those related to gas components, those related to furnace heat, and those related to furnace body temperature / pressure, so one unit (total 4 units) should be associated with each. ing. As shown in FIG. 3, when considering the k-th layer and the i-th unit, first, each unit output of the k−1-th layer, To each weight (mutual coupling coefficient) The sum of those multiplied by, Take a function However, output the product of T: constant It is what

However, θ is a threshold value.

A unit that always takes 1 is provided on the input side, and its weighting coefficient is set to −θ so that it is reflected in the above formula (12), and the threshold value can be corrected by learning.

Since the function is a sigmoid function as in the above equation (13), the threshold value comes to a position where f (x) = 0.5 as shown in FIG. Further, by using such a sigmoid function, all the outputs of each unit take a value of 0 to 1. In addition to normalizing, the aim is to match the concept of confidence in the output layer.

When data as shown in FIG. 2 is given to the input layer, each layer and each unit sequentially propagates the output from the lower layer to the upper layer in accordance with the equation (14).

FIG. 7 shows an example in which the output of the output layer is viewed as time series data. The output of the three units is shown as a trend. As described above, each has a value of 0 to 1, and the closer the value is to 1, the higher the probability is, and the closer to 0, the lower the probability is. With this value, the transition of furnace heat can be predicted qualitatively and quantitatively.

Although the above is the embodiment of prediction, learning of the mutual coupling coefficient of the network is performed as follows.

The learning method is the backpropagation method of Rummelhardt et al.

The learning rule is as follows.

Correction amount Where ε and α are constants Amount of correction of the weighting factor from the j-th layer j unit to the K-th layer i unit at time t When the teacher output of the third layer i unit is y _i , the learning coefficient d _i is
In the output layer, Otherwise, And f is a sigmoid function, f ′ (x) = f (x) (1-f (x)) (18) Good.

The correction proceeds counter-propagating from the output layer to the input layer.
In general, learning converges by iteratively calculating forward propagation and backward propagation multiple times based on reference input data given to the input layer and teacher output.

First, the teacher output data for learning is created as follows.
For the prediction of furnace heat level transition, the actual results are used to determine the hot metal temperature (HMT
(Abbreviated) is evaluated by the actual trend. For example, assume that the trend of HMT and the prediction result is obtained as shown in FIG. Predetermine how many hours ahead the forecast will be. 3 here
Time. Current HMT to HMT _t , HM 3 hours ago
When the T and _{HMT t-3,} it is possible to evaluate the difference HMT _t -HMT _t-3 as Jitsuseki.

This value is multiplied by the function shown in FIG. In FIG. 9, the functions of Ff ₁ , f ₂ and f ₃ having values in the interval [0.1] are prepared. The respective meanings are: f ₁ : became hot f ₂ : remained unchanged f ₃ : became younger. In the example of FIG. 8, HMT _t −HMT _t-3 is +
The temperature is 15 ° C., and at this time, f ₁ = 0.5, f ₂ = 0.5, and f ₃ = 0. That is, this is the teacher output.

Then, the input data three hours ago and the output data of the prediction result are brought over again and subjected to the learning calculation of the network together with the teacher output.

If this is performed in a certain cycle, the mutual coupling coefficient of the network can be automatically generated in each cycle. In this example, the HMT actual value was used to generate the teacher output data, but there are many other indexes that represent the state of furnace heat. The furnace heat index H ₀ or the like may be used. Don't worry about the fixed cycle. Learning may be triggered when good teacher data is generated.

Once the initial weighting coefficient has been created for learning, the frequency of learning for modification can be relatively low since fine modification is sufficient thereafter.

FIG. 10 shows the overall flow. The normal prediction flow is a fixed cycle start, and the learning flow is also a fixed cycle start. In the embodiment, the former is activated every 20 minutes and the latter is activated at regular intervals of every day.

FIG. 11 shows the trend of the prediction result by the method of the present invention and the trend of the result of the HMT transition in the actual operation. It can be seen that the prediction is accurate. Moreover, based on this prediction result, the trend when the action of the blast temperature is taken
Shown in the figure. The hot metal temperature is now controlled to the proper level.

〔The invention's effect〕

INDUSTRIAL APPLICABILITY According to the present invention, the prediction accuracy of furnace heat is improved, and when this is used to construct a furnace heat control model and the blast temperature and coke ratio actions are appropriately taken, the hot metal temperature is stably controlled at an appropriate level. I can do it now.

In addition, since the learning method of neural network is used for the prediction system, maintainability such as parameter tuning and development efficiency are improved compared with the conventional model.

[Brief description of drawings]

FIG. 1 is a block diagram of a system for carrying out the method of the present invention, FIG. 2 is an explanatory diagram of an embodiment of a neural network, FIG. 3 is an explanatory diagram of the structure of a unit of the neural network, and FIG. 4 is a neural network. Schematic diagram of
Fig. 5 is a hierarchical structure diagram of the neural network, Fig. 6 is a graph of the internal function of the unit of the neural network, Fig. 7 is a graph showing an example of the output layer output of the neural network, and Fig. 8 is the furnace heat level. FIG. 9 is a graph showing a trend of the hot metal temperature, FIG. 9 is a graph of a function serving as a reference of teacher output, and FIG.
11 and 12 are graphs showing the prediction results in the examples. 1 ... Blast furnace operation data 2 ... Input means 3 ... Processing means 4 ... Neural network 5 ... Network prediction result 6 ... Learning reference input data 7 ... Learning teacher output 8 ... Learning pattern creation processing.

Front page continuation (72) Inventor Keiji Kobayashi 1 Kawasaki-cho, Chiba-shi, Chiba Kawasaki Steel Co., Ltd. Chiba Works (72) Inventor Nobuhiro Takashima 2-3-2 Uchisaiwaicho, Chiyoda-ku, Tokyo Kawasaki Steel Co., Ltd. Tokyo Head office

Claims (2)

[Claims] 1. A method for predicting a future transition of blast furnace heat level, comprising two or more layers pre-learned with an input layer for inputting blast furnace operation data and an output layer for outputting a transition heat level prediction result. Prepare a hierarchical neural network of
The blast furnace operation data is read from various sensors installed in the blast furnace at a predetermined timing, and these data are processed into various data indicating the recent fluctuation status and input to the input layer of the neural network, Internally, each unit in each layer inputs the links from all units in the previous layer, sums them, multiplies the result by an internal function to obtain one output, and outputs the output to all units in the next layer. A blast furnace heat prediction method characterized by quantitatively and qualitatively predicting the future transition of the furnace heat level by quantitatively outputting the prediction result from at least one unit finally provided in the output layer. 2. The predicting method according to claim 1, wherein the actual results of transition of the furnace heat level are quantified by the recent change amount of the measured value representing the furnace heat, and when the change amount is large in the positive direction, the furnace heat is It is determined that the temperature has risen, and when it is large in the negative direction, it is determined that the furnace heat has decreased, and teacher data of the network prediction result is created, along with the input value at that time and the actual prediction result by the network, for learning the neural network. A method for predicting heat in a blast furnace characterized in that data is automatically generated, and learning is actually performed using this data to automatically generate the mutual coupling coefficient of the neural network.