JPH11707A - Method for controlling rolling mill and controller for rolling mill - Google Patents

Abstract

PROBLEM TO BE SOLVED: To estimate model error taking rolling variables which becomes the occurrence factor of more model error into consideration without using many tables. SOLUTION: Specifications regarding rolling such as the kind of steel, thickness, width, dimensions of a slab and finishing temp. are inputted by a rolling specification inputting part 7. In the rolling load calculation processing part 2, rolling load is calculated by a rolling load formula based on these input data. A part of collected data is stored in a long term learing input value storage device 11, inputted into a neural network learning controller 12, processed with the neural network and the degree of error which is given to rolling load calculation processing part by variation such as the kind of steel and component value is calculated. The calculated parameters are posted to the long term learning processing part 6. The output of the long term learning processing part 6 is inputted into a rolling load learning correcting/processing part 3 and used for correcting the error in the rolling load calculation processing part 2.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates generally to a control method and a control apparatus for various rolling mills, and more particularly to a rolling method for accurately correcting model errors by learning control in a setting calculation having a rolling model and performing accurate control. The present invention relates to a machine control method and a control device.

[0002]

2. Description of the Related Art When calculating the setting of a rolling mill, for example, regarding the control of the thickness of the rolling mill, the thickness accuracy of the steel sheet tip portion, which is an unsteady portion, depends greatly on the calculation accuracy of the roll gap in the setting calculation. doing. Generally, the method of setting the roll gap is given by the solution of the rolling load equation (1) describing the plastic properties of the material and the cage meter equation (2) describing the elastic properties of the rolling mill. High model accuracy is required. Examples of the gauge meter type and the rolling load type are shown below.

[0003]

(Equation 1) P = K (h−S) (2) where P is a rolling load, b is a plate width, R ′ is a flat roll radius, H is a stand-in side plate thickness, h is a stand-out side plate thickness, k
_P is the average deformation resistance, Q _P is the rolling load coefficient, α is the tension distribution coefficient, t _f is the front unit tension, t _b is the rear unit tension, K is the mill longitudinal rigidity constant, and S is the roll gap.

[0004] However, for identification which is important for improving the model accuracy, it is necessary to regress the coefficients with high accuracy based on the rolling variables constituting the model. However, in many cases, the rolling variables cannot be measured. An error occurs. Therefore, in the setting calculation, learning correction to increase the rolling accuracy in the next and subsequent times is generally performed by grasping the error caused by the model in advance and correcting the model. The learning correction includes two methods, a short-term learning correction and a long-term learning correction. For the former, when the same material and the material in the same rod are continuously rolled, the error of the model calculated from the most recent rolling results is reflected in the set-up calculation value, for example, to change the temperature of the rolling mill, roll wear, etc. This is to absorb a model error caused by the accompanying change in the rolling state.

[0005] In the latter case, when the specification of the rolled material changes, the error of the model stored in the past when the same material and the material in the same rod are rolled is reflected in the set calculation value.

In the following, the learning method will be specifically described by taking a rolling load equation as an example.

The load model error ratio Z _k is described by the ratio (3) between the actual calculated load P _{A obtained by} recalculating the load and the measured load value P using actual measured values such as plate thickness and temperature.

Z _k = P / P _A (3) Further, assuming that the rolling order number of the rolled material is i, the short-term learning coefficient Zi of the i-th material is calculated as an exponential smoothing between the short-term learning coefficient of the preceding material and the model error ratio. A method of determining by processing is general, and assuming that the smoothing constant is a, it is obtained by equation (4).

Z _i = (1−a) Z _i−1 + aZ _k (4) Since the portion related to the deformation resistance and the rolling load coefficient among the model errors is uniquely determined by the steel type, composition and size, the front material When the specifications are the same, the error can be sufficiently absorbed by the short-term learning correction.

On the other hand, if the specification changes and the model error changes significantly, the model error cannot be absorbed by the short-term learning correction, so that the long-term learning correction is required. Long-term learning modification, of the rolled material in the past of the same material and the same in the rod, using the model error Z _k shown in equation (3), a method of absorbing the model error. For that purpose, first, in order to determine the category of materials having the same specification, it is necessary to perform stratification by using material properties such as steel type and size as parameters, and to store past model error ratios.

[0011]

However, in order to accurately predict the model error, it is necessary to classify the sections in detail according to the parameters of the material properties, and an enormous table must be created. Further, mere segmentation reduces the number of data stored in the table, which hinders improvement in learning accuracy.

Incidentally, the determination of the correction coefficient in the long-term learning is described, for example, in Japanese Patent Application Laid-Open No. 61-202711.
The method disclosed in Japanese Unexamined Patent Application Publication No. 2005-189686, "Rolling Mill Learning Control Method and Apparatus", is used to determine a long-term learning correction coefficient based on a past model error of a rolled material in the same material and same rod. In addition, the learning gain is determined in consideration of the past number of rolls, and it is considered effective when the table is subdivided and the accumulated data is reduced.

However, there is a problem in that subdivision and adjustment of the learning gain in each layer of the table require a large amount of adjustment work, so that efficient and highly accurate learning cannot be expected.

As described above, when learning control of a model error peculiar to a material is performed, one effective method is to store a correction coefficient of the correction factor in a table subdivided into rolling variables having a high degree of influence that causes a model error. Although it is a means, a huge table is required, not only a great deal of adjustment labor is required, but also the accumulated results are rare, so that there is a problem that a sufficient learning effect cannot be expected.

Therefore, the present invention firstly makes it possible to estimate a model error in consideration of a rolling variable which becomes a more model error generating factor without using many tables, and secondly, makes a great deal of adjustment. It is an object of the present invention to be able to accurately estimate a model error without requiring great effort.

[0016]

SUMMARY OF THE INVENTION The object of the present invention is to determine each set value of a rolling mill in accordance with a mathematical model in which the characteristics of a rolling mill are mathematically modeled, and to feed back a rolling result to correct errors in the mathematical model. A method of controlling a rolling mill that performs rolling, wherein a model error of a rolled material of the same specification or a rolled material of the same type is obtained based on past model error results of a rolled material of the same specification or a rolled material of the same type. And estimate
Long-term learning control that corrects the mathematical model calculation value, and model errors caused by changes in rolling conditions due to short-term changes such as temperature changes and wear of the rolling mill, are estimated from the actual model error of the nearest rolling material, A method for controlling a rolling mill, wherein the long-term learning control is performed using a neural network, wherein the short-term learning control for correcting the model calculation value is reflected in a rolling model. Is solved by

By using a neural network,
How the change of each parameter used for the long-term learning control affects the error of the mathematical model can be accurately grasped as the learning progresses. Therefore, the long-term learning control becomes accurate, and the control of the rolling mill with high accuracy becomes possible.

Further, by using the neural network, it is not necessary to finely classify the rolled material.
The need for a large number of tables is eliminated. Furthermore, since it is not necessary to accurately estimate the error factors in advance,
There is no need to determine the exact mathematical model skeleton in advance.

Since the calculation of the neural network requires an extremely large amount of calculation, if the calculation is performed by a control computer, there is a possibility that a problem such as a delayed response of the control may be caused. Therefore, performing the long-term learning control using a neural network offline, accumulating a part of the rolling results in a rolling results collection device, and performing the long-term learning control offline using the rolling results stored in the rolling results collection device as a teacher signal. These problems can be solved by using a neural network and using the result obtained by the neural network in an online controller (claim 2).

A major factor that causes a model error specific to a rolled material of the same specification or a rolled material of the same type is a steel type or a component value. Therefore, at least one of these
By using a parameter for estimating an error of a model peculiar to a rolled material of the same specification or a rolled material of the same type (claim 3), accurate long-term learning can be performed.

Further, if the sheet thickness, the sheet width, and the rolling reduction are included in addition to these parameters (claim 4), more accurate long-term learning can be performed.

As a method of combining the long-term learning control and the short-term learning control, if the formula value calculated by the long-term learning control is further corrected by the short-term learning control (claim 5). Even if the learning control includes an error, the error can be quickly compensated for by the short-term learning control.

Further, the above-mentioned problems are caused by a rolling specification input section for inputting a rolling specification, a rolling load calculation processing section for calculating a rolling load based on data from the rolling specification input section, and a calculation result of the rolling load calculation processing section. Based on the roll gap calculation processing unit that determines the roll gap of the rolling mill based on the rolling results collection processing unit that collects the rolling results, based on the rolling results,
A short-term learning processing unit that learns short-term changes such as temperature changes and wear of the rolling mill to correct the calculation result of the rolling load calculation processing unit, and a long-term learning input value storage device that collects actual data of the rolled material, A neural network learning processing device that calculates, using a neural network, a parameter for correcting an error in the rolling load calculation processing unit, using the data stored in the long-term learning input storage device as a teacher signal, and inputs the data to the long-term learning input storage device. The long-term learning input unit for inputting the standard value data of the rolled material corresponding to the actual data to be transferred, and having the parameters calculated by the neural network learning processing device transcribed therein, based on the data from the long-term learning input unit. A long-term learning processing unit that corrects the error of the rolling load calculation processing unit; Control apparatus comprising and a rolling load learning correction processing unit that performs heavy learning corrected rolling mill (claim 6)
Is solved by

The rolling specification input unit inputs rolling specifications such as steel type, plate thickness, plate width, slab dimensions, and finishing temperature. The rolling load calculation processing unit calculates a rolling load by a rolling load formula based on these input data. Of the input rolling specifications, those to be subjected to long-term learning, such as steel type and component values, are input to a long-term learning processing unit via a long-term learning input unit.

On the other hand, past data on rolled materials
The data is collected by the rolling result collection processing unit, input to the rolling load learning correction processing unit via the short-term learning processing unit, and the short-term error of the rolling load calculation processing unit is corrected.

A part of the collected data is stored in the long-term learning input value storage device, input to the neural network learning processing device and processed by the neural network, and the variation of steel type, component value, etc. is rolled. The degree of error given to the load calculation processing unit is calculated. Then, the calculated parameters are transcribed to the long-term learning processing unit, and the output of the long-term learning processing unit is input to the rolling load learning correction processing unit and used to correct an error of the rolling load calculation processing unit.

In the configuration of the control device, a long-term learning input value storage device and a neural network learning processing device are
By configuring the control device independent of the other parts of the control device (claim 7), the amount of data of the control device is not increased and the calculation of the neural network is not hindered. An enormous amount of calculation required for the calculation can be performed.

After the rolling load learning correction processing unit performs learning correction of the rolling load based on the output result of the long-term learning processing unit, the result is further corrected based on the output result of the short-term learning processing unit. By providing a function to perform learning correction of the rolling load (claim 8), even if the long-term learning control includes an error, the error is quickly compensated by the short-term learning control. be able to.

[0029]

Embodiments of the present invention will be described below in detail with reference to the drawings.

FIG. 1 is a block diagram showing an example of a thickness control device for a continuous rolling mill according to the present invention.

In FIG. 1, 1 is a thickness control device, 2 is a rolling load calculation processing unit, 3 is a rolling load learning correction processing unit, 4 is a roll gap calculation processing unit, 5 is a short-term learning processing unit, and 6 is a long-term learning process. A processing unit, 7 is a rolling specification input unit, 8 is a long-term learning input unit, 9 is a rolling performance collection processing unit, 10 is a rolling mill, 11 is a long-term learning input value storage device, and 12 is a neural network learning processing device. In this embodiment, the long-term learning input storage device 11 and the neural network learning processing device 12 are configured independently of the thickness control device 1.

The rolling specification input section 7 includes a steel type, a sheet thickness, a sheet width,
Enter specifications related to rolling, such as slab dimensions and finishing temperature. The rolling load calculation processing unit 2 calculates a rolling load by a rolling load formula based on these input data. Of the input rolling specifications, those to be subjected to long-term learning are input to the long-term learning processing unit 6 via the long-term learning input unit 8.

The long-term learning input section 8 basically includes rolling variables such as sheet thickness, sheet width, rolling reduction and the like which have a high influence on the load model calculation results, and elements dominant in the deformation resistance formula included in the load model. A certain temperature or component is selected and stored.

On the other hand, past data on rolled materials
The data is collected by the rolling result collection processing unit 9 and is input to the rolling load learning correction processing unit 3 via the short-term learning processing unit 5 to correct a short-term error of the rolling load calculation processing unit 2.

A part of the collected data is stored in the long-term learning input value storage device 11, is input to the neural network learning control device 12, and is processed by the neural network. Is given to the rolling load calculation processing unit. Then, the calculated parameters are transferred to the long-term learning processing unit 6. The output of the long-term learning processing unit 6 is input to the rolling load learning correction processing unit 3 and is used to correct an error of the rolling load calculation processing unit 2.

In this embodiment, the long-term learning input value storage unit 11 and the neural network learning processing unit 1
2 is composed of a control device independent of the other parts of the control device, so that the data amount of the thickness control device is not enormous, and the calculation of the neural network is not hindered. An enormous amount of calculation required for the calculation can be performed.

However, if the on-line controller has sufficient speed and capacity, the long-term learning input value storage unit 11 and the neural network learning processing unit 12
Need not be a separate controller. In this case, the neural network in the long-term learning processing unit 6 and the neural network in the neural network learning processing unit 12 can be the same.

[0038] Based on the information of the rolling specifications input unit 7, the load calculated value P _C calculated by the rolling load calculation processing unit 2, the learning correction amount obtained by the long-term learning processing unit 6 and the short-term learning processing unit 5 According to Z _N and Z _i , the rolling load learning correction processing unit 3 makes correction. Roll gap calculation processing section 4, the modified load P _S, performing roll gap setting of the rolling mill 10. Also, determine the model error Z _K by calculating the ratio of the load model in rolling result collection processing unit 9 P _A and the load actual, is reflected in the short-term learning process in the subsequent material, together with further long learning input items, a new Is stored in the long-term learning input storage device 11 in order to perform the learning calculation of the offline neural network. In addition, the learning process of the rolling state in the short-term learning processing unit 5 is obtained by Expression (4).

FIG. 2 is a block diagram for explaining the long-term learning processing section 6 and the neural network learning processing device 12 in detail. In the following drawings, the same components as those shown in the preceding drawings are denoted by the same reference numerals, and description thereof will be omitted. The long-term learning processing unit 6 classifies all the rolling specification materials into the same type of material such as a steel type, and includes a plurality of long-term learning neural network processing units 6-b constructed for each characteristic of each material group. It is composed of an input value conversion unit 6-a and an output value conversion unit 6-c that perform normalized numerical processing, and obtains a long-term learning correction coefficient 6-d as a final output value. Note that the neural network employs a three-layer structure including an input layer, a middle layer, and an output layer. However, the same can be applied to a multi-layer or non-hierarchical type.

The feature of this embodiment is that the long-term learning neural network processing unit 6-b does not have a function of updating neural network parameters such as a synaptic connection weight coefficient and a sigmoid function threshold value by itself.
The output from the input through the neural network is obtained using these parameters provided from the neural network learning processing device 12.

Neural network learning processor 12
Is composed of a plurality of neural network learning processing units 12-b configured for each material group similar to the long-term learning processing unit 6, and an input value conversion unit 12-a that performs normalized numerical processing of input values. Used by the long-term learning neural network processing unit 6-b, a model error Z _K updating the neural network parameters, such as synaptic weight coefficient and sigmoid function threshold and a teacher signal 12-c, the neural network parameter adjustment processor At 12-d, parameter learning is performed by the back propagation method.
Note that the back propagation method is a general method for learning a neural network, and a description thereof will be omitted. The learned neural network parameter 12-e obtained by the learning is arbitrarily set to the long-term learning neural network processing unit 6-b of the long-term learning processing unit 6 so that a new learning target material can be flexibly handled.
And the neural network calculation corresponding to the transmitted neural network parameters such as the synaptic connection weighting coefficient and the sigmoid function threshold is executed.

Table 1 shows an example of division between the long-term learning neural network processing unit 6-b and the neural network learning processing unit 12-b.

[0043]

[Table 1]

The NN number indicates a neural network number, and eight long-term learning neural network processing units 6-b and eight neural network learning processing units 12-b are independently provided.
It is selected and used according to the corresponding steel type.

FIG. 3 shows a neural network learning processing unit 6
b and the contents of the neural network learning processing unit 12-b are shown in a simplified manner. As the input, the sheet thickness, the sheet width, the rolling reduction, the amounts of C, Mn, and Si are input to the input layer (i), and are combined with the intermediate layer (j) with a predetermined weight. Note that FIG.
In FIG. 7, some connections between the input layer (i) and the intermediate layer (j) are omitted. The output of the middle layer (j) is coupled to the output layer with a specified weight and output as the Z term. The Z term is described as Z _N in FIG.
The predicted rolling load calculated by the rolling load calculation processing unit 2 is Z
_{The value obtained} by multiplying _N is given to the rolling load learning correction processing unit 3 as a corrected rolling load, and is used for correcting the rolling load.

In the present embodiment, the neural network learning processor 6-b only calculates and outputs the Z term based on the parameters input to the input layer, and outputs the parameters of the neural network itself. No correction, that is, learning correction such as a synaptic connection weight coefficient or a sigmoid function threshold is performed.

On the other hand, in the neural network learning processing unit 12-b, the Z term is made to match the actual Z _k obtained through the input conversion unit 12-a in FIG. Learning correction such as a synaptic connection weight coefficient and a sigmoid function threshold is performed as 12-C. These calculations require a large amount of memory and calculation time. However, since the neural network learning processing device 12 is an off-line device, it does not disturb online control.

In the above description, the long-term learning control is
The short-term learning control was used when the rolling conditions such as steel type and sheet thickness category were significantly changed, and the short-term learning control was used within the range where the steel type and sheet thickness category etc. did not change significantly. Is advantageous in that even if there are errors in the model of the long-term learning control, those errors are quickly corrected by the short-term learning control.

That is, the rolling load calculation processing unit 2 calculates
The calculated load of the load model_CAnd the long-term learning processing unit 6
The learning correction amount calculated by_NThen, for long-term learning control
The more modified calculated load is Z_N× P_CBut this is
Further, the learning correction amount Z calculated by the short-term learning processing unit 5_iBy
Corrected and modified calculated load P_STo P_S= Z_i× Z_N× P_C … (5)  And the corrected calculated load P_SBy low
The gap calculation processing unit 4 determines the roll gap of the rolling mill 10.
Make settings. This modified calculated load P_SThe calculation of the pressure
This is performed by the extension load learning correction processing unit 3.

In this case, if the short-term learning process is performed based on the equation (4), the short-term learning process will be performed on data containing data related to the long-term learning process. Therefore, in order to remove the influence of the long-term learning process from the short-term learning process, the exponential smoothing process is performed by the following formula (6) instead of the formula (4). Z _i = (1−a) Z _i−1 + a (Z _k / Z _N ) (6) That is, the load model error ratio Z _k calculated by the equation (3)
Is divided by Z _N to cancel the effect of the long-term learning process and exponentially smooth the one converted to the learning coefficient related to the short-term learning.

FIG. 4 shows a graph of a load model error prediction result using one embodiment of the present invention. FIG. 4 shows the final stand F7 in a continuous hot strip rolling mill of seven continuous machines.
The figure shows the prediction of the load model error of the stand and the actual value of the load model error for a thin high-tensile material (thickness of 3.5 mm or less). In FIG. 4, the horizontal axis represents the rolled material layered for each sheet thickness, and the numerals are the coil numbers. The vertical axis shows the model error, and the black circles show how much the actual load value differs from the load value calculated by the static model (P / P _C ). The white circles indicate the prediction of the load error indicated by the mathematical model, that is, the value calculated by the model of the present invention (P /
It shows the predicted value of P _C). The Δ marks indicate the residuals, and indicate the differences between the values of the black circles and the white circles.

In FIG. 4, the portions indicated by arrows A, B, C, D, E, and F correspond to the plate widths and plate thicknesses indicated by the reference symbols in Table 2, respectively. For example,
The portion indicated by arrow D corresponds to a material having a board thickness of 2.65 mm and a board width of 1121 mm.

[0053]

[Table 2]

Of these data, for the materials C, D, and E, for which a large amount of data is collected before learning the data of FIG. 4 and learning is advanced, the rolling load is predicted with an accuracy within 10%. is made of. It is considered that the reason why the prediction accuracy of the rolling load is poor in the portions A, B, and F is that the data of the sheet thickness and the sheet width is not accumulated, and the learning is not advanced.

Table 3 shows the results of FIG.
As a result, the difference of the load model error prediction based on the actual load model error is -0.8
% And the standard deviation was 9.3%. From this, it is understood that the model error can be accurately estimated within 10% by the device of the present invention.

[0056]

[Table 3]

[0057]

As described above, according to the present invention, in learning control in setting calculation of a rolling mill, a model error peculiar to a material is obtained by a neural network, so that an enormous table and a large amount of adjustment labor are required. Without needing,
Prediction can be made with high accuracy. In particular, by performing the calculation of the neural network off-line, accurate control can be performed without disturbing the on-line control.

[Brief description of the drawings]

FIG. 1 is a block diagram showing a rolling mill control device according to one embodiment of the present invention.

FIG. 2 is a block diagram illustrating an outline of a neural network processing unit according to the first embodiment of the present invention.

FIG. 3 is a diagram showing simplified contents of a neural network learning processing unit and a neural network learning processing unit;

FIG. 4 is a diagram showing the load model error prediction and the actual value of the load model error for a thin material of a high-tensile material.

[Explanation of symbols]

 1 Sheet thickness control device 2 Rolling load calculation processing unit 3 Rolling load learning correction processing unit 4 Roll gap calculation processing unit 5 Short term learning processing unit 6 Long term learning processing unit 7 Rolling specification input unit 8 Long term learning input unit 9 Rolling result collection processing unit Reference Signs List 10 rolling mill 11 long-term learning input value storage device 12 neural network learning processing device

Claims (8)

[Claims] A control method for a rolling mill that determines each set value of a rolling mill according to a mathematical model in which the characteristics of the rolling mill are mathematically modeled, feeds back a rolling result, and corrects an error in the mathematical model to perform rolling. From the past results of the model error of the rolled material of the same specification or the same type of specification, the error of the model specific to the rolled material of the same specification or the same type of specification is estimated, and the mathematical model calculation value Model errors caused by changes in rolling conditions due to short-term changes such as temperature changes and wear of rolling mills are estimated from the model errors of the closest rolled material, and corrected to model calculation values. A rolling mill control method for reflecting the short-term learning control to the rolling model, wherein the long-term learning control is performed using a neural network. . 2. A control method for a rolling mill that determines each set value of a rolling mill in accordance with a mathematical model in which characteristics of the rolling mill are mathematically modeled, feeds back a rolling result, and corrects an error in the mathematical model to perform rolling. In the past, from the results of the model error of the rolled material of the same specification or rolled material of the same specification in the past, the error of the model specific to the rolled material of the same specification or the rolled material of the same specification is estimated, and the model calculation value is calculated. Estimate model errors caused by changes in rolling conditions due to short-term changes such as temperature changes and wear of rolling mills from long-term learning control to add corrections, and estimate the model errors from the model errors of the closest rolled material to model calculation values. In the rolling mill learning control method that reflects the short-term learning control to the rolling model, a part of the rolling results is stored in the rolling results collection device, and the rolling results stored in the rolling results collection device are taught. A method for controlling a rolling mill, wherein the long-term learning control is performed offline using a neural network as a teacher signal, and the result obtained by the neural network is used in an online control device. 3. A parameter for estimating an error of a model peculiar to a rolled material of the same specification or a rolled material of the same type includes at least one of a steel type and a component value of the material. A method for controlling a rolling mill according to claim 2. 4. Parameters for estimating an error of a model specific to a rolled material of the same specification or a rolled material of the same type include sheet thickness, sheet width, reduction ratio, and at least the steel type or component value of the material. The method for controlling a rolling mill according to claim 1, wherein the control method includes one of them. 5. The method according to claim 1, wherein the numerical model calculation value corrected by the long-term learning control is further corrected by the short-term learning control.
The control method of a rolling mill according to the paragraph. 6. A rolling specification input unit for inputting a rolling specification,
A rolling load calculation processing unit that calculates a rolling load based on data from a rolling specification input unit; a roll gap calculation processing unit that determines a roll gap of a rolling mill based on a calculation result of the rolling load calculation processing unit; And a short-term learning processing unit that learns short-term changes such as temperature changes and wear of the rolling mill and corrects the calculation result of the rolling load calculation processing unit based on the rolling results. A long-term learning input value storage device that collects actual data of the rolled material, and using the data stored in the long-term learning input storage device as a teacher signal, using a neural network to correct the error of the rolling load calculation processing unit. Neural network learning processing device to calculate and long-term learning to input standard value data of rolled material corresponding to actual data input to long-term learning input storage device A force section, a long-term learning processing section that transcribes parameters calculated by the neural network learning processing device, corrects an error of the rolling load calculation processing section based on data from the long-term learning input section, A control device for a rolling mill, comprising: a processing unit; and a rolling load learning correction processing unit that performs learning correction of a rolling load based on an output result of the long-term learning processing unit. 7. The control device for a rolling mill according to claim 6, wherein the long-term learning input value storage unit and the neural network learning processing device comprise a control device independent of other portions of the control device. 8. The rolling load learning correction processing unit performs learning correction of the rolling load based on the output result of the long-term learning processing unit, and further, based on the result, based on the output result of the short-term learning processing unit. The control device for a rolling mill according to claim 6 or 7, wherein learning control of the rolling load is performed.