CN114798763B - Method and system for predicting rough rolling outlet temperature of heating furnace tapping plate blank - Google Patents
Method and system for predicting rough rolling outlet temperature of heating furnace tapping plate blank Download PDFInfo
- Publication number
- CN114798763B CN114798763B CN202110126990.3A CN202110126990A CN114798763B CN 114798763 B CN114798763 B CN 114798763B CN 202110126990 A CN202110126990 A CN 202110126990A CN 114798763 B CN114798763 B CN 114798763B
- Authority
- CN
- China
- Prior art keywords
- rough rolling
- temperature
- production process
- process data
- outlet temperature
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
- 238000005096 rolling process Methods 0.000 title claims abstract description 164
- 238000010438 heat treatment Methods 0.000 title claims abstract description 107
- 238000000034 method Methods 0.000 title claims abstract description 47
- 238000010079 rubber tapping Methods 0.000 title claims abstract description 25
- 238000004519 manufacturing process Methods 0.000 claims abstract description 104
- 238000007637 random forest analysis Methods 0.000 claims abstract description 41
- 238000012216 screening Methods 0.000 claims abstract description 27
- 238000004422 calculation algorithm Methods 0.000 claims abstract description 21
- 238000005098 hot rolling Methods 0.000 claims abstract description 19
- 238000013528 artificial neural network Methods 0.000 claims abstract description 18
- 230000000694 effects Effects 0.000 claims abstract description 10
- 230000007787 long-term memory Effects 0.000 claims abstract description 7
- 238000007781 pre-processing Methods 0.000 claims abstract description 6
- 230000006403 short-term memory Effects 0.000 claims abstract description 6
- 229910000831 Steel Inorganic materials 0.000 claims description 28
- 239000010959 steel Substances 0.000 claims description 28
- 239000000463 material Substances 0.000 claims description 26
- 238000002791 soaking Methods 0.000 claims description 21
- 238000007599 discharging Methods 0.000 claims description 20
- 238000012545 processing Methods 0.000 claims description 9
- 238000010606 normalization Methods 0.000 claims description 8
- 230000015654 memory Effects 0.000 claims description 7
- 238000012163 sequencing technique Methods 0.000 claims description 4
- 238000012360 testing method Methods 0.000 description 19
- 210000002569 neuron Anatomy 0.000 description 16
- 238000012549 training Methods 0.000 description 16
- 230000006870 function Effects 0.000 description 13
- 230000008569 process Effects 0.000 description 13
- 230000001133 acceleration Effects 0.000 description 11
- 230000004913 activation Effects 0.000 description 11
- 230000011218 segmentation Effects 0.000 description 10
- 238000004364 calculation method Methods 0.000 description 8
- 230000007774 longterm Effects 0.000 description 7
- 230000004044 response Effects 0.000 description 6
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 description 5
- 230000008859 change Effects 0.000 description 5
- 210000004027 cell Anatomy 0.000 description 4
- 238000012937 correction Methods 0.000 description 4
- 238000003066 decision tree Methods 0.000 description 3
- 238000010586 diagram Methods 0.000 description 3
- 238000002474 experimental method Methods 0.000 description 3
- 239000011159 matrix material Substances 0.000 description 3
- 230000000717 retained effect Effects 0.000 description 3
- 238000012952 Resampling Methods 0.000 description 2
- 230000009286 beneficial effect Effects 0.000 description 2
- 230000008901 benefit Effects 0.000 description 2
- 230000033228 biological regulation Effects 0.000 description 2
- 230000007423 decrease Effects 0.000 description 2
- 238000012217 deletion Methods 0.000 description 2
- 230000037430 deletion Effects 0.000 description 2
- 238000005516 engineering process Methods 0.000 description 2
- 238000011156 evaluation Methods 0.000 description 2
- 238000005457 optimization Methods 0.000 description 2
- 230000000306 recurrent effect Effects 0.000 description 2
- 230000003044 adaptive effect Effects 0.000 description 1
- 238000004458 analytical method Methods 0.000 description 1
- 238000013473 artificial intelligence Methods 0.000 description 1
- 238000009529 body temperature measurement Methods 0.000 description 1
- 238000013527 convolutional neural network Methods 0.000 description 1
- 238000007405 data analysis Methods 0.000 description 1
- 238000013075 data extraction Methods 0.000 description 1
- 230000003247 decreasing effect Effects 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 230000007123 defense Effects 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 238000001514 detection method Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 230000018109 developmental process Effects 0.000 description 1
- 238000001914 filtration Methods 0.000 description 1
- 238000010801 machine learning Methods 0.000 description 1
- 230000014759 maintenance of location Effects 0.000 description 1
- 238000013178 mathematical model Methods 0.000 description 1
- 238000005259 measurement Methods 0.000 description 1
- 230000007246 mechanism Effects 0.000 description 1
- 239000000203 mixture Substances 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 230000001537 neural effect Effects 0.000 description 1
- 238000004321 preservation Methods 0.000 description 1
- 238000013138 pruning Methods 0.000 description 1
- 230000009467 reduction Effects 0.000 description 1
- 238000005070 sampling Methods 0.000 description 1
- 238000012795 verification Methods 0.000 description 1
Classifications
-
- B—PERFORMING OPERATIONS; TRANSPORTING
- B21—MECHANICAL METAL-WORKING WITHOUT ESSENTIALLY REMOVING MATERIAL; PUNCHING METAL
- B21B—ROLLING OF METAL
- B21B38/00—Methods or devices for measuring, detecting or monitoring specially adapted for metal-rolling mills, e.g. position detection, inspection of the product
- B21B38/006—Methods or devices for measuring, detecting or monitoring specially adapted for metal-rolling mills, e.g. position detection, inspection of the product for measuring temperature
Landscapes
- Engineering & Computer Science (AREA)
- Mechanical Engineering (AREA)
- Control Of Metal Rolling (AREA)
Abstract
The invention discloses a method and a system for predicting rough rolling outlet temperature of a heating furnace tapping plate blank, wherein the method comprises the following steps: collecting production process data of a hot-rolled slab; preprocessing production process data; constructing a rough rolling outlet temperature prediction model based on a long-term and short-term memory neural network; adding the rough rolling outlet temperature at the historical moment as a new input variable for predicting the rough rolling outlet temperature of the hot rolling heating furnace slab at the current moment; selecting a temperature prediction model with the best prediction effect as a selected model; ranking the importance of the production process data by using a random forest algorithm; gradually removing the production process data with low importance scores through random forest screening, and re-inputting the rest production process data into the selected model to predict rough rolling outlet temperature until screening the production process data combination with highest precision of the selected model. The invention collects variable factors which possibly have influence on the outlet temperature of the rough rolling of the blank, screens the variable factors and builds a temperature prediction model according to the variable factors.
Description
Technical Field
The invention relates to a hot rolling prediction method and a hot rolling prediction system, in particular to a hot rolling heating furnace tapping plate blank rough rolling outlet temperature prediction method and a hot rolling heating furnace tapping plate blank rough rolling outlet temperature prediction system.
Background
The hot rolled strip steel is one of important products in the steel industry and plays an important role in the fields of national defense equipment, automobile manufacturing, aerospace and the like. The technical process of the hot continuous rolling production process of the strip steel is complex. The change of temperature is a very important factor, and can change the microstructure inside the hot rolled strip steel, so that the mechanical property of the strip steel product is changed, and finally the property and quality of the finished strip steel are affected.
For a given specification of steel, the tapping temperature of the hot-rolled heating furnace slab directly influences the rough rolling outlet temperature. Unfortunately, the tapping temperature of the heating furnace plate blank is not directly measured at present, and only can be indirectly calculated by adopting a mathematical model, and meanwhile, an operator adjusts the tapping temperature by experience. Due to the different operating teams, the level is different, leading to fluctuation of the rough rolling outlet temperature, thereby affecting the dimensional accuracy and performance of the product.
The operation of the heating furnace mainly refers to the rough rolling outlet temperature, the temperature of the furnace gas of the heating furnace is adjusted according to the actual measured temperature of the rough rolling outlet intermediate slab, and the deviation of the temperature of the rough rolling outlet billet is reduced as much as possible according to experience, so that the temperature is close to the target temperature of the rough rolling outlet specified by the process. However, in actual production, a relatively long delay time is required for the hot-rolled heating furnace to draw out the slab to reach the roughing mill outlet, resulting in a delay in the adjustment by the operator, and thus a relatively large fluctuation in temperature.
For the steel grade with a given specification, the actual data of the heating furnace and the historical data of the rough rolling outlet temperature are utilized to accurately predict the temperature of the slab to be extracted from the heating furnace at the rough rolling outlet, and the temperature is pre-determined in advance to guide the heating operation, so that the accurate control of the temperature of the slab is facilitated, and the quality of hot rolled products is improved.
At present, some published patent documents also appear on a temperature prediction model of hot continuous rolling strip steel, for example, a method described in a method for improving the calculation accuracy of the head temperature of a hot rolling intermediate billet (patent number: ZL 201510668578.9) document: and expanding a linear calculation interval, adopting a mode of combining a linear algorithm and an average algorithm, filtering temperature false numbers in an automatic first stage and an automatic second stage, and subdividing a temperature value defaulted by a second-stage model according to the thickness of the finished strip steel, thereby improving the accuracy of finish rolling pre-calculation. The method described in the document named 'a method for controlling the final rolling temperature of hot rolled strip based on speed adjustment' (patent number: ZL 201710194635.3): setting a corresponding table of the target thickness of the strip steel and the first acceleration, a corresponding table of the final rolling temperature deviation and the second acceleration, and a corresponding table of the target thickness of the strip steel and the acceleration correction coefficient; searching corresponding first acceleration according to the target thickness of the strip steel, and carrying out speed-up rolling on the strip steel by using the first acceleration; searching corresponding second acceleration according to the final rolling temperature deviation of the strip steel; searching a corresponding acceleration correction coefficient according to the target thickness of the strip steel; multiplying the second acceleration by an acceleration correction coefficient to obtain a third acceleration; the speed regulation is carried out by using the third acceleration, and the hot rolled strip steel finish rolling temperature control method based on the speed regulation can effectively improve the rolling temperature control precision. The limitations of the above patents are: the modeling is not carried out by utilizing a large amount of historical data accumulated in the heating furnace and the rolling production process, the model rules contained in the historical data are not fully mined, and a forecast model based on the temperature of the steel slab to be tapped of the heating furnace driven by data at the rough rolling outlet is not established. The limitations of the above patents are: the modeling is not performed by utilizing a large amount of historical data accumulated in the rolling production process, and model rules contained in the historical data are not fully mined.
In recent years, with the rapid development of artificial intelligence and big data analysis technology, data-driven prediction methods are becoming popular and favored by various nationalities. In the hot continuous rolling production process, a large amount of actual production process data are collected by a computer and a sensor, and the temperature of a slab at the outlet of a roughing stand can be predicted from the large amount of noisy and fuzzy actual production data by using a machine learning method.
Disclosure of Invention
Aiming at the problems in the prior art, the invention aims to provide a method and a system for predicting the rough rolling outlet temperature of a heating furnace tapping plate blank, so that the temperature prediction accuracy is improved.
In order to achieve the above purpose, the invention adopts the following technical scheme:
A method for predicting rough rolling outlet temperature of a heating furnace tapping plate blank comprises the following steps: collecting production process data of a hot-rolled slab; preprocessing production process data; constructing a rough rolling outlet temperature prediction model based on a long-term and short-term memory neural network; adding the rough rolling outlet temperature at the historical moment as a new input variable for predicting the rough rolling outlet temperature of the hot rolling heating furnace slab at the current moment; selecting a temperature prediction model with the best prediction effect as a selected model; ranking the importance of the production process data by using a random forest algorithm; gradually removing the production process data with low importance scores through random forest screening, and re-inputting the rest production process data into the selected model to predict rough rolling outlet temperature until screening the production process data combination with highest precision of the selected model.
Further, the production process data includes: the actual width of the material, the actual thickness of the material, the length of the material, the actual weight of the material, the weight of the charged plate blank, the number of heating furnaces, the number of rows in the furnace, the time in the furnace, the actual charging temperature, the actual discharging temperature, the target discharging temperature value, the negative discharging temperature tolerance, the positive discharging temperature tolerance, the end temperature of the preheating section, the heating time of the first heating section, the heating time of the second heating section, the heating time of the soaking section, the end temperature of the first heating section, the end temperature of the second heating section, the end temperature of the soaking section and the rough rolling outlet temperature.
Further, the preprocessing includes: removing the strip steel records with missing items in the production process data; sequencing each piece of production process data from first to last according to the production time; and carrying out maximum and minimum normalization processing on the production process data.
Further, the rough rolling outlet temperature corresponding to each slab is made into a time sequence, and the rough rolling outlet temperatures of the continuous N slabs before the current slab are intercepted as new input variables.
Further, if the rough rolling outlet temperature data of the previous slab of the current slab to be predicted can be measured, the rough rolling outlet temperatures at the times of t-1, t-2, … and t-N are taken as N new input variables; if the rough rolling outlet temperature data of the second slab before the current slab to be predicted can be measured, taking the rough rolling outlet temperatures at the times of t-2, t-3, … and t-N as N new input variables; if the rough rolling outlet temperature data of the third slab before the current slab to be predicted can be measured, taking the rough rolling outlet temperatures at the times of t-3, t-4, … and t-N as N new input variables; wherein t represents the rough rolling outlet temperature of the slab at the current t moment.
Further, according to the prediction precision of the temperature prediction model, selecting an N value corresponding to the model with the highest precision, and taking the temperature prediction model corresponding to the N value as the selected model.
Further, eliminating production process data with low importance scores through random forest screening, inputting the residual variables into a selected model for predicting rough rolling outlet temperature, and recording the accuracy of a short-term selected model; repeating the above operation until the precision of the selected model is obviously and continuously lowered, and stopping deleting the production process data; comparing the precision of all the selected models, and selecting the selected model with the highest precision; and combining the production process data corresponding to the selected model with the highest precision as final production process data.
In order to achieve the above purpose, the present invention further adopts the following technical scheme:
a predictive system for the rough rolling outlet temperature of a heating furnace tapping slab, comprising: the data acquisition module acquires production process data of the hot-rolled plate blank and preprocesses the production process data; the prediction module is used for constructing a rough rolling outlet temperature prediction model based on a long-short-period memory neural network, adding the rough rolling outlet temperature at the historical moment as a new input variable, predicting the rough rolling outlet temperature of the hot rolling heating furnace slab at the current moment, and finally selecting a temperature prediction model with the best prediction effect as a selected model; and the sorting and screening module is used for sorting the importance of the production process data by using a random forest algorithm, gradually removing the production process data with low importance scores screened by the random forest, and re-inputting the rest production process data into the selected model of the prediction module for predicting the rough rolling outlet temperature until screening the production process data combination with highest precision of the selected model.
Further, the production process data includes: the actual width of the material, the actual thickness of the material, the length of the material, the actual weight of the material, the weight of the charged plate blank, the number of heating furnaces, the number of rows in the furnace, the time in the furnace, the actual charging temperature, the actual discharging temperature, the target discharging temperature value, the negative discharging temperature tolerance, the positive discharging temperature tolerance, the end temperature of the preheating section, the heating time of the first heating section, the heating time of the second heating section, the heating time of the soaking section, the end temperature of the first heating section, the end temperature of the second heating section, the end temperature of the soaking section and the rough rolling outlet temperature.
Further, the data acquisition module eliminates strip steel records with missing items in the production process data, sorts each piece of production process data from first to second according to production time, and then carries out maximum and minimum normalization processing on the production process data.
Further, the prediction module takes the rough rolling outlet temperature corresponding to each slab as a time sequence, intercepts rough rolling outlet temperatures of continuous N slabs before the current slab as new input variables, selects an N value corresponding to a model with highest precision according to the prediction precision of the temperature prediction model, and takes the temperature prediction model corresponding to the N value as a selected model.
Further, the sorting and screening module rejects production process data with low importance scores subjected to random forest screening, inputs the residual variables into the selected model for predicting rough rolling outlet temperature, and records the accuracy of the short-term selected model; the sorting and screening module repeats the above operation until the precision of the selected model is obviously and continuously lowered, and then the deletion of the production process data is stopped, at this time, the precision of all the selected models is compared, the selected model with the highest precision is selected, and finally the production process data corresponding to the selected model with the highest precision is used as the final production process data combination.
In the technical scheme, variable factors which possibly affect the temperature of a rough rolling outlet of a blank are collected from the data of the hot continuous rolling production process of the strip steel, important influencing variables are screened out from the variables by utilizing a random forest algorithm, redundant variables are removed, the dimension reduction of model input data is realized, and a temperature prediction model of a hot rolling heating furnace plate blank at the outlet of a rough rolling frame is established on the basis.
Drawings
FIG. 1 is a flow chart of the method of the present invention;
FIG. 2 is a schematic diagram of an LSTM network neuron structure;
FIG. 3 is an overall expanded schematic of an LSTM predictive model;
FIG. 4 is an architecture diagram of the system of the present invention;
FIG. 5 is a graph showing mtry and random forest model decision coefficients in accordance with one embodiment of the present invention.
Detailed Description
The technical scheme of the invention is further described below with reference to the accompanying drawings and examples.
The invention aims to provide a temperature prediction method and a temperature prediction system for a slab to be tapped from a heating furnace at an outlet of a roughing stand, so that an operator of the heating furnace can predict the temperature of the slab to be extracted at the outlet of the roughing stand in advance, thereby guiding the operation of the heating furnace in time, overcoming the defects of lack of direct temperature measurement and untimely adjustment of the temperature of the slab of the heating furnace, improving the control level of the temperature of the slab, realizing more accurate temperature control of the slab of the roughing, and being beneficial to solving the technical problems of unstable quality and performance of hot rolled strip steel products caused by large temperature fluctuation of the slab of the roughing in the hot continuous rolling production process.
Accordingly, referring to fig. 1, the present invention firstly discloses a temperature prediction method of a slab to be tapped from a heating furnace at an outlet of a roughing stand, which mainly comprises the steps of:
S1: and collecting production process data of the hot rolled plate blank.
S2: preprocessing the production process data.
S3: and constructing a rough rolling outlet temperature prediction model based on a long-short-term memory neural network (LSTM).
S4: the rough rolling outlet temperature at the historical moment is added as a new input variable for predicting the rough rolling outlet temperature of the hot rolling heating furnace slab at the current moment.
S5: the temperature prediction model with the best prediction effect is selected as the selected model of the LSTM.
S6: the production process data is ranked for importance using a random forest algorithm.
S7: progressively eliminating the production process data with low importance scores after random forest screening, namely eliminating the last N variables with importance row, wherein initial N=1.
S8: the remaining production process data is re-entered into the LSTM-based selected model for predicting rough roll outlet temperature.
S9: the iterative step S8 is repeated until a predetermined number of iterations is reached.
S10: the accuracy of the selected model is recorded.
S11: the value of N is increased, i.e. n=n+1.
S12: as the value of N increases, it is determined whether the accuracy of the selected model becomes significantly lower continuously? If yes, go to step S13, if not, return to step S7.
S13: stopping training, comparing the precision of each selected model, and screening out the production process data combination with the highest precision of the selected model.
As shown in fig. 1, the present invention firstly performs step S1, that is, collects the hot rolled slab production process data and pre-processes the actual production process data.
In order to predict the temperature of a slab to be tapped from a hot rolling heating furnace at a rough rolling outlet, the collected relevant data comprise: the response variables include 21 independent variables such as the actual width of the material, the actual thickness of the material, the length of the material, the actual weight of the material, the weight of the charged slab, the number of heating furnaces, the number of rows in the furnaces, the time in the furnaces, the actual charging temperature, the actual discharging temperature, the target discharging temperature, the negative discharging temperature tolerance, the positive discharging temperature tolerance, the end temperature of the preheating section, the heating time of the first heating section, the heating time of the second heating section, the heating time of the soaking section, the end temperature of the first heating section, the end temperature of the second heating section, the end temperature of the soaking section, and the rough rolling outlet temperature. As shown in table 1:
TABLE 1 experimental data composition
Remarks: the number of roughing stands is 4, and the invention takes R4 as an example to collect the outlet temperature as the roughing outlet temperature.
Next, step S2 is executed, namely, the collected data of the hot continuous rolling rough rolling slab production process is subjected to necessary pretreatment:
(1) Removing strip steel records with missing item data in the data records;
(2) Sequencing the data of each hot continuous rolling slab from first to last according to the production time;
(3) Carrying out maximum and minimum normalization processing on the data to enable the original data to be linearly converted into the range of [0,1], wherein the normalization formula is as follows:
Is the value after maximum and minimum normalization processing for x ij, x ij is the value of the j-th variable of the ith data, x j min is the minimum value of all data of the j-th variable, and x j max is the maximum value of all data of the j-th variable.
After the above process is completed, step S3 is executed, namely, a rough rolling outlet temperature prediction model based on a long short term memory neural network (LSTM) is constructed.
The invention adopts a long and short term memory neural network (LSTM) to predict the rough rolling outlet temperature, and the LSTM network is a special Recurrent Neural Network (RNN). Compared with the feedforward neural network such as BP, CNN and the like, signals of each layer of neurons can only propagate to the upper layer, the processing of samples is independent at all times, the RNN has a self-feedback mechanism, signals generated by previous input can be transmitted downwards, and time sequence data can be processed better. The LSTM network is based on a common circulating neural network (RNN), memory units are added in each neural unit of the hidden layer, so that memory information on a time sequence is controllable, and the memory and forgetting degree of previous information and current information can be controlled through a plurality of controllable gates (forgetting gates, input gates and output gates) when each time is transmitted among each unit of the hidden layer, thereby the RNN network has a long-term memory function and can solve the problem of long-term dependence. In the hot continuous rolling production process, the temperature of the first slab influences the temperature of the next slab, a time sequence relation exists between analysis samples from a global angle, and the LSTM network can be utilized to mine the time sequence characteristics between the data samples, so that the prediction of the rough rolling outlet temperature of the hot continuous rolling slab is facilitated.
The hidden layer of the original RNN has only one state, h, which is very sensitive to short-term input. LSTM adds a cell state, C, to the RNN to preserve long-term state. The key to LSTM is how to control long term state C. Here, the LSTM idea is to use three control switches. The first switch is a forgetting door and is responsible for controlling the continuous preservation of the long-term state C; the second switch is an input gate and is responsible for controlling the input of the instant state into the long-term state C; the third switch is the output gate, responsible for controlling whether the long-term state C is taken as the output of the current LSTM. The LSTM network keeps and transmits useful information for a long time through 3 gates, and filters out useless information, so that the network has long-term memory and learning capacity, can effectively solve the long-term dependence problem, and excavates time sequence characteristics among data samples.
The LSTM network is a special Recurrent Neural Network (RNN), the core of the LSTM is the neuron state, C t and C t-1 represent the neuron state at the current moment and the neuron state at the last moment, f t、it、Ot respectively represent the output results of the forgetting gate, the input gate and the output gate at the moment t,Is an internal hidden state. At the same time, an LSTM tuple contains three inputs, the current time input x t, the last time output h t-1, and the last time neuron state C t-1. Sigma represents a sigmoid activation function, and tanh represents a tanh activation function, as shown in fig. 2.
The sigmoid activation function formula is shown below:
The sigmoid activation function is able to "compress" the successive real values of the input to between 0 and 1.
The tanh activation function formula is shown below:
the tanh activation function is capable of "compressing" the continuous real values of the input to between-1 and 1.
LSTM network neurons consist of three parts: forget gate (forget gate), input gate (input gate), output gate (output gate).
Forgetting the door: the forgetting gate takes the output h t-1 of the previous layer and the sequence data x t to be input in the current layer as inputs, and obtains the output value of f t.ft in the [0,1] interval through an activation function sigmoid, which indicates the forgetting probability of the state of the previous layer, wherein 1 is "complete retention", and 0 is "complete rejection". The forget gate will decide to discard information that it deems unimportant, allowing useful information to be retained. The specific calculation formula is as follows:
ft＝σ(Wf·[ht-1,xt]+bf) (4)
An input door: the input gate comprises two parts, the first part uses a sigmoid activation function and outputs as i t, the second part uses a tanh activation function and outputs as I t is a value in the interval of [0,1], which representsTo the extent that the information in (a) is retained,Representing new information that the layer is kept.
it＝σ(Wi·[ht-1,xt]+bi) (5)
F t is the output of the forgetting gate, controls the degree to which the upper cell state C t-1 is forgotten,For two output multiplications of the input gate, it is indicated how much new information is retained, based on which the neuron state C t of this layer can be updated with new information.
Output door: the output gate is used to control how much of the neuron state of the layer is filtered. Firstly, using a sigmoid activation function to obtain O t with a value of [0,1] interval, then processing the neuron state C t through a tanh activation function, and multiplying O t, namely the output h t of the layer.
Ot＝WO·[ht-1,xt]+bO (8)
ht＝Ot*tanh(Ct) (9)
W f、Wi、WC、WO in formulas (11) - (15) represents the weight matrix of the forget gate, the input gate, the cell state, and the output gate, respectively, and b f、bi、bC、bO represents the bias vector of the forget gate, the input gate, the cell state, and the output gate, respectively.
On the basis of the LSTM network neuron structure, the invention sets the time step as 20, constructs 20 nodes for memorizing the network state at the past moment, sets the number of neurons of a neural network layer as 10, uses an Adam algorithm as an optimization algorithm, and Adam (Adaptive Moment Estimation) is a first-order optimization algorithm capable of replacing the traditional random gradient descent process, and can iteratively update the neural network weight based on training data. Adam's algorithm differs from the traditional random gradient descent. Random gradient descent keeps a single learning rate to update all weights, the learning rate does not change during training, and Adam dynamically adjusts the learning rate of each parameter by computing the first and second moment estimates of the gradient. The method has the advantages that after offset correction, each iteration learning rate has a certain range, so that the parameters are stable.
The overall LSTM predictive model expansion diagram is shown in fig. 3, with the LSTM network trained with the test set, and then model accuracy tested with the test set. The invention adopts MAE (mean absolute error) and RMSE (root mean square error) as precision evaluation indexes.
MAE: the average absolute error is the average value of the absolute errors, and can better reflect the actual condition of the predicted value error.
RMSE: root mean square error, the deviation between the observed value and the true value is measured.
In order to balance MAE and RMSE, the overall accuracy Acc of the defined model is weighted by MAE and RMSE, and the calculation formula is as follows:
Acc＝0.6*MAE+0.4*RMSE (12)
After the above-mentioned flow is completed, step S4 is performed, i.e., on the basis of the above-mentioned influencing factors, of adding the rough rolling outlet temperature at the historical moment as a new input variable for predicting the rough rolling outlet temperature of the hot rolling heating furnace slab at the present moment.
Considering the time sequence of the production data of the hot continuous rolling slab, the outlet temperature of the hot rolling slab roughing stand can be regarded as a time sequence, and each slab has a roughing outlet temperature at a corresponding moment. If the rough rolling outlet temperature of the current slab is taken as the rough rolling outlet temperature at the time t, the rough rolling outlet temperature of the previous block of the current slab is taken as the rough rolling outlet temperature at the time t-1, and the rough rolling outlet temperature of the second block of the previous block of the current slab is taken as the rough rolling outlet temperature at the time t-2.
Therefore, to predict the rough rolling outlet temperature of the slab at the current t moment, on the basis of the original 21 influencing factors, the rough rolling outlet temperatures at the t-1, t-2, … and t-N moments can be continuously added as N new input variables to be used for predicting the rough rolling outlet temperature of the slab at the current t moment together.
The initial value of N can be 10, and the specific value is optimized and determined by subsequent experimental verification.
It is noted that when the rough rolling outlet temperature at the historical moment is added as a new input variable, the temperature data of the previous slab of the current slab to be predicted may not be obtained in time due to the production time relationship, and even the temperature data of the first two slabs of the current slab to be predicted may not be obtained. Therefore, three different models need to be built according to the actual situation to predict the rough rolling outlet temperature, in particular:
(1) The rough rolling outlet temperature data of the plate blank before the plate blank to be predicted can be directly obtained.
In this case, the outlet temperatures of the current slab are predicted using the rough rolling outlet temperatures at times t-1, t-2, … and t-N as N new input variables.
(2) The outlet temperature data of the previous slab of the current slab to be predicted cannot be obtained, but the outlet temperature data of the other slabs in front can be obtained.
In this case, the rough rolling outlet temperatures at times t-2, t-3, … and t-N are used as N new input variables, and the rough rolling outlet temperature of the current slab is predicted.
(3) The outlet temperature data of the first two slabs of the current slab to be predicted cannot be obtained, but the outlet temperature data of the other slabs in front can be obtained.
In this case, the rough rolling outlet temperatures at times t-3, t-4, … and t-N are used as N new input variables, and the rough rolling outlet temperature of the current slab is predicted.
After the above procedure is completed, step S5 is executed, in which the LSTM temperature prediction model with the best prediction effect is selected as the selected model according to the accuracy of the temperature prediction model prediction. The LSTM temperature prediction model with the best prediction effect is the model with the highest precision, the N value corresponding to the model is recorded, the temperature prediction model corresponding to the N value is used as a selected model, and the selected model is applied to the LSTM temperature prediction model in the subsequent step as the LSTM temperature prediction model in the subsequent step.
After the above procedure is completed, step S6 is performed, in which the relevant process variables are ranked in importance by using Random Forest (RF) algorithm.
(1) Generating a autogenous dataset D n: for the original data set, a bootstrap resampling method is used for replacing random data extraction, and a autogenous data set D n with the same size as the original data set is formed.
(2) Generating a regression tree T n from D n: setting the total number of independent variables of original data as p, randomly selecting mtry independent variables (1-mtry-p) at each node of each tree as alternative branch variables, selecting optimal branches based on the alternative branch variables, and generating a regression tree without pruning by adopting CART (classification and regression tree) algorithm;
the CART generated regression tree algorithm is as follows:
In the input space of the training data set, recursively dividing each region into two sub-regions and deciding the output value on each sub-region, constructing a binary decision tree:
① Selecting an optimal segmentation variable j and a segmentation point s, and solving:
R1(j,s)＝{x|x(j)≤s},R2(j,s)＝{x|x(j)＞s} (14)
For the traversal variable j, recursively selecting a segmentation point s, dividing samples into two sub-areas of R 1、R2 respectively, namely dividing samples with values of the variable j smaller than the segmentation point s into an R 1 area, dividing samples with values larger than the segmentation point s into an R 1 area, and c 1、c2 is the average value of output variables y (namely R4 outlet temperature) of all samples in the R 1、R2 area respectively. The segmentation point s at which the equation (13) takes the minimum value is taken as the optimal segmentation point of the variable j.
② Traversing the variable j, selecting the variable which enables the equation (13) to obtain the minimum value and the optimal segmentation point s thereof, and taking the variable as the final optimal segmentation variable j and the segmentation point s.
③ And (3) predicting: according to the previous splitting rule (optimal splitting variable j and splitting point s), judging which leaf node (sub-region) the sample x belongs to, and taking the average value (namely c) of measured values of all samples in the leaf node (sub-region) as the predicted value of the sample x.
(3) Evaluation of Process variable (parameter) importance in Hot rolled slab production Using off-bag data when generating a self-generated dataset Using boottrap resampling method, the probability of each data in the raw dataset not being extracted isWhere k is the total amount of data in the original dataset. When k is sufficiently largeConverging toThis shows that approximately 36.8% of the data in the original dataset does not appear in the autogenous dataset during each round of random sampling, which is called out of bag (OOB) data, and that the performance of the model is estimated using the out of bag data is called OOB estimation.
An important application of the out-of-bag data is to evaluate the importance of the independent variable, the main idea is to randomly replace a certain independent variable in the OOB sample, calculate the change of the prediction error of the OOB sample before and after the replacement of the independent variable, and if the change of the error is large, the importance of the independent variable is high. The specific process is as follows:
① When random forest After the establishment is completed, for the nth regression tree T n in the forest, the corresponding out-of-bag data is recorded as OOB n, and thenWhen the original data amount N is sufficiently large, k≡0.368×n. The OOB n is predicted using the regression tree T n, resulting in a predicted mean square error MSE n for OOB n:
Where y i is the ith measured value of the response variable in OOB n, Is the i-th predictor of the response variable in OOB n.
Thus, the predicted mean square error of ntree out-of-bag data is available for the entire random forest, namely:
[MSE1…MSEn…MSEntree]
② Aiming at an independent variable X j (j is more than or equal to 1 and less than or equal to p), keeping other column data unchanged in ntree OOB samples, randomly replacing jth column data in each sample to form ntree new OOB samples, and calculating the prediction mean square error of each new OOB sample to obtain the following steps:
[ MSE j1…MSEjn…MSEjntree ], then there is the following predicted mean square error matrix for all the arguments:
③ Subtracting [ MSE 1…MSEn…MSEntree ] from the j-th row vector of the matrix, dividing the average by the standard error of the independent variable X j in the original data set to obtain the importance score (Increase of Mean Squared Error, incMSE) of the independent variable X j, namely:
where X i is the ith actual measurement of the argument X j, N is the total amount of data in the original dataset, which is the mean of the arguments X j. In particular, when equation (18) is equal to 0, the importance score IncMSE j need not be divided by the corresponding standard deviation Std, and the importance score of the corresponding argument in this case is almost always 0.
④ The importance scores of all the variables are output, and then all the variables are ranked in order of importance scores from large to small.
As can be seen from the process of the random forest algorithm, there are mainly two parameters that need to be optimized:
1) Number ntree of decision trees in random forest. A larger ntree can obtain more stable and reliable results, and ntree should be increased as much as possible, so the invention sets ntree to 5000, i.e. 5000 decision trees are included in the random forest.
2) Each tree randomly selects the number mtry of branch variables serving as alternatives at each node. Setting ntree =5000, respectively constructing mtry random forest models when different values (p is more than or equal to mtry and is equal to the total input variable number), calculating the decision coefficients (R-square, rsq) of each random forest model based on corresponding out-of-bag data, and taking mtry corresponding to the random forest model with the largest decision coefficient as mtry final value, wherein the final calculation formula is as follows:
Wherein y i is the ith measured value of the response variable, In order to respond to the mean value of the variable,K should be the total number of test set samples in response to the ith predictor of the variable.
After the above process is completed, step S7 is executed, namely, the related variables with importance scores screened by random forests arranged at the back are gradually removed, the remaining variables are input into the LSTM model to predict rough rolling outlet temperature, and the related variable combination with highest LSTM model precision is screened.
For example, importance ranking such as: the furnace row number is less than the negative tolerance of the tapping temperature, less than … and less than the end temperature of the soaking section. Then the importance score of 'in-furnace rank number' is firstly removed and arranged at the rearmost characteristic variable, the rest characteristic variable is input into the LSTM predictive model to predict the R4 outlet temperature, and the model precision is recorded. And then eliminating the characteristic variables of the two importance scores of 'furnace row number' and 'discharge temperature negative tolerance', arranging the characteristic variables at the rearmost part, inputting the rest characteristic variables into an LSTM predictive model to predict the R4 outlet temperature, and recording model accuracy. Then the final importance scores are rejected and ranked at the rearmost: after the above procedure is completed, step S8 is executed, i.e., the remaining feature variables are input to the LSTM prediction model to predict the R4 outlet temperature, and the iteration is continued for a certain number of times, for example, 800 times (step S9), and the model accuracy is recorded after the iteration is completed (step S10).
After the above-described flow is completed, step S11 is performed, i.e., the value of N is increased so that n=n+1.
When the model accuracy is significantly reduced due to excessive variables removed, step S12 is performed, i.e., the deletion of the feature variables is stopped. Finally, step S13 is executed, i.e. the model with highest accuracy is selected by comparing the accuracy of the previously recorded models, and the combination of the input variables of the model is used as the final combination of the input variables.
After the above-mentioned steps S1-S13 are all completed, the selected model is applied online.
The production process data and rough rolling temperature historical data of the hot rolling heating furnace are collected under normal working conditions, wherein the final temperature of a soaking section, the actual tapping temperature and the soaking section of the steel plate blank of the heating furnace are in section time, a predicted calculation value of a heating furnace model control system can be adopted, and final temperature of the soaking section, the actual tapping temperature and the soaking section of the steel plate blank can also be adopted; and then, adopting the LSTM rough rolling outlet temperature prediction model established in the prior art to predict the temperature of the heating furnace slab to be tapped at the outlet of the rough rolling mill on line. The predicted value can be used for evaluating whether the slab to be tapped of the heating furnace can reach the target temperature specified by the rough rolling outlet or not, and providing advance guidance for the precise control of the heating furnace.
It should be noted that, according to the actual production situation, the embodiment of the present invention enumerates the temperature prediction model applying three scenarios:
(1) The rough rolling outlet temperature data of the plate blank before the plate blank to be predicted can be directly obtained.
Model 1 is applied in this case: the rough rolling outlet temperature at the times of t-1, t-2, … and t-N is added as a rough rolling outlet temperature prediction model of a new input variable.
(2) The outlet temperature data of the previous slab of the current slab to be predicted cannot be obtained, but the outlet temperature data of the other slabs in front can be obtained.
Model 2 is applied in this case: the rough rolling outlet temperature at the times of t-2, t-3, … and t-N is added as a rough rolling outlet temperature prediction model of a new input variable.
(3) The outlet temperature data of the first two slabs of the current slab to be predicted cannot be obtained, but the outlet temperature data of the other slabs in front can be obtained.
Model 3 is applied in this case: the rough rolling outlet temperature at the times of t-3, t-4, … and t-N is added as a rough rolling outlet temperature prediction model of a new input variable.
Referring to fig. 4, in addition to the above method, the invention also discloses a temperature prediction system of a heating furnace slab to be tapped at an outlet of a roughing stand, which mainly comprises a data acquisition module, a prediction module and a sequencing screening module.
As shown in fig. 4, the data acquisition module performs steps S1 and S2 of the method of the present invention, that is, acquires the production process data of the hot rolled slab and performs preprocessing on the production process data.
S3-S5 of the method is executed by the prediction module, namely a rough rolling outlet temperature prediction model based on a long-short-period memory neural network is constructed, the rough rolling outlet temperature at the historical moment is added as a new input variable for predicting the rough rolling outlet temperature of the hot rolling heating furnace slab at the current moment, and finally the temperature prediction model with the best prediction effect is selected as a selected model;
The sorting and screening module executes S6-S13 of the method, namely, the random forest algorithm is used for sorting the importance of the production process data, the production process data with low importance scores through random forest screening are gradually removed, the rest production process data are input into the selected model of the prediction module again to predict the rough rolling outlet temperature, and the production process data combination with highest precision of the selected model is screened.
The data acquisition module, the prediction module and the sorting and screening module in the system execute the flow of the corresponding method, so that the description is omitted here.
The technical solution of the present invention is further described below by means of several examples.
The invention collects the production process data of the continuous hot continuous rolling strip steel of a certain large-scale hot continuous rolling unit in China for one year, and deletes the missing data therein, and the missing data is 65533 pieces of data in total. Each hot rolled slab data has 21 process variables in addition to the roughing outlet temperature. The invention performs specific modeling and prediction experiments on the data set. The overall flow of the rough rolling outlet temperature prediction model is shown in fig. 1.
Example 1
1. And sorting the hot-rolled plate blank data according to the production time, and eliminating the missing item data.
When data is acquired, the values of variables on a few acquisition points are easily missed, and the representation of the values of the missed acquisition points in the data is replaced by 0. As can be seen from equation (1), if the data containing 0 is not eliminated, the 0-value data will affect the normalization result, and thus the convergence of the model and the final accuracy of the model. Therefore, the data having the variable value of 0 is discarded. In addition, in consideration of the time continuity of the collected data of each strip blank, each piece of data is arranged according to the sequence of the production time, and the time correlation of the data is enhanced.
2. And constructing an LSTM prediction model and training the LSTM network weight.
The data set consisting of all variables of the hot rolled slab is divided according to training set: test set=7:3, the training set is trained by using LSTM network, the learning rate is set to 0.0006, the time step is 20, the number of neurons of the neural network layer is set to 10, the iteration coefficient is set to 800, the values of the parameter W f,bf,Wi,bi,WC,bC,WO,bO in the above formulas (9) - (14) are learned, and the network parameters are optimized by using Adam optimizer.
Modeling experiments were performed on the rough rolling outlet temperature at the moment when no history was added. The results show that: when the rough rolling outlet temperature at the historical moment is not added, the prediction accuracy index MAE of the model on the test set is 9.87, the RMSE is 12.67, and the Acc is 10.99.
3. The hot-rolled slab rough rolling outlet temperature at the current moment is predicted by adding the hot-rolled slab rough rolling outlet temperature at the historical moment as a new input variable.
And predicting the rough rolling outlet temperature of the hot-rolled slab at the current time t, and adding the rough rolling outlet temperatures of the hot-rolled slab at the times t-1, t-2, … and t-N. I.e. predicting the rough rolling outlet temperature of a certain slab, and adding the rough rolling outlet temperature of the N slabs in front of the certain slab as a new input variable. The variation of the prediction accuracy of the model on the test set with the size of N is shown in the following table 3:
table 3 model prediction accuracy over test set as a function of N value
As can be seen from table 3:
1) The MAE decrease rate of the prediction model was about 2 (the MAE of the model was 9.87 when the history was not added) and the RMSE was decreased by about 2.3 (the RMSE of the model was 12.67 when the history was not added) with respect to the rough rolling outlet temperature when the history was not added.
2) The model accuracy is highest when n=3. Therefore, the rough rolling outlet temperature of the first three slabs needing to be predicted is taken as a new input variable, and the prediction model has the best effect.
4. Based on the experimental dataset, a random forest was used to calculate the importance ranking for each influencing factor.
Setting ntree =5000, and respectively constructing random forest models when mtry takes different values (1-mtry-24), wherein the relation between mtry and the decision coefficients of the random forest models is shown in figure 2. As can be seen from fig. 2, when mtry takes a value of 9, the decision coefficient reaches the maximum, and the established random forest model has the strongest interpretation on the response variable, so the invention takes a value of mtry as 9.
FIG. 5 shows mtry the relationship of the decision coefficients to each random forest model. As shown in fig. 5, the importance ranking of the relevant variables obtained using the random forest model is shown in table 4 with ntree =5000, mtry =9.
TABLE 4 importance scoring of hot rolled slab roughing outlet temperature related variables
5. And gradually removing related variables with importance scores which are screened by random forests and are arranged at the back, forming a new training set and a new testing set by the residual variables, retraining an LSTM prediction model and testing the precision of the new model, and repeating the process until the related variable combination with the highest LSTM model precision is screened, and obtaining the LSTM rough rolling R4 outlet temperature prediction model with the highest precision. The specific process is as follows:
1) M variables with importance scores arranged at the rearmost are removed, M=1 is started, the data composed of the rest variables are divided into a training set and a testing set, and the training set, the testing set and the testing set are divided according to the training set, the testing set and the testing set=7:3.
2) Putting the training set into an LSTM network for training, setting the learning rate to be 0.0006, setting the time step to be 20, setting the neuron number of the neural network layer to be 10, setting the iteration coefficient to be 800, testing the model precision by using the test set after training, storing the model precision, and then enabling M=M+1. The model accuracy Acc is weighted by MAE and RMSE, and the calculation formula is shown in formula (12).
3) Repeating the steps 1) and 2), and stopping the experiment if the model precision is obviously and continuously reduced along with the increase of M.
4) And comparing the precision of all the models, and screening out the model with the highest precision and the related variable combination thereof.
The M variables with importance scores at the last are removed, and as M becomes larger, the experimental results are shown in the following table 5:
table 5 LSTM predictive model test set accuracy variation
From the above table, it can be seen that:
1) The LSTM prediction model is most accurate when the importance scores are culled in the last 9 variables. Namely, the input variables are: the temperature of the outlet at the time t-1, the temperature of the outlet at the time t-2, the temperature of the outlet at the time t-3, the end temperature of the soaking section, the actual thickness of the material, the actual tapping temperature, the heating time of the soaking section, the end temperature of the second heating section, the length of the material, the number of the heating furnace, the heating time of the first heating section, the time in the furnace, the heating time of the second heating section, the actual charging temperature and the heating time of the preheating section are 15 variables.
The model accuracy at this time is: mae=7.45 (representing an average prediction error of 7.45 ℃ for the rough rolling R4 outlet temperature), rmse=9.98, acc=8.46. When the number of variables to be eliminated exceeds 9, the model precision gradually decreases.
2) Compared with the model accuracy without eliminating any variables, after eliminating the importance scores and ranking the last 9 variables, the MAE is reduced by 0.21, the RMSE is reduced by 0.25, and the model accuracy is improved, which shows that the model can better explain the relation between the rough rolling outlet temperature and each variable.
Example 2:
Considering that the rough rolling outlet temperature of the hot rolled slab at the time of addition history is taken as a new input variable, the rough rolling outlet temperature of the first three slabs to be predicted is theoretically added as a new input variable, but due to the time relationship, the outlet temperature data of the previous slab of the current slab to be predicted may not be taken in time, and the outlet temperature data can only be taken from the second slab in front of the current slab to be predicted. At this time, different model predictions are established according to actual conditions to predict the rough rolling outlet temperature of the hot rolled slab.
In this case, since the outlet temperature data of the previous slab of the current slab to be obtained cannot be obtained, the outlet temperatures of the second, third and fourth slabs in front of the current slab to be obtained can be used as new input variables, and the outlet temperature of the current slab can be predicted. Namely, the input variables are: the temperature of the outlet at the time t-2, the temperature of the outlet at the time t-3, the temperature of the outlet at the time t-4, the end temperature of the soaking section, the actual thickness of the material, the actual tapping temperature, the heating time of the soaking section, the end temperature of the second heating section, the length of the material, the number of the heating furnace, the heating time of the first heating section, the time in the furnace, the heating time of the second heating section, the actual charging temperature and the heating time of the preheating section are 15 variables.
The LSTM predictive model network structure remained consistent with example 1; the super parameters also remain the same: the number of hidden layer neurons is 10, the learning rate is 0.0006, the time step is 20, and the training iteration number is 800.
In this case, the model accuracy is reduced compared to case (1) because the historical outlet temperature is taken as a new input variable by separating one slab, and the RF-LSTM model predicts the accuracy on the test set as: mae=8.10, rmse=10.80.
Example 3:
Considering that the outlet temperature data of the previous slab of the current slab to be predicted cannot be obtained in time due to the time relationship, even the outlet temperature data of the two previous slabs of the current slab to be predicted cannot be obtained, the outlet temperature data can only be obtained from the third slab in front of the current slab to be predicted. At this time, different model predictions are established according to actual conditions to predict the rough rolling outlet temperature of the hot rolled slab.
In this case, since the outlet temperature data of the first two slabs of the current slab to be processed cannot be obtained, the outlet temperatures of the third, fourth and fifth slabs in front of the current slab to be processed can be used as new input variables, and the outlet temperature of the current slab can be predicted. Namely, the input variables are: the temperature of the outlet at the time t-2, the temperature of the outlet at the time t-3, the temperature of the outlet at the time t-4, the end temperature of the soaking section, the actual thickness of the material, the actual tapping temperature, the heating time of the soaking section, the end temperature of the second heating section, the length of the material, the number of the heating furnace, the heating time of the first heating section, the time in the furnace, the heating time of the second heating section, the actual charging temperature and the heating time of the preheating section are 15 variables.
The LSTM predictive model network structure remained consistent with example 1; the super parameters also remain the same: the number of hidden layer neurons is 10, the learning rate is 0.0006, the time step is 20, and the training iteration number is 800.
In this case, since the historical outlet temperature is taken as a new input variable by separating two slabs, the model accuracy is reduced compared with the cases (1) and (2), and the prediction accuracy of the RF-LSTM model on the test set is as follows: mae=8.47, rmse=11.23.
Finally, to illustrate the beneficial effects of the present invention. The present invention also compares the MLP (multi-layer perceptron) network prediction model with the SVR (support vector regression) prediction model, and the comparison results are shown in table 6 below. The experimental result shows that the LSTM prediction model and the RF-LSTM prediction model (the LSTM prediction model based on the random forest screening related variable) have obviously better precision than the MLP and SVR algorithm, which proves that the LSTM prediction model based on the random forest screening related variable has obvious advantages and pertinence in the aspect of predicting the rough rolling outlet temperature of the hot rolled slab.
Table 6 comparison of different algorithm detection accuracy
Advanced technology
1. In the hot continuous rolling production process, each slab immediately follows the last slab to carry out the same working procedure under almost the same conditions, and the hot rolling production data collected by the computer and the sensor are also collected from one slab to another slab, and have sequential property in time. According to the characteristics, an LSTM (long and short term memory neural network) model is selected, and the temperature of a rough rolling outlet of a heating furnace tapping plate blank is predicted. Because LSTM is well suited for addressing problems highly related to time series due to its unique design structure.
And as can be seen from table 6, the LSTM model is significantly more accurate than the MLP and SVR models in predicting the rough rolling outlet temperature. Obviously, the LSTM model is more targeted in processing hot rolled production data.
2. The hot rolling production data collected by the computer and the sensor has approximately 30 variables, and excessive variables inevitably introduce a large amount of noise and redundancy factors to influence the accuracy of the model. Therefore, some irrelevant variables, even disturbance variables, should be eliminated, useful variables are extracted, and model accuracy and interpretation strength are improved. According to the method, importance scoring is carried out on all variables by using a random forest algorithm, the importance of each variable on the predicted rough rolling outlet temperature is evaluated, and therefore partial variables with lower correlation are screened out, and more accurate rough rolling outlet temperature prediction is realized.
In addition, as shown in tables 4 and 5, after 9 variables with lower importance scores are screened out by using random forests, MAE (mean absolute error) is reduced by 0.21, RMSE (root mean square error) is reduced by 0.25, and the prediction accuracy of a visible model is obviously improved, so that the advancement of the invention is proved.
It will be appreciated by persons skilled in the art that the above embodiments are provided for illustration only and not for limitation of the invention, and that variations and modifications of the above described embodiments are intended to fall within the scope of the claims of the invention as long as they fall within the true spirit of the invention.
Claims (7)
1. A method for predicting a rough rolling outlet temperature of a heating furnace tapping plate blank, comprising the steps of:
Collecting production process data of a hot-rolled slab;
Preprocessing the production process data;
constructing a rough rolling outlet temperature prediction model based on a long-term and short-term memory neural network;
adding the rough rolling outlet temperature at the historical moment as a new input variable for predicting the rough rolling outlet temperature of the hot rolling heating furnace slab at the current moment;
Selecting a temperature prediction model with the best prediction effect as a selected model;
Ranking the importance of the production process data using a random forest algorithm;
Gradually removing production process data with low importance scores through random forest screening, re-inputting the rest production process data into the selected model to predict rough rolling outlet temperature, and recording the precision of a short-term selected model until screening the production process data combination with highest precision of the selected model;
The rough rolling outlet temperature corresponding to each slab is used as a time sequence, and the rough rolling outlet temperatures of continuous N slabs before the current slab are intercepted as new input variables;
According to the prediction precision of the temperature prediction model, selecting an N value corresponding to the model with highest precision, and taking the temperature prediction model corresponding to the N value as a selected model;
Repeating the above operation until the precision of the selected model is obviously and continuously lowered, and stopping deleting the production process data;
comparing the precision of all the selected models, and selecting the selected model with the highest precision;
and combining the production process data corresponding to the selected model with the highest precision as final production process data.
2. A method for predicting rough rolling outlet temperature of a heating-furnace tapping plate blank according to claim 1, wherein the production process data comprises:
The actual width of the material, the actual thickness of the material, the length of the material, the actual weight of the material, the weight of the charged plate blank, the number of heating furnaces, the number of rows in the furnace, the time in the furnace, the actual charging temperature, the actual discharging temperature, the target discharging temperature value, the negative discharging temperature tolerance, the positive discharging temperature tolerance, the end temperature of the preheating section, the heating time of the first heating section, the heating time of the second heating section, the heating time of the soaking section, the end temperature of the first heating section, the end temperature of the second heating section, the end temperature of the soaking section and the rough rolling outlet temperature.
3. A method for predicting the rough rolling outlet temperature of a heating-furnace tapping plate blank according to claim 1, wherein said pretreatment comprises:
Removing the strip steel records with missing items in the production process data;
Sequencing each piece of production process data from first to last according to the production time;
And carrying out maximum and minimum normalization processing on the production process data.
4. A method for predicting the rough rolling outlet temperature of a heating-furnace tapping plate blank according to claim 1, wherein:
If the rough rolling outlet temperature data of the previous slab of the current slab to be predicted can be measured, taking the rough rolling outlet temperatures at the times of t-1, t-2, … and t-N as N new input variables;
If the rough rolling outlet temperature data of the second slab before the current slab to be predicted can be measured, taking the rough rolling outlet temperatures at the times of t-2, t-3, … and t-N as new input variables;
If the rough rolling outlet temperature data of the third slab before the current slab to be predicted can be measured, the rough rolling outlet temperatures at the times of t-3, t-4, … and t-N are taken as new input variables.
5. A system for predicting a rough rolling outlet temperature of a heating furnace tapping plate blank, comprising:
The data acquisition module acquires production process data of the hot-rolled plate blank and preprocesses the production process data;
The prediction module is used for constructing a rough rolling outlet temperature prediction model based on a long-short-period memory neural network, adding the rough rolling outlet temperature at the historical moment as a new input variable, predicting the rough rolling outlet temperature of the hot rolling heating furnace slab at the current moment, and finally selecting a temperature prediction model with the best prediction effect as a selected model;
The sorting and screening module uses a random forest algorithm to sort the importance of the production process data, gradually eliminates the production process data with low importance scores through random forest screening, and re-inputs the rest of the production process data into a selected model of the prediction module for predicting rough rolling outlet temperature until screening the production process data combination with highest precision of the selected model;
The prediction module takes the rough rolling outlet temperature corresponding to each slab as a time sequence, intercepts the rough rolling outlet temperatures of continuous N slabs before the current slab as new input variables, selects an N value corresponding to a model with highest precision according to the prediction precision of a temperature prediction model, and takes the temperature prediction model corresponding to the N value as a selected model;
The sorting and screening module eliminates production process data with low importance scores after random forest screening, inputs the residual variables into a selected model for predicting rough rolling outlet temperature, and records the accuracy of a short-term selected model; and the sorting and screening module repeats the above operation until the precision of the selected model is obviously and continuously lowered, and stops deleting the production process data, at the moment, the precision of all the selected models is compared, the selected model with the highest precision is selected, and finally the production process data corresponding to the selected model with the highest precision is used as the final production process data combination.
6. The system for predicting the rough rolling outlet temperature of a heating-furnace tapping plate blank according to claim 5, wherein the production process data comprises:
The actual width of the material, the actual thickness of the material, the length of the material, the actual weight of the material, the weight of the charged plate blank, the number of heating furnaces, the number of rows in the furnace, the time in the furnace, the actual charging temperature, the actual discharging temperature, the target discharging temperature value, the negative discharging temperature tolerance, the positive discharging temperature tolerance, the end temperature of the preheating section, the heating time of the first heating section, the heating time of the second heating section, the heating time of the soaking section, the end temperature of the first heating section, the end temperature of the second heating section, the end temperature of the soaking section and the rough rolling outlet temperature.
7. A system for predicting the rough rolling outlet temperature of a heating-furnace tapping plate blank according to claim 5, wherein:
The data acquisition module eliminates strip steel records with missing items in production process data, sorts each piece of production process data from first to second according to production time, and then carries out maximum and minimum normalization processing on the production process data.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110126990.3A CN114798763B (en) | 2021-01-29 | 2021-01-29 | Method and system for predicting rough rolling outlet temperature of heating furnace tapping plate blank |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110126990.3A CN114798763B (en) | 2021-01-29 | 2021-01-29 | Method and system for predicting rough rolling outlet temperature of heating furnace tapping plate blank |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN114798763A CN114798763A (en) | 2022-07-29 |
| CN114798763B true CN114798763B (en) | 2024-07-12 |
Family
ID=82525924
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202110126990.3A Active CN114798763B (en) | 2021-01-29 | 2021-01-29 | Method and system for predicting rough rolling outlet temperature of heating furnace tapping plate blank |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN114798763B (en) |
Families Citing this family (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116822367A (en) * | 2023-07-03 | 2023-09-29 | 东北大学 | IF steel mechanical property prediction method based on LSTM model |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108687140A (en) * | 2018-05-11 | 2018-10-23 | 鞍钢股份有限公司 | Self-learning control method for slab temperature of hot rolling heating furnace |
| CN110909508A (en) * | 2019-10-29 | 2020-03-24 | 中国石油化工股份有限公司 | Heating furnace temperature field real-time prediction method based on convolution long and short term memory network |
Family Cites Families (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| KR20020052723A (en) * | 2000-12-26 | 2002-07-04 | 이구택 | coiling temperature control method of hot strip using learning method |
| KR100832975B1 (en) * | 2001-12-24 | 2008-05-27 | 주식회사 포스코 | System for optimizating target discharge temperature of reheating furnace |
| JP4997263B2 (en) * | 2009-03-19 | 2012-08-08 | 株式会社日立製作所 | Hot rolling simulation apparatus and rolling history simulation method |
| JP5610819B2 (en) * | 2010-04-02 | 2014-10-22 | 株式会社神戸製鋼所 | Steel temperature prediction method considering transformation heat generation |
| JP6435234B2 (en) * | 2015-05-20 | 2018-12-05 | 株式会社日立製作所 | Hot roll finishing mill outlet temperature control device and control method thereof |
| EP3969966A4 (en) * | 2019-05-17 | 2023-06-21 | Tata Consultancy Services Limited | Method and system for adaptive learning of models for manufacturing systems |
| CN112139261B (en) * | 2019-06-27 | 2022-08-16 | 宝山钢铁股份有限公司 | Target tapping temperature prediction control method for hot rolling heating furnace |
-
2021
- 2021-01-29 CN CN202110126990.3A patent/CN114798763B/en active Active
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108687140A (en) * | 2018-05-11 | 2018-10-23 | 鞍钢股份有限公司 | Self-learning control method for slab temperature of hot rolling heating furnace |
| CN110909508A (en) * | 2019-10-29 | 2020-03-24 | 中国石油化工股份有限公司 | Heating furnace temperature field real-time prediction method based on convolution long and short term memory network |
Also Published As
| Publication number | Publication date |
|---|---|
| CN114798763A (en) | 2022-07-29 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108764517B (en) | Method, equipment and storage medium for predicting change trend of silicon content in molten iron of blast furnace | |
| CN111241657B (en) | Cold-rolled strip steel plate shape prediction method based on deep learning | |
| CN109002686B (en) | Multi-grade chemical process soft measurement modeling method capable of automatically generating samples | |
| CN104200268B (en) | PSO (Particle Swarm Optimization) extremity learning machine based strip steel exit thickness predicting method | |
| CN109992921B (en) | On-line soft measurement method and system for thermal efficiency of boiler of coal-fired power plant | |
| CN115456460A (en) | Multi-quality index output prediction method and system for tobacco shred loosening and conditioning process | |
| CN110739031B (en) | Supervised prediction method and device for metallurgical sintering process and storage medium | |
| CN112100745B (en) | Automobile girder steel mechanical property prediction method based on LDA theory | |
| CN111144552B (en) | Multi-index grain quality prediction method and device | |
| CN114897227B (en) | Multi-steel mechanical property forecasting method based on improved random forest algorithm | |
| CN114282443B (en) | Residual service life prediction method based on MLP-LSTM supervised joint model | |
| CN113011796A (en) | Edible oil safety early warning method based on hierarchical analysis-neural network | |
| CN113128124B (en) | Multi-grade C-Mn steel mechanical property prediction method based on improved neural network | |
| CN113204925A (en) | Strip steel thickness prediction method for improving whale optimization least square support vector machine | |
| CN114798763B (en) | Method and system for predicting rough rolling outlet temperature of heating furnace tapping plate blank | |
| CN114015825A (en) | Method for monitoring abnormal state of blast furnace heat load based on attention mechanism | |
| CN112614021A (en) | Tunnel surrounding rock geological information prediction method based on built tunnel information intelligent identification | |
| CN115879369A (en) | Coal mill fault early warning method based on optimized LightGBM algorithm | |
| CN110110447B (en) | Method for predicting thickness of strip steel of mixed frog leaping feedback extreme learning machine | |
| CN115096357A (en) | Indoor environment quality prediction method based on CEEMDAN-PCA-LSTM | |
| Zheng et al. | Dynamic rolling force prediction of reversible cold rolling mill based on BP neural network with improved PSO | |
| CN117312770A (en) | Aero-engine residual life prediction method based on multi-level graph feature fusion | |
| CN116689502A (en) | Roll bending force parameter setting method of finishing mill group process of hot continuous rolling mill based on SA-SCSO-SCN neural network | |
| CN109214513B (en) | Solid-liquid waste intelligent coupling treatment method based on adaptive deep belief network | |
| CN109146007B (en) | Solid waste intelligent treatment method based on dynamic deep belief network |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |