A data-driven method for calculating the adaptive coefficient of heat flux density

Abstract

A data-driven calculation method for the adaptive coefficient of heat flux density according to the present invention includes: collecting cooling process parameters, strip information parameters, and cooling water temperature as input feature parameters, taking the adaptive coefficient of heat flux density as the output feature parameter, and forming an original data set from the input feature parameters and the output feature parameter; performing data cleaning on the original data set and dividing it into a training set, a test set, and a validation set; constructing a stacking model for the adaptive coefficient of heat flux density, and using a chaotic Harris hawk optimization algorithm to obtain the optimal number and type of base learners; using a chaotic Harris hawk optimization algorithm to optimize the hyperparameters of the optimal base learners; inputting the training set into the stacking model for training; using the test set to verify the generalization performance of the stacking model; collecting the current strip information parameters, and using the tested stacking model to predict the adaptive coefficient of heat flux density of the current strip.

Description

Technical Field

[0001] The present invention belongs to the technical field of hot rolling, and relates to a method for calculating an adaptive coefficient of heat flux density based on data driving. Background Art

[0002] Post-rolling cooling is an important process means in the rolling production of hot-rolled steel. By controlling the temperature and cooling path of the strip during the cooling process, hot-rolled products with expected microstructures and mechanical properties can be obtained. With the continuous improvement of the market's quality requirements for hot-rolled products, the temperature control accuracy of the strip during the cooling process has also been improved. However, post-rolling cooling is a process involving multiple variables and factors, and these variables have characteristics such as strong coupling, nonlinearity, and time-variation. It is very difficult to quantitatively describe this process with high precision only by a simplified heat transfer model. Therefore, the temperature control of the post-rolling cooling process has always been a key and difficult problem in this industry. Especially for the control of the strip head temperature, due to the inability to perform feedback control, the control accuracy of the head temperature is even lower. Introducing a model correction coefficient in actual production has become a simple and effective method. Since this coefficient generally directly participates in the calculation of the heat transfer model, for strips of different steel grades and specifications, the magnitude of the initial value of this coefficient directly determines the calculation accuracy of the heat transfer model, and thus determines the temperature control accuracy of the strip head.

[0003] In order to improve the calculation accuracy of the heat transfer model during the post-rolling cooling process, domestic scholars have successively proposed different calculation methods for the model correction coefficient for different control systems. Chinese Patent "CN105032951B A Control Method for Improving the Accuracy of the Ultra-Fast Cooling Temperature Model and the Self-Learning Efficiency" comprehensively considers the influence of factors such as steel grade physical properties, steel plate specifications, medium water temperature, steel plate temperature, aging, etc. on the heat transfer coefficient. Construct a multi-dimensional space relationship, which effectively describes the connection between the heat transfer coefficients under the action of different influencing factors. Quickly and accurately predict the value of the heat transfer coefficient under the target influencing conditions, and finally achieve the precise control of the temperature model. Chinese Patent "CN111814861B An Online Cooling Control Method Based on a Dual Self-Learning Model" combines the characteristics of an unsupervised space clustering model and a supervised deep neural network prediction model to establish a system structure with parallel dual self-learning models and shared weights. It realizes fast combined self-learning under the condition of low-cost data volume, and improves the robustness and learning efficiency of the overall cooling control system. Chinese Patent "CN108446454B A Method for Improving the Calculation Accuracy of the Layer Cooling Model Setting" considers the influence of the threading speed on heat transfer, and establishes a threading speed influence coefficient function according to the steel grade of the strip, the cooling strategy, the target thickness layer, and the target coiling temperature layer. The threading speed influence coefficient is used to correct the water cooling self-learning coefficient. Effectively solves the problem of low calculation accuracy of the layer cooling model setting. Chinese Patent "CN104942025B A Compensation Method for the Genetic Coefficient after Stopping Rolling in the Hot Rolling Coiling Temperature Model" aims at the compensation method for the genetic coefficient after stopping rolling in the hot rolling coiling temperature model, which can appropriately compensate the genetic coefficient when the rolling line stops rolling. When the rolling line resumes rolling, the coiling temperature can be basically controlled within the tolerance range, thereby reducing quality losses.

[0004] The above research has improved the calculation accuracy of the heat transfer model to a certain extent, but basically only considers the influence of a single factor on heat transfer. The temperature change of the strip in the cooling zone is affected by various factors such as rolling speed, steel grade, specifications, and cooling water temperature, pressure, cooling mode, etc. Especially as the market orders tend to be small-batch and diversified, the steel grade and specifications change frequently in production, and the impact on the calculation accuracy of the heat transfer model is more significant. However, the above inventions do not consider the steel grade, specifications, cooling water state, and cooling mode. Summary of the Invention

[0005] To solve the above technical problems, the object of the present invention is to provide a data-driven calculation method for the adaptive coefficient of heat flux density.

[0006] The present invention provides a data-driven calculation method for the adaptive coefficient of heat flux density, including:

[0007] Step 1: Collect the cooling process parameters, strip information parameters, and cooling water temperature as input feature parameters, and use the heat flux density adaptive coefficient as the output feature parameter to form the original dataset from the input and output feature parameters;

[0008] Step 2: Clean the original dataset and divide the cleaned dataset into a training set, a test set, and a validation set according to the ratio of 7:2:1;

[0009] Step 3: Build a heat flux density adaptive coefficient stacking model and use the chaotic Harris hawk optimization algorithm to obtain the optimal number and type of base learners;

[0010] Step 4: Use the chaotic Harris hawk optimization algorithm to optimize the hyperparameters of the obtained optimal base learners to obtain the optimal hyperparameter combination;

[0011] Step 5: Input the training set into the heat flux density adaptive coefficient stacking model for training to obtain the trained heat flux density adaptive coefficient stacking model;

[0012] Step 6: Use the test set to verify the generalization performance of the heat flux density adaptive coefficient stacking model;

[0013] Step 7: Collect the current strip information parameters and use the tested heat flux density adaptive coefficient stacking model to predict the heat flux density adaptive coefficient of the current strip.

[0014] A data-driven calculation method for the heat flux density adaptive coefficient of the present invention has the following beneficial effects:

[0015] The calculation method of the present invention is based on a large amount of historical production data, combines the chaotic Harris hawk optimization algorithm and the stacking model, and establishes a prediction model for the heat flux density adaptive coefficient. Compared with the case-based reasoning model used on-site, the model proposed by the present invention has higher accuracy, improves the calculation accuracy of the heat transfer model in the post-rolling cooling process, and ensures the control accuracy and quality of the strip head temperature. Brief Description of the Drawings

[0016] Figure 1 is a flowchart of a data-driven calculation method for the heat flux density adaptive coefficient of the present invention;

[0017] Figure 2 is a change diagram of the first objective function in the optimization process for determining the optimal structure of the stacking model in the embodiment of the present invention;

[0018] Figure 3It is a change diagram of the second objective function in the optimization process for determining the hyperparameters of the stacking model in the embodiments of the present invention;

[0019] Figure 4 It is a comparison diagram of temperature control of the stacking model and the on-site model of the present invention within the first 50 m of the strip head;

[0020] Figure 5 It is a comparison diagram of temperature control of the stacking model and the on-site model of the present invention within the first 100 m of the strip head. Detailed implementation manners

[0021] The following combines the accompanying drawings and embodiments to further describe in detail the specific implementation manners of the present invention. The following embodiments are used to illustrate the present invention, but are not used to limit the scope of the present invention.

[0022] As Figure 1 shown, a data-driven method for calculating the adaptive coefficient of heat flux density of the present invention includes:

[0023] Step 1: Collect cooling process parameters, strip information parameters, and cooling water temperature as input feature parameters, and use the adaptive coefficient of heat flux density as the output feature parameter. The original data set is composed of the input feature parameters and the output feature parameter.

[0024] The cooling process parameters include: the measured temperature at the final rolling, the measured temperature in the middle, the measured temperature at coiling, the rolling speed, the upper valve cooling mode, the lower valve cooling mode, the upper starting valve position, the lower starting valve position, and the cooling direction. The strip information parameters include: the mass percentage content of chemical components, the strip width, and the strip thickness. The 28 main factors affecting heat transfer, including 9 cooling process parameters, 18 strip information parameters, and the cooling water temperature, are used as input feature parameters. The adaptive coefficient of heat flux density is used as the output feature parameter, that is, the sample label.

[0025] The specific content of Step 1 is:

[0026] Step 1.1: Collect the characteristic parameters of the strip from the hot strip rolling production site. Since the main function of the present invention is for the prediction of the strip head, the data collection position should be as close as possible to the strip head but also need to consider the influence of the poor strip shape at the strip head on the measured value. Considering the relationship between the two, the characteristic parameter collection position of the strip is calculated according to the following formula:

[0027] ;

[0028] ;

[0029] In the formula, l bis the starting position of data acquisition, i.e., the length from the strip head, in m; l c is the distance from the pyrometer before coiling to the coiler, in m; Δ b is a fixed constant; h is the strip thickness, in m; l e is the ending position of data acquisition, which refers to the length from the strip head, in m; N is the number of sampling points; v is the strip running speed, in m / s; Δ t is the sampling period, in s.

[0030] In this embodiment, l c is the distance from the pyrometer before coiling to the coiler, with a value of 26.22 m. Δ b takes a value of 0.08. N takes a value of 50. Δ t takes a value of 0.2 s. Assume that the thickness of a certain strip is 0.004 m and the running speed is constantly 7 m / s. Then the distances from the strip head to the sampling start position and the ending position are respectively:

[0031] ;

[0032] ;

[0033] Step 1.2: Calculate the characteristic parameters of the strip according to the following formula:

[0034] ;

[0035] In the formula, x j is the value of the j th characteristic of the strip; x k,j is the k th sampling point's j th characteristic; M is the total number of input characteristic parameters and output characteristic parameters, with a value of 29.

[0036] Step 1.3: Repeat Step 1.1 and Step 1.2 to collect the cooling process parameters, information parameters, and cooling water temperature of different strips, and then multiple samples can be obtained, thus constituting the original dataset.

[0037] In this embodiment, the number of samples in the original dataset is 60,000.

[0038] Step 2: Clean the original dataset and divide the cleaned dataset into a training set, a test set, and a validation set according to the ratio of 7:2:1. Specifically:

[0039] Step 2.1: Since the data collected from the hot strip rolling production site contains a large amount of noisy data, it is necessary to perform preliminary data cleaning on the original dataset to eliminate obviously unrealistic abnormal data. Specifically, eliminate abnormal data according to the following formula:

[0040] ;

[0041] In the formula, x max and x min are the upper and lower limits of the characteristic parameters respectively; x i,j is the i th j characteristic of the n th sample;

[0042] In this embodiment, the values of the upper and lower limits of each characteristic parameter are shown in Table 1:

[0043] Table 1 Setting range of characteristic parameters

[0044]

[0045] Step 2.2: Further clean the data cleaned in Step 2.1 using the 3-sigma rule to eliminate the characteristic parameter samples that do not conform to the following formula, so that the distribution of each characteristic parameter in the original dataset is more in line with the normal distribution:

[0046] ;

[0047] In the formula, is the average value of the j th characteristic, σ is the standard deviation of the j th characteristic.

[0048] Step 2.3: Divide the cleaned dataset into a training set, a test set, and a validation set according to the ratio of 7:2:1. The training set is used for stacking model training, the test set is used for performance evaluation of the stacking model, and the validation set is used for validation when determining the hyperparameters to be optimized during the training process.

[0049] In this embodiment, after Steps 2.1 and 2.2, the number of samples in the original dataset changes from 60,000 to 54,518. Then, after dividing it into a training set, a test set, and a validation set according to the ratio of 7:2:1, the number of samples in each dataset is 38,163, 10,903, and 5,452 respectively.

[0050] Step 3: Construct a stacking model for the heat flux density adaptive coefficient, and use the chaotic Harris hawk optimization algorithm to obtain the optimal number and type of base learners, specifically as follows:

[0051] Step 3.1: Use the extremely randomized tree regressor (ETR), random forest (RF), extreme gradient boosting (XGBoost), categorical gradient boosting (CatBoost), gradient boosting decision tree (GBDT), histogram gradient boosting regressor (HGBR), bagging method (Bagging), adaptive boosting algorithm (AdaBoost), light gradient boosting machine (LightGBM), and decision tree regressor (DTR) as candidate base learners.

[0052] Step 3.2: Use linear regression (LR) as the meta-learner.

[0053] Step 3.3: Take the number and type of base learners as the parameters to be optimized.

[0054] In this embodiment, when determining the optimal stacking model structure, the parameters to be optimized and their search spaces are shown in Table 2:

[0055] Table 2 Parameters to be optimized and search spaces

[0056]

[0057] Step 3.4: Use the following formula as the first objective function to be optimized:

[0058] ;

[0059] In the formula, num is the number of base learners; type is the type of each base learner; m is the number of samples in the validation set; and y i are respectively i the model prediction value and the actual value (label) of the

[0060] th sample.

[0061] In this embodiment, when determining the optimal structure of the stacking model, the change of the value of the first objective function with the iteration process is as shown in Figure 2As shown, the parameter combination corresponding to the minimum value of the first objective function is the optimal parameter, as shown in Table 3:

[0062] Table 3 Optimal Parameter Values

[0063]

[0064] Step 4: Use the chaotic Harris hawk optimization algorithm to optimize the hyperparameters of the obtained optimal base learner to obtain the optimal hyperparameter combination, specifically:

[0065] Step 4.1: Determine the hyperparameters to be optimized and the optimization interval according to the type of the optimal base learner obtained in Step 3.

[0066] In this embodiment, the base learners are ETR, RF, CatBoost, and Bagging, and the meta-learner is LR. The key hyperparameters and search spaces of each learner are shown in Table 4:

[0067] Table 4 Hyperparameters and Search Spaces

[0068]

[0069] Step 4.2: Use the coefficient of determination as the second objective function to be optimized:

[0070] ;

[0071] In the formula, R 2 is the coefficient of determination, and the closer it is to 1, the better the fitting effect of the model; m is the number of samples in the validation set; , y i and are the model prediction value, actual value, and average value of the actual values of the i th sample, respectively.

[0072] Step 4.3: Aim to obtain the maximum value of the second objective function, and use the chaotic Harris hawk optimization algorithm to optimize the hyperparameters in Step 4.1 to obtain the optimal hyperparameter combination.

[0073] In this embodiment, during the hyperparameter optimization process, the change of the value of the second objective function with the iteration process is as Figure 3 shown. The parameter combination corresponding to the minimum value of the second objective function is the optimal parameter, as shown in Table 5:

[0074] Table 5 Hyperparameters and Search Spaces

[0075]

[0076] Step 5: Input the training set into the stacking model of the heat flux density adaptive coefficient for training to obtain a trained stacking model of the heat flux density adaptive coefficient.

[0077] Step 6: Verify the generalization performance of the stacking model of the heat flux density adaptive coefficient using the test set, specifically:

[0078] Use the coefficient of determination R 2 , mean square error MSE, and mean absolute error MAS as evaluation indicators to verify the generalization performance of the stacking model of the heat flux density adaptive coefficient obtained in Step 4.

[0079] ;

[0080] ;

[0081] where p is the number of samples in the test set.

[0082] Step 7: Collect the current strip information parameters and use the tested stacking model of the heat flux density adaptive coefficient to predict the heat flux density adaptive coefficient of the current strip.

[0083] In this embodiment, the comparison of the prediction results of each base learner, the prediction results of the on-site case-based reasoning model, and the prediction results of the stacking model of the present invention is shown in Table 6:

[0084] Table 6 Model Performance Evaluation

[0085]

[0086] After adopting the stacking model established by the present invention, the temperature control accuracy within the first 50 m and 100 m at the head of the production line is as Figure 4 and Figure 5 shown. It can be clearly seen from the figure that compared with the case-based reasoning method previously used on this production line, the stacking model proposed by the present invention has a much higher prediction accuracy, resulting in a significant improvement in the temperature control accuracy at the head. Among them, 88% and 93% of the samples within the first 50 m and 100 m at the head have a control accuracy within ±20°C, which is 26% and 24% higher than before, respectively.

[0087] The above are only the preferred embodiments of the present invention and are not intended to limit the idea of the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A data-driven heat flux adaptive coefficient calculation method, characterized in that: include: Step 1: Collect cooling process parameters, strip information parameters and cooling water temperature as input characteristic parameters, and use heat flux adaptive coefficient as output characteristic parameter. The input characteristic parameters and output characteristic parameters constitute the original data set. Step 2: Clean the original data set and divide the cleaned data set into training set, test set and validation set in a ratio of 7:2:1; Step 3: Construct a heat flux adaptive coefficient stacking model and use the chaos Harris Eagle optimization algorithm to obtain the optimal number and type of base learners, specifically: Step 3.1: Construct a candidate pool of base learners: extreme random tree regressor, random forest, extreme gradient boosting, categorical gradient boosting, gradient boosted decision tree, histogram gradient boosting regressor, bagging, adaptive boosting algorithm, lightweight gradient boosting machine and decision tree regressor as candidate base learners; Step 3.2: Use linear regression as a meta-learner; Step 3.3: The number and type of base learners are taken as the parameters to be optimized; Step 3.4: The following formula is used as the first objective function to be optimized: In the formula, num is the number of base learners; type is the type of each base learner; m is the number of samples in the validation set; and i are the model predicted value and actual value of the i-th sample respectively; Step 3.5: Taking the validation set as input and taking the minimum value of the first objective function as the goal, the Chaos Harris Eagle Optimization Algorithm is used to optimize the parameters to be optimized in step 3.3, and the optimal parameter combination is obtained, which is the optimal structure of the heat flux adaptive coefficient stacking model; Step 4: Use the Chaos Harris Eagle Optimization Algorithm to optimize the hyperparameters of the obtained optimal base learner to obtain the optimal hyperparameter combination; Step 5: Input the training set into the heat flux density adaptive coefficient stacking model for training to obtain the trained heat flux density adaptive coefficient stacking model; Step 6: Use the test set to verify the generalization performance of the heat flux adaptive coefficient stacking model; Step 7: Collect the current strip information parameters and use the tested heat flux density adaptive coefficient stacking model to predict the current heat flux density adaptive coefficient of the strip.

2. The method for calculating heat flux adaptive coefficient based on data drive according to claim 1, characterized in that: The cooling process parameters include: final rolling measured temperature, intermediate measured temperature, coiling measured temperature, rolling speed, upper valve cooling mode, lower valve cooling mode, upper start valve position, lower start valve position and cooling direction; The strip steel information parameters include: mass percentage of chemical composition, strip steel width and strip steel thickness.

3. The method for calculating heat flux adaptive coefficient based on data drive according to claim 1, characterized in that: The step 1 is specifically as follows: Step 1.1: Collect the characteristic parameters of the strip from the hot rolling production site, and calculate the collection position of the characteristic parameters of the strip according to the following formula: L e l b +Nv△t; In the formula, l b is the starting position of data collection, i.e. the length from the head of the strip, in m; l c is the distance from the pyrometer to the coiler before coiling, in m; Δb is a fixed constant; h is the strip thickness, in m; l e is the end position of data collection, which refers to the length from the head of the strip, in m; N is the number of sampling points; v is the running speed of the strip, in m / s; Δt is the sampling period, in s; Step 1.2: Calculate the characteristic parameters of the strip according to the following formula: In the formula, x j is the value of the jth characteristic of the strip; x k,j is the jth feature of the kth sampling point; M is the total number of input feature parameters and output feature parameters; Step 1.3: Repeat steps 1.1 and 1.2 to collect cooling process parameters, information parameters and cooling water temperature of different steel strips, so as to obtain multiple samples and form an original data set.

4. The method for calculating heat flux adaptive coefficient based on data drive according to claim 1, characterized in that: The step 2 is specifically as follows: Step 2.1: Preliminarily eliminate abnormal data according to the following formula: x min <x i，j <x max ,i=1,2,...,n; In the formula, x max and x min are the upper and lower limits of the characteristic parameters respectively; x i,j is the jth feature of the i-th sample; n is the number of samples; Step 2.2: The data cleaned in step 2.1 is further cleaned using the Laida criterion to remove samples of characteristic parameters that do not conform to the following formula, so that the distribution of each characteristic parameter in the original data set is more consistent with the normal distribution: In the formula, is the mean value of the j-th feature, and σ is the standard deviation of the j-th feature; Step 2.3: Divide the cleaned dataset into training set, test set, and validation set in a ratio of 7:2:

1.

5. The method for calculating heat flux adaptive coefficient based on data drive according to claim 1, characterized in that: The step 4 is specifically as follows: Step 4.1: Determine the hyperparameters to be optimized and the optimization interval according to the type of the optimal base learner obtained in step 3; Step 4.2: Take the coefficient of determination as the second objective function to be optimized: In the formula, R 2 is the determination coefficient, the closer it is to 1, the better the model fit is; m is the number of samples in the validation set; and are the model predicted value, actual value and the average of the actual value of the i-th sample respectively; Step 4.3: To obtain the maximum value of the second objective function, the Chaos Harris Eagle Optimization Algorithm is used to optimize the hyperparameters in step 4.1 to obtain the optimal hyperparameter combination.

6. The method for calculating heat flux adaptive coefficient based on data drive according to claim 1, characterized in that: The step 6 is specifically as follows: The coefficient of determination R 2 , mean square error (MSE) and mean absolute error (MAS) are used as evaluation indicators to verify the generalization performance of the heat flux adaptive coefficient stacking model obtained in step 5.

Images