A method for predicting the mixing ratio of blast furnace upper charge based on linear regression

Abstract

The present invention provides a method for predicting the mixing ratio of the upper charge of a blast furnace based on linear regression. The method expresses the mixing ratio of the charge as the weighted sum of the characteristics of various charge operating parameters, establishes a hypothesis model, determines a loss function to quantify the difference between the actual value of the mixing ratio and the predicted value, fits the hypothesis model to update the parameters and calculates the minimum loss mean to obtain an optimal prediction model, and selects the parameters of the optimal model to obtain the predicted mixing ratio. The present invention makes a linear hypothesis based on the charge operating parameters to obtain a hypothesis model of the charge mixing ratio. The method uses the measured charge operating parameters and the corresponding mixing ratio in blast furnace simulation experimental data for training. The method seeks the optimal coefficient solution through the gradient descent method to obtain the optimal mixing ratio prediction model. The method outputs the relationship between the mixing ratio and the charge operating parameters. The method refers to the mixing ratio under different conditions to provide guidance for different blast furnace operations.

Description

Technical Field

[0001] The invention relates to a blast furnace charge, and in particular to a blast furnace charge mixing ratio prediction method. Background Art

[0002] The upper portion of a blast furnace typically utilizes an alternating charge distribution system. As the charge descends, smaller ore particles penetrate the larger coke layer, forming a so-called mixed layer. The mixing ratio describes the degree of intermixing between the ore and coke particles. Different mixing ratios can lead to varying gas flow distributions within the blast furnace. A high mixing ratio can obstruct gas flow, impacting smooth operation, productivity, thermal efficiency, and other key performance indicators. Therefore, it's essential to design a reasonable mixing ratio based on the specific charge type.

[0003] Direct numerical simulations require significant time and effort, and the numerous charge operations involved in the process all have varying degrees of impact on the mix ratio. Therefore, there is an urgent need to develop a method for predicting the mix ratio of the upper charge in a blast furnace that can quickly predict the magnitude of changes in the mix ratio resulting from different charge operations. Summary of the Invention

[0004] Purpose of the invention: In view of the shortcomings of the prior art, the present invention provides a method for predicting the mixing ratio of the upper charge of a blast furnace based on linear regression, which can be quickly predicted under complex blast furnace operations.

[0005] Technical solution: A method for predicting the mixing ratio of the upper charge of a blast furnace based on linear regression, comprising the following steps:

[0006] (1) Obtain the charge operating parameters and express the charge mixing ratio as the weighted sum of the characteristics of each charge operating parameter, that is, assume the model M = α1x1 + α2x2 + α3x3 + b, where x1, x2, x3 are charge operating parameters, α1, α2, α3 are the coefficients of the corresponding charge operating parameters, and b is the bias term;

[0007] (2) Randomly initialize the values ​​of α1, α2, α3, and b, obtain the blast furnace simulation experimental data and input it into the hypothetical model for fitting, use the gradient descent algorithm to update α1, α2, α3, and b to reduce the error, obtain the optimal model training α and b, α is [α1, α2, α3], and calculate the minimum loss mean L(α, b);

[0008] (3) Substitute α and b corresponding to the minimum loss mean L(α,b) to obtain the optimal mixing ratio prediction model

[0009] (4) Select the furnace charge operating parameters x1, x2, and x3 that need to be tested, and solve to obtain the predicted mixing ratio M.

[0010] Furthermore, in the furnace charge operating parameters of step (1), x1 is the batch weight, x2 is the ore-coke particle size ratio, and x3 is the particle emission rate.

[0011] Furthermore, step (2) the gradient descent algorithm includes the following steps:

[0012] ①Determine a loss function l (i) (α, b), to quantify the gap between the actual and predicted values ​​of the mixing ratio, i represents the i-th sample in the data set [M(i), x1, x2, x3] obtained from the blast furnace simulation experiment;

[0013]

[0014] in, is the predicted result of sample i, M(i) is the actual value of sample i;

[0015] ② Iteratively extract a small batch of samples from all data sets, and then update the parameters in the direction of the negative gradient using the following formula:

[0016]

[0017]

[0018] Among them, η is the learning rate and n is the number of selected samples.

[0019] Furthermore, step (3) calculates the minimum loss mean L(α,b):

[0020]

[0021] L(α,b)=min{L 1 (α,b),L 2 (α,b),…,L m (α,b)}

[0022] Among them, L m (α, b) represents the mean loss calculated for the mth time.

[0023] Furthermore, the furnace operating parameters x1, x2, and x3 tested in step (5) are independent of each other. When any two parameters remain unchanged, the variable parameter x can be output. i The relationship with the change of mixing ratio M.

[0024] Beneficial effects: The present invention makes a linear assumption through charge operating parameters to obtain a hypothetical model of the charge mixing ratio, uses the measured charge operating parameters and the corresponding mixing ratio in the blast furnace simulation experimental data for training, seeks the optimal coefficient solution through the gradient descent method, obtains the optimal mixing ratio prediction model, outputs the relationship between the mixing ratio and the change of charge operating parameters, and refers to the mixing ratio under different conditions, thereby providing guidance for different blast furnace operations. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 Flow chart of the method of the present invention;

[0026] Figure 2 The figure is a comparison chart of the predicted value and the actual value of the embodiment. DETAILED DESCRIPTION

[0027] A prediction method for the mixture ratio of the upper charge of a blast furnace based on linear regression, such as Figure 1 As shown, the following steps are included:

[0028] (1) Obtaining the charge operating parameters. The charge mixing ratio can be expressed as the weighted sum of the characteristics of each charge operating parameter, that is, assuming the model M = α1x1 + α2x2 + α3x3 + b, where x1, x2, and x3 are charge operating parameters, α1, α2, and α3 are coefficients corresponding to the charge operating parameters, and b is a bias term;

[0029] There are many charge operation parameters. Based on previous simulation experience, we selected batch weight x1, ore-coke particle size ratio x2, and particle emission rate x3. b is the set bias item. Although there will not be a time when any charge operation parameter is 0, this setting is conducive to improving the accuracy of the prediction model.

[0030] It should be noted that when large particles are placed on top of smaller particles, mixing is essentially nonexistent (or negligible). Therefore, the mixing ratio described herein refers to smaller ore particles distributed on larger coke particles. Batch weight refers to the weight of material placed in each batch, and particle discharge rate refers to the rate at which particles are discharged to simulate coke combustion and ore melting in a simulated blast furnace.

[0031] (2) Randomly initialize the values ​​of α1, α2, α3, and b, select the furnace operating parameters in some data sets, input them into the curve for fitting, and use the gradient descent algorithm to update the model parameters to reduce the error:

[0032] Given a data set A = [M(i), x1, x2, x3], the i-th group of data will produce a loss function l (i)(α, b) is used to quantify the difference between the actual and predicted values ​​of the mixture ratio, which is used to judge the usability of the model. A non-negative loss is used here. The smaller the value, the more accurate the prediction. The loss is 0 for a perfect prediction. The loss function determined for the i-th group of data is as follows:

[0033]

[0034] in, is the prediction result of sample i, M(i) is the actual value of sample i, and α is [α1, α2, α3].

[0035] The dataset here comes from a blast furnace simulation experiment. In a blast furnace simulation, the charge operating parameters, x1, x2, and x3, are required as inputs. When the blast furnace simulation is complete, the charge mix ratio M is calculated based on the simulation results, resulting in a data set [M, x1, x2, x3]. Repeating the blast furnace simulation multiple times, each with different input settings, yields different data. By combining these data, we obtain the dataset [M(i), x1, x2, x3].

[0036] Iteratively extract a small batch of samples from all data sets, and then update the parameters in the direction of the negative gradient. The update calculation for the i-th group of data is as follows:

[0037]

[0038]

[0039] Among them, η is the learning rate, which aims to slow down the convergence speed and is generally taken as 0.03; n is the number of selected samples.

[0040] When the model parameters α and b satisfy the minimum loss mean L(α, b) of the training data set, that is, the total loss value of all training samples is minimized, the minimum loss mean is obtained as follows:

[0041]

[0042] L(α,b)=min{L 1 (α,b),L 2 (α,b),…,L m (α,b)}

[0043] Among them, L m (α, b) represents the mean loss calculated for the mth time.

[0044] The mean loss is actually the sum of all the calculated loss values ​​of i data, and then the average. The so-called minimum means that the gradient descent function needs to be used multiple times. The loss value and the mean obtained each time are different. Here we need to find the minimum value as much as possible. The data set corresponding to the minimum loss mean is considered to be the optimal one.

[0045] (4) According to the minimum loss mean, the optimal model parameters α and b are written into the mixing ratio prediction model The trained model is helpful for early prediction of the mixing ratio, saving the time cost of simulation calculations. It can also be used to observe the changing relationship between a certain operating parameter and the mixing ratio, providing guidance for further blast furnace operations.

[0046] (5) Select the charge operating parameters x1, x2, and x3 to be tested, and the predicted mixing ratio M can be obtained. In this embodiment, the charge operating parameters include batch weight x1, ore-coke particle size ratio x2, and particle emission rate x3. These are the most important and frequently changing operating parameters in the simulated blast furnace, thus ensuring that the obtained model can be used to calculate the mixing ratio to be predicted, as well as the relationship between the mixing ratio and the change of any charge operating parameters.

[0047] Here, we can also select the charge operating parameters x1 and x3 as fixed values ​​to derive the relationship between the mixing ratio M and the ore-coke particle size ratio x2. Given a set of charge operating parameters under the aforementioned constraints, we can obtain a corresponding predicted mixing ratio value, which can be used to adjust the charge operating parameters. By observing the prediction model, we can see that the ore-coke particle size ratio x2 has a significant impact on the entire prediction model. This is because as the particle size ratio increases, small ore particles are more likely to penetrate into large coke particles. This is consistent with the actual simulation results and ensures the accuracy of the prediction model.

[0048] The entire prediction model utilizes a pre-existing mixture ratio dataset to predict the relationship between the mixture ratio and three blast furnace parameters. The predicted relationship between the mixture ratio and blast furnace parameters requires training with as many samples as possible. The accuracy of the prediction model is closely related to the diversity of the three blast furnace parameters. A well-trained model facilitates early prediction of the mixture ratio, saving time and computational time. It can also be used to observe the relationship between a specific operating parameter and the mixture ratio, providing guidance for future blast furnace operations.

[0049] According to the above method and steps, an example calculation is performed: the data set involved in the training [M(i), x1, x2, x3] is as follows:

[0050]

[0051]

[0052] Assume that the model M = α1x1 + α2x2 + α3x3 + b, initialize the values ​​of α1, α2, α3, and b to [0.1, 0.1, 0.1, 0]; read the data set for training, set the learning rate η to 0.03, select the number of repeated calculations to be m = 3, and the number of selected small-batch gradient descent algorithms to be n = 10; the α and b corresponding to the minimum loss mean L(α, b) calculated according to the above method are [α1, α2, α3] = [0.158, 0.336, -0.107], b = 0.027; that is, the model for this prediction is:

[0053] Select some data from the dataset as the test set, and bring the test set values ​​into the prediction model for verification. The test set is as follows:

[0054]

[0055] The results are as follows Figure 2 As shown in the figure, the predicted mix ratio by the prediction model is close to the actual mix ratio. Overall, the predicted value is slightly lower than the actual value. However, the maximum error of the test shown in the figure does not exceed 1%. Therefore, the prediction results of this prediction method for the mix ratio of the upper blast furnace charge are accurate and reliable.

Claims

1. A method for predicting the mixing ratio of blast furnace upper charge based on linear regression, characterized in that: The following steps are involved: (1) Obtain the charge operating parameters and express the charge mixing ratio as the weighted sum of the characteristics of each charge operating parameter, that is, assume that the model ,in is the charge operating parameter, is the coefficient corresponding to the furnace operating parameters, is the bias term; in the furnace charge operating parameters For batch weight, is the ore-coke particle size ratio, is the particle emission rate; (2) Random initialization The value of , obtains the blast furnace simulation experimental data and inputs it into the hypothesis model for fitting, and uses the gradient descent algorithm to update Reduce the error and obtain the optimal model training , for , and calculate the minimum loss mean ; (3) The minimum loss mean is obtained Corresponding Substitute into the optimal mixing ratio prediction model ; Calculate the minimum loss mean : in, Indicates the mean loss calculated for the mth time; (4) Select the furnace operating parameters to be tested , solve to get the predicted mixing ratio .

2. The method for predicting the mixing ratio of the upper charge of a blast furnace based on linear regression according to claim 1, wherein: Step (2) The gradient descent algorithm includes the following steps: ①Determine a loss function , to quantify the gap between the actual and predicted values ​​of the mixing ratio, where represents the first samples; in, ② Iteratively extract a small batch of samples from all data sets, and then update the parameters in the direction of the negative gradient using the following formula: in, ; 。 3. The method for predicting the mixing ratio of the upper charge of a blast furnace based on linear regression according to claim 1, wherein: Charge operating parameters tested in step (4) Independent of each other, when any two parameters remain unchanged, the variable parameters can be output and mixing ratio M The changing relationship.

Images