Refining furnace molten steel temperature prediction method based on dynamic selection Gaussian process regression

Abstract

The invention provides a refining furnace molten steel temperature prediction method based on dynamic selection Gaussian process regression, and relates to the technical field of iron and steel manufacturing. The method comprises two stages of integrated training and integrated fusion. In an integrated training stage, a Feature Bagging technology is utilized to create a plurality of feature subspaces from an original feature space, training data is projected into the feature subspaces to obtain training subsets, and a Gaussian process algorithm is utilized to obtain a corresponding number of base learners on the training subsets; in the integrated fusion stage, an improved KNN algorithm is used to determine a competition domain of test data, a meta learning theory is used to establish a competitiveness estimation model of the base learners, the competitiveness of all the base learners is estimated in the competition domain, and the optimal base learner is selected based on the estimated value of the competitiveness. The predicted value is used as a final molten steel temperature prediction result of the refining furnace. According to the method, engineering requirements are met in precision, statistical interpretation is provided for forecast results, and the engineering application value of the method is improved.

Description

Technical Field

[0001] The invention relates to the technical field of iron and steel manufacturing, and in particular to a method for predicting the temperature of molten steel in a refining furnace based on dynamic selection of Gaussian process regression. Background Art

[0002] The Ladle Refining Furnace (LF) is an important secondary refining equipment in the iron and steel manufacturing process. The main purpose of the LF is to ensure that the temperature and composition of the molten steel meet the specified requirements. However, in actual engineering, it is not realistic to continuously measure the temperature of the molten steel due to the limitations of current temperature measurement technology. Engineers responsible for the LF usually estimate the molten steel temperature based on experience to determine the tapping time. If the current estimate is considered to reach the required tapping temperature, a special thermocouple will be used for manual measurement until the actual temperature reaches the required value, and the tapping operation will not be performed.

[0003] Due to the many factors that affect the temperature of molten steel, the prediction of molten steel temperature is very complicated, and the estimation based on the experience of engineers is usually unstable. If the estimated value is much lower than the actual value, the current measurement is meaningless and another measurement is required. The energy loss caused by manual measurement will delay the steel tapping time. On the other hand, if the estimation exceeds the required steel tapping temperature, it will lead to a waste of energy. In short, both overestimation and underestimation of molten steel temperature will lead to economic losses, so an accurate molten steel temperature prediction model is very necessary.

[0004] The existing molten steel temperature prediction technology still has deficiencies in accuracy, which limits its practical application. In addition, most of the existing technologies are point predictions, that is, they only predict the temperature of a single point, and the credibility of the prediction results cannot be evaluated, so the reference value in practical applications is limited. Summary of the invention

[0005] The technical problem to be solved by the present invention is to provide a method for predicting the molten steel temperature of a refining furnace based on dynamically selected Gaussian process regression in response to the above-mentioned deficiencies in the prior art, which meets engineering requirements in terms of accuracy, provides statistical explanations for the forecast results, and enhances its engineering application value.

[0006] In order to solve the above technical problems, the technical solution adopted by the present invention is:

[0007] A method for predicting the temperature of molten steel in a refining furnace based on dynamic selection Gaussian process regression is proposed. A dynamic selection Gaussian process model is proposed. The model dynamically selects the best individual from the ensemble each time the temperature is predicted. The method includes two stages: ensemble training and ensemble fusion. In the ensemble training stage, Feature Bagging technology is first used to create several feature subspaces from the original feature space, and the training data is projected into the feature subspace to obtain the training subset. Then, the Gaussian process algorithm is used to obtain a corresponding number of base learners on the training subset. In the ensemble fusion stage, the improved KNN algorithm is first used to determine the competition domain of the test data. Then, the competitiveness estimation model of the base learner is established using the meta-learning theory, and the competitiveness of all base learners is estimated in the competition domain. Finally, the best base learner is selected based on the estimated value of the competitiveness, and its predicted value is used as the final prediction result of the temperature of molten steel in the refining furnace.

[0008] Furthermore, the specific steps of the integrated training stage are as follows:

[0009] Step 1.1: Use the Feature Bagging algorithm to generate n feature subspaces based on the original feature space. The specific steps are:

[0010] Step 1.1.1: Generate a random integer r, satisfying r∈[d / 2,d], where d is the dimension of the original feature space;

[0011] Step 1.1.2: Randomly extract r features from the original features to form a feature subspace;

[0012] Step 1.1.3: Repeat the above two steps n times to obtain n feature subspaces;

[0013] Step 1.2: Split the actual production data set of the refinery into a training set S train and the test set S test , of which training samples account for 70% and test samples account for 30%; in the actual production data set, the input data include initial molten steel temperature, power consumption, ladle age, empty ladle temperature, conversion time, molten steel weight, argon blowing amount, temperature measurement time and the weight of 8 additives, and the output data is the tapping temperature;

[0014] Step 1.3: Set the training set S train Projected onto n feature subspaces respectively, we get n feature subsets {S1,…,S n};

[0015] Step 1.4: Specify m different types of covariance functions for the Gaussian process algorithm to obtain m Gaussian process models {GP1,…,GP m} as base learner;

[0016] Step 1.5: For each base learner, use a feature subset to train it, and get a total of n×m trained base learners

[0017] Furthermore, the specific steps of the integration and fusion stage are as follows:

[0018] Step 2.1: From the test set S test Traverse all test samples;

[0019] Step 2.2: For the selected test sample d test First, the improved K nearest neighbor algorithm is used to determine the competitive domain R of the base learner test , the specific steps are:

[0020] Step 2.2.1: Use the RReliefF algorithm to calculate the weights of all input features; the weights of different predictions, different attributes, and different predictions and different attributes are collected in N dC 、N dA [A] and N dC&dA [A]; the final estimated value of each attribute W[A] is calculated by the following formula:

[0021] W[A]=N dC&dA [A] / N dC -(N dA [A]-N dC&dA [A]) / (mN dC );

[0022] Step 2.2.2: Use the weighted Euclidean distance formula to calculate the distance between all training data and the current test sample d test distance;

[0023] Step 2.2.3: Select the K training data with the closest distance to form the competition domain R test ;

[0024] Step 2.3: Use meta-learning theory to build an estimator by calculating the test The performance on the test data is used to estimate the competitiveness of the base learner on the current test data. The specific steps are:

[0025] Step 2.3.1: Randomly extract samples from the original training set to form a training set S for the meta-learner M-train ;

[0026] Step 2.3.2: Construct the following five base learner performance evaluation indicators:

[0027] m1: RMSE; each nearest neighbor instance To test point x i The distance between x and x is used as the weight of the corresponding neighbor, where j = 1, 2, ..., k; the following formula is used to calculate the distance between x and x. i arrive Distance:

[0028]

[0029] in,

[0030]

[0031] rank(i,j) represents the distance ranking from the sample in the competition domain to the test sample; σ is the parameter of the control distance, which represents the number of samples in the competition domain;

[0032] m1 is calculated as follows:

[0033]

[0034] in, Represent the true output value and the predicted output value respectively;

[0035] m2: The variance of the K nearest neighbor predictions, as follows:

[0036]

[0037] m3: The prediction variance of the test point itself. This capability metric is only available in ensemble learning. If there is the same prediction mean, the GPR model with a smaller prediction variance will be preferred.

[0038] m4: The correlation between the observed value and the predicted value of the instance in the capability region, which is calculated as follows:

[0039]

[0040] m5: Similarity of neighboring points, that is, the sum of the differences between the predicted value of the test point and the observed value of the instance in the capability region. This capability metric is calculated as follows:

[0041]

[0042] Step 2.3.3: Using S M-train The samples in calculate the 5 index values ​​of all base learners, thus obtaining a set of 5-dimensional data as meta-feature input data;

[0043] Step 2.3.4: Calculate the absolute value of the difference between the base learner's predicted value and the actual value as the meta-feature output data;

[0044] Step 2.3.5: Use the meta-training data from the first two steps, i.e., meta-feature input data and meta-feature output data, to train the meta-learner M in combination with the Gaussian process algorithm. meta , as an estimator of competitiveness;

[0045] Step 2.4: Using the meta-learner M meta Estimate the competitiveness of all base learners

[0046] Step 2.5: Select the most competitive base learner and determine its prediction value As a test sample d test The predicted value of .

[0047] The beneficial effect of adopting the above technical scheme is that the method for predicting the temperature of molten steel in a refining furnace based on dynamic selection of Gaussian process regression provided by the present invention first constructs a Gaussian process model of dynamic regression integration, and the dynamic integration selects the most competitive individual for each test mode according to the ability estimated by the information neighbors. Secondly, a distance metric based on RReliefF is constructed to search these neighbors instead of the traditional K nearest neighbor method. Several evaluation indicators are combined through a meta-regressor to achieve a more robust estimate of the ability. Finally, a Bayesian non-parametric model is used for integrated generation to obtain statistical predictions, and the molten steel temperature of the ladle refining furnace is successfully predicted. The method proposed in the present invention can effectively improve the prediction accuracy, meet the needs of actual projects, avoid energy waste caused by manual estimation, and ensure the safety of on-site workers at the same time. It can provide a reference for related technologies such as molten steel temperature prediction in the actual application of ladle refining furnaces, and is of great significance to the development of steelmaking technology. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 A flow chart of a method for predicting molten steel temperature in a refining furnace based on dynamic selection of Gaussian process regression provided by an embodiment of the present invention;

[0049] Figure 2 A schematic diagram of an integrated generation technology in a prediction model provided by an embodiment of the present invention;

[0050] Figure 3 A schematic diagram of an integrated fusion technology based on meta-learning provided in an embodiment of the present invention;

[0051] Figure 4 The temperature prediction effect diagram of a 300-ton refining furnace is provided in an embodiment of the present invention; wherein, Figure (4a) is a comparison diagram of the predicted temperature and the actual temperature of 100 data points, and Figure (4b) is a corresponding prediction error curve diagram;

[0052] Figure 5A schematic diagram of the application of the molten steel temperature prediction model provided by an embodiment of the present invention in a secondary system. DETAILED DESCRIPTION

[0053] The specific implementation of the present invention is further described in detail below in conjunction with the accompanying drawings and examples. The following examples are used to illustrate the present invention, but are not intended to limit the scope of the present invention.

[0054] This embodiment proposes a method for predicting the temperature of molten steel in a refining furnace based on dynamic selection Gaussian process regression. It integrates dynamic integration and Gaussian process to propose a dynamic selection Gaussian process model. This model dynamically selects the best individual from the integration each time the temperature is predicted. In order to ensure that the best individual can be correctly selected, innovative designs are proposed in terms of constructing the "competition domain" and estimating the "competitiveness". Figure 1 As shown, the method of this embodiment is described as follows.

[0055] Step 1: Base learner training (ensemble generation phase), such as Figure 2 shown.

[0056] Step 1.1: Use the Feature Bagging algorithm to generate 6 feature subspaces based on the original feature space. The specific steps are:

[0057] Step 1.1.1: Generate a random integer r, satisfying r∈[d / 2,d], where d=17 is the dimension of the original feature space;

[0058] Step 1.1.2: Randomly extract r features from the original features to form a feature subspace;

[0059] Step 1.1.3: Repeat the above two steps 6 times to obtain 6 feature subspaces.

[0060] Step 1.2: Split the actual production data of a steel plant's 300-ton refining furnace (761 groups in total) into a training set S train and the test set S test , of which 816 are training samples and 100 are testing samples. In the actual production data set, the input data include initial molten steel temperature, power consumption, ladle age, empty ladle temperature, conversion time, molten steel weight, argon blowing amount, temperature measurement time, and the weight of 8 additives, and the output data is the tapping temperature.

[0061] Step 1.3: Set the training set S train Project them onto the feature subspace respectively and obtain 6 feature subsets {S1,…,S6}.

[0062] Step 1.4: Specify five different types of covariance functions for the Gaussian Process (GP) algorithm and obtain five Gaussian process models {GP1,…,GP5} as base learners.

[0063] Step 1.5: For each base learner, use a feature subset for training, and get a total of 30 trained base learners

[0064] Step 2: Testing phase (integration and fusion phase).

[0065] Step 2.1: From the test set S test Iterate over all test samples.

[0066] Step 2.2: For the selected test sample d test First, the improved K-nearest neighbor (KNN) algorithm is used to determine the competitive domain R of the base learner test , the specific steps are:

[0067] Step 2.2.1: Use the RReliefF (Regressional ReliefF) algorithm to calculate the weights of all input features; the weights of different predictions, different attributes, and different predictions and different attributes are collected in containers N dC 、N dA [A] and N dC&dA [A]. The final estimated value of each attribute W[A] is calculated by the following formula:

[0068] W[A]=N dC&dA [A] / N dC -(N dA [A]-N dC&dA [A]) / (mN dC ).

[0069] Step 2.2.2: Use the weighted Euclidean distance formula to calculate the distance between all training data and the current test sample d test distance.

[0070] Step 2.2.3: Select the K training data with the closest distance to form the competition domain R test .

[0071] Step 2.3: Use meta-learning theory to build an estimator by calculating the test The performance of the base learner on the current test data is estimated to be competitive Process such as Figure 3 As shown, the specific steps are:

[0072] Step 2.3.1: Randomly extract samples from the original training set to form a training set S for the meta-learner M-train .

[0073] Step 2.3.2: Construct the following five base learner performance evaluation indicators:

[0074] m1: RMSE; each nearest neighbor instance To test point x i The distance between x and x is used as the weight of the corresponding neighbor, where j = 1, 2, ..., k; the following formula is used to calculate the distance between x and x. i arrive Distance:

[0075]

[0076] in,

[0077]

[0078] rank(i,j) represents the distance ranking from the sample in the competition domain to the test sample; σ is the parameter set to control the distance, which represents the number of samples in the competition domain.

[0079] m1 is calculated as follows:

[0080]

[0081] in, represent the true output value and the predicted output value respectively.

[0082] m2: Variance of the predictions of the K nearest neighbors; if the prediction variance of regressor A in the capability region is smaller than that of regressor B, then A can be said to be more stable than B. Assuming that the other metrics of the two regressors are the same, regressor A will be preferred. That is, if a base regressor is uncertain about the predictions of the K nearest neighbors of a test point, then it should also be less competent in its own predictions. Therefore, the variance of the predictions of the K nearest neighbors can also be a reasonable capability metric, calculated as follows:

[0083]

[0084] m3: The prediction variance of the test point itself; this capability measure is only available in ensemble learning because GP is used as an inductive algorithm for ensemble generation. GP can output a Gaussian distribution for each input vector. In other words, GP treats the target variable as a random variable that follows a Gaussian distribution specified by mean and variance. For two Gaussian variables with the same mean, the one with a larger variance can take values ​​far away from the mean with a higher probability. However, the best pointwise estimate of GP is the predicted mean. Therefore, if there is an identical predicted mean, the GR model with a smaller predicted variance will be preferred.

[0085] m4: The correlation between the instance observation and the predicted value in the capability region. This correlation is usually used to evaluate the regressor and is calculated as follows:

[0086]

[0087] m5: Similarity of neighboring points, that is, the sum of the differences between the predicted value of the test point and the observed value of the instance in the capability region; this capability metric is calculated as follows:

[0088]

[0089] Step 2.3.3: Using S M-train The samples in calculate the 5 index values ​​of all base learners, thus obtaining a set of 5-dimensional data as meta-feature input data.

[0090] Step 2.3.4: Calculate the absolute value of the difference between the base learner's predicted value and the actual value as the meta-feature output data.

[0091] Step 2.3.5: Use the meta-training data from the first two steps (including meta-feature input data and meta-feature output data) to train the meta-learner M in combination with the GP algorithm. meat , as an estimator of competitiveness.

[0092] Step 2.4: Using M meta Estimate the competitiveness of all base learners

[0093] Step 2.5: Select the most competitive base learner and determine its prediction value As a test sample d test The predicted value of .

[0094] In this embodiment, the simulation verification results are as follows: Figure 4 As shown, the application effect of the method proposed in the present invention in the temperature forecast of a 300-ton refining furnace is shown, wherein Figure (4a) is a comparison diagram of the predicted temperature and the actual temperature of 100 data points, and Figure (4b) is a corresponding prediction error curve diagram. By comparing the method of this embodiment with the prior art, it can be found that the prediction accuracy has been improved to a certain extent, meeting the needs of actual engineering. On this basis, Figure 5 The application scenario of the proposed temperature prediction model in the secondary control system was demonstrated. With the support of the data prediction model, the primary control system can automatically complete the steel-tapping operation, avoiding the energy waste caused by manual estimation while ensuring the safety of on-site workers.

[0095] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope defined by the claims of the present invention.

Claims

1. A method for predicting the temperature of molten steel in a refining furnace based on dynamic selection Gaussian process regression, characterized in that: The method proposes a dynamic selection Gaussian process model, which dynamically selects the best individual from the ensemble each time the temperature is predicted. The method includes two stages: ensemble training and ensemble fusion. In the ensemble training stage, Feature Bagging technology is first used to create several feature subspaces from the original feature space, and the training data is projected into the feature subspace to obtain the training subset, and then the Gaussian process algorithm is used to obtain a corresponding number of base learners on the training subset. In the ensemble fusion stage, the improved KNN algorithm is first used to determine the competition domain of the test data, and then the meta-learning theory is used to establish the competitiveness estimation model of the base learner, and the competitiveness of all base learners is estimated in the competition domain. Finally, the best base learner is selected based on the estimated value of the competitiveness, and its prediction value is used as the final prediction result of the molten steel temperature of the refining furnace.

2. The method for predicting the temperature of molten steel in a refining furnace based on dynamic selection of Gaussian process regression according to claim 1 is characterized in that: The specific steps of the integrated training phase are as follows: Step 1.1: Use the Feature Bagging algorithm to generate n feature subspaces based on the original feature space. The specific steps are: Step 1.1.1: Generate a random integer r, satisfying r∈[d / 2,d], where d is the dimension of the original feature space; Step 1.1.2: Randomly extract r features from the original features to form a feature subspace; Step 1.1.3: Repeat the above two steps n times to obtain n feature subspaces; Step 1.2: Split the actual production data set of the refinery into a training set S train and the test set S test , of which training samples account for 70% and test samples account for 30%; in the actual production data set, the input data include initial molten steel temperature, power consumption, ladle age, empty ladle temperature, conversion time, molten steel weight, argon blowing amount, temperature measurement time and the weight of 8 additives, and the output data is the tapping temperature; Step 1.3: Set the training set S train Projected onto n feature subspaces respectively, we get n feature subsets {S1,…,S n }; Step 1.4: Specify m different types of covariance functions for the Gaussian process algorithm to obtain m Gaussian process models {GP1,…,GP m } as base learner; Step 1.5: For each base learner, use a feature subset to train it, and get a total of n×m trained base learners 3. The method for predicting the temperature of molten steel in a refining furnace based on dynamic selection of Gaussian process regression according to claim 2 is characterized in that: The specific steps of the integration and fusion stage are as follows: Step 2.1: From the test set S test Traverse all test samples; Step 2.2: For the selected test sample d test First, the improved K nearest neighbor algorithm is used to determine the competitive domain R of the base learner test , the specific steps are: Step 2.2.1: Use the RReliefF algorithm to calculate the weights of all input features; the weights of different predictions, different attributes, and different predictions and different attributes are collected in N dC 、N dA [A] and N dC&dA [A]; the final estimated value of each attribute W[A] is calculated by the following formula: W[A]=N dC&dA [A] / N dC -(N dA [A]-N dC&dA [A]) / (m-N dC ); Step 2.2.2: Use the weighted Euclidean distance formula to calculate the distance between all training data and the current test sample d test distance; Step 2.2.3: Select the K training data with the closest distance to form the competition domain R test ; Step 2.3: Use meta-learning theory to build an estimator by calculating the test The performance on the test data is used to estimate the competitiveness of the base learner on the current test data. The specific steps are: Step 2.3.1: Randomly extract samples from the original training set to form a training set S for the meta-learner M-train ; Step 2.3.2: Construct the following five base learner performance evaluation indicators: m1: RMSE; for each nearest neighbor instance To test point x i The distance between x and x is used as the weight of the corresponding neighbor, where j = 1, 2, ..., k; the following formula is used to calculate the distance between x and x. i arrive Distance: in, rank(i,j) represents the distance ranking from the sample in the competition domain to the test sample; σ is the parameter of the control distance, which represents the number of samples in the competition domain; m1 is calculated as follows: in, Represent the true output value and the predicted output value respectively; m2: The variance of the K nearest neighbor predictions, as follows: m3: The prediction variance of the test point itself. This capability metric is only available in ensemble learning. If there is the same prediction mean, the GPR model with a smaller prediction variance will be preferred. m4: The correlation between the observed value and the predicted value of the instance in the capability region, which is calculated as follows: m5: Similarity of neighboring points, that is, the sum of the differences between the predicted value of the test point and the observed value of the instance in the capability region. This capability metric is calculated as follows: Step 2.3.3: Using S M-train The samples in calculate the 5 index values ​​of all base learners, thus obtaining a set of 5-dimensional data as meta-feature input data; Step 2.3.4: Calculate the absolute value of the difference between the base learner's predicted value and the actual value as the meta-feature output data; Step 2.3.5: Use the meta-training data from the first two steps, i.e., meta-feature input data and meta-feature output data, to train the meta-learner M in combination with the Gaussian process algorithm. meta , as an estimator of competitiveness; Step 2.4: Using the meta-learner M meta Estimate the competitiveness of all base learners Step 2.5: Select the most competitive base learner and determine its prediction value As a test sample d test The predicted value of .

Images