A steel ingot rate prediction method based on WOA optimized LightGBM

Abstract

The present invention discloses a method for predicting steel ingot yield based on WOA-optimized LightGBM, including data collection, data cleaning, building a Light GBM regression model, using WOA to optimize the LightGBM model, model training, inputting characteristic data of steel with ingot yield to be predicted into the trained WOA-LightGBM model, and obtaining the predicted value of the steel ingot yield. The present invention uses the LightGBM model to process data without time series relationship, uses WOA to optimize the hyperparameters of the LightGBM model, and controls the global convergence of the model according to probability by controlling the parameters of the shrinkage and encirclement mechanism, so that the model can more reliably find the global optimal solution and reduce the model error.

Description

Technical Field

[0001] The present invention belongs to the technical field of steel ingot yield prediction, and specifically relates to a steel ingot yield prediction method based on WOA optimized LightGBM. Background Art

[0002] The steel ingot yield is one of the key factors that determine the comprehensive yield of steel. Existing research mainly predicts a specific mechanical or chemical property of steel based on real-time detection data of steel smelting with time series. For example, the PSO-LSTM model is used to predict the mechanical properties of steel continuous casting embryos based on characteristics such as drawing speed, carbon content, phosphorus content, and sulfur content on a data set with a time series relationship. Although this method can predict the mechanical properties of steel well, the LSTM model is a time recursive neural network and is not suitable for processing data sets without time series relationships. In addition, some scholars have analyzed the Markov chain composed of particles in the PSO algorithm and proved that the PSO algorithm does not have global convergence from the perspective of transition probability. Therefore, when using this model to predict the steel ingot yield, it is easy to fall into a local optimal solution and it is difficult to ensure the prediction accuracy.

[0003] With the development of industrial intelligence, it is necessary to propose a solution that can shorten the research and development cycle of smelting products, improve quality stability and production efficiency, and is suitable for predicting the steel ingot rate. Summary of the invention

[0004] The purpose of the present invention is to address the problems that the existing PSO-LSTM model cannot process data sets without time series relationships, the PSO algorithm does not have global convergence for the search of hyperparameters, and is prone to fall into local optimal solutions, which makes it difficult for the existing model to find the global optimal solution and the prediction error is high. A steel ingot rate prediction scheme is proposed, which is more suitable for processing data sets without time series relationships, is easier to jump out of local optimal solutions, and has more accurate prediction results.

[0005] To achieve the above object, the present invention adopts the following technical solution:

[0006] A method for predicting steel ingot yield based on WOA-optimized LightGBM, comprising the following steps:

[0007] Step 1: Data collection: Collect several sample steels with ingot yield values, use the characteristic data of these samples as a sample set, and extract the characteristic data of the steel with ingot yield to be predicted;

[0008] Step 2: Data cleaning: standardizing, performing KS test, outlier detection, missing value filling and normalization on the sample data in step 1;

[0009] Step 3: Build the WOA-LightGBM regression model. Build a decision tree based on the GBDT algorithm, and build a Light GBM model using the histogram algorithm and the leaf growth strategy with depth restrictions. Use the WOA whale algorithm to optimize LightGBM, and limit the parameters A and C in the WOA algorithm so that the model can obtain the optimal hyperparameters during training. Input the sample data cleaned in step 2 into the WOA-LightGBM model for training. Use the K-fold cross-validation method during training to obtain a trained WOA-LightGBM model.

[0010] Step 4: Input the characteristic data of the steel whose ingot yield is to be predicted into the WOA-LightGBM model trained in step 3 to obtain the predicted value of the ingot yield of the steel.

[0011] As a preferred embodiment, the sample set described in step 1 and the characteristic data of the steel to be predicted ingot yield include: ingot specifications, scrap steel input, returned alloy steel, carbon steel, scars, short length, white spots, other surface features, Mo, Ni, Cr, Mn, C and Si, a total of 14 characteristics.

[0012] Specifically, step 2 includes:

[0013] Step 2.1: Delete the samples with a success rate of 1 and standardize the data to obtain standardized data set samples;

[0014] Step 2.2: Perform KS test on the standardized data set samples;

[0015] Step 2.3 performs outlier detection on the data set that has passed the KS test and clears the outlier values;

[0016] Step 2.4 fills the missing values ​​for the cleared values;

[0017] Step 2.5 normalizes the padded data set to obtain the cleaned data set.

[0018] Furthermore, in step 2.3, if the KS test results meet the characteristics of normal distribution, outliers are detected according to the 3σ principle and the outliers are cleared; if the characteristics of non-normal distribution are met, outliers are detected using the quartile method and the outliers are cleared.

[0019] Furthermore, the method of step 2.4 is as follows: the missing values ​​of the feature data of the data set in step 2.3 are processed by using the mean interpolation method to fill in the values ​​cleared in step 2.3.

[0020] Specifically, the method of limiting the value range of parameters A and C in the WOA algorithm in step 3 is: first calculate the distance vector between the current optimal position and a solution in the current whale population, and use different formulas to update the values ​​of parameters A and C according to the size of the distance vector.

[0021] Furthermore, in step 3, the K-fold cross-validation method is used in the model training process to divide the data set into K subsets, each subset data is used as a validation set, and the remaining K-1 groups of subset data are used as training sets. K rounds of training are performed to obtain the trained WOA-LightGBM model.

[0022] LightGBM is a model optimized on the GBDT algorithm with excellent results. It has been applied to housing price evaluation, user churn, disease prediction and other issues. It is difficult to detect the various mechanical properties of steel in real time during the steel smelting process, so the relevant data usually does not include time series relationships. The LightGBM model does not require the data set to have a time series relationship, and has a better effect when processing high-dimensional data. It is suitable for the prediction of steel ingot rate. WOA is a heuristic algorithm. Existing studies have proved through Markov chains that the contraction and encirclement mechanism in the WOA algorithm is a key factor in determining whether the whale optimization algorithm converges. Therefore, the parameters of the contraction and encirclement mechanism can be controlled to make WOA converge globally and find the global optimal solution. The WOA algorithm has the ability to make up for the PSO algorithm. The present invention uses WOA to optimize the hyperparameters of LightGBM to further improve the accuracy of the model. In addition, in view of the problem that the LightGBM model is sensitive to data noise, the present invention uses a variety of methods to handle data exceptions and reduce the interference of noise on the LightGBM model.

[0023] Using WOA to optimize LightGBM for steel ingot rate prediction can make predictions in the absence of time series; the model uses the WOA algorithm to optimize hyperparameters, which can more reliably find the global optimal solution compared to the original PSO algorithm. The present invention further improves the prediction effect by optimizing hyperparameters through WOA.

[0024] The beneficial effects of the present invention are:

[0025] The present invention provides a method for predicting steel ingot yield that can handle data without time series relationships, can more reliably find the global optimal solution, and has better prediction effects. The present invention uses the LightGBM model to process data without time series relationships, uses WOA to optimize the hyperparameters of the LightGBM model, and controls the model to converge globally based on probability by controlling the parameters of the shrinkage and encirclement mechanism, so that the model can more reliably find the global optimal solution and reduce model errors. In addition, when processing data, the present invention combines prior knowledge with multiple methods to eliminate abnormal data, reduces the impact of noise on the model, and further reduces model errors. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 It is the process of the method of the present invention;

[0027] Figure 2 is a box plot of characteristic data of non-normal distribution of steel ingot yield under the quartile method in the embodiment;

[0028] Figure 3 It is the shrinking and surrounding mechanism of the WOA algorithm;

[0029] Figure 4 It is the spiral update position mechanism of the WOA algorithm;

[0030] Figure 5 This is the prediction effect diagram of WOA-LightGBM;

[0031] Figure 6 It is the comparison of the root mean square error of multiple models in predicting the steel ingot yield;

[0032] Figure 7 It is the comparison of mean absolute errors of multiple models in predicting steel ingot yield;

[0033] Figure 8 It is the comparison of the average absolute percentage error of multiple models in predicting the steel ingot yield; DETAILED DESCRIPTION

[0034] like Figure 1 , the prediction method of the present invention comprises:

[0035] Step 1. Data collection. Samples with steel ingot yield values ​​are selected from the steel ingot yield data set. The sample features are steel ingot specifications (unit: tons, for example, 24 steel nails are put into a furnace of steel), scrap steel input (unit: %), returned alloy steel (unit: %), which is a kind of steel waste that does not become a product in the production process of a steel plant. It can generally be directly loaded into the furnace for smelting steel), carbon steel (unit: %), an iron-carbon alloy, scarring (one of the surface defects of steel, manifested as scar-like metal flakes on the surface of the product, the unit is hidden), short length (an inspection item for the surface quality of steel ingots, which refers to a length that is less than the lower limit of the indefinite length specified in the standard, but not less than the shortest allowed length, units omitted), white spots (a check item for the surface quality of steel ingots, a microcrack inside steel or forgings, units omitted), other surfaces (a check item for the surface quality of steel ingots, the name of a surface defect of steel, units omitted), Mo (the content of chemical element Mo, units omitted), Ni (the content of chemical element Ni, units omitted), Cr (the content of chemical element Cr, units omitted), Mn (the content of chemical element Mn, units omitted), C (the content of chemical element C, units omitted), Si (the content of chemical element Si, units omitted) There are 14 features in total, a total of 200 samples. Generate sample set (F, y) = [f 1 ,f 2 ,f 3 ,……,f 14 ,y]. F=[f 1 ,f 2 ,f 3 ,……,f 14 ] is the feature vector set, and y is the target vector. Some sample data are selected and shown in Table 1:

[0036] Table 1

[0037]

[0038] Step 2. Data cleaning

[0039] In the process of steel smelting, the ingot rate value usually cannot reach 1. Therefore, the sample with an ingot rate of 1 is abnormal data and is deleted by a simple elimination method, such as sample 15 in Table 1. The data is standardized to obtain a new sample set, which is the standardized data set sample. Taking Table 1 as an example, the standardized data set sample is shown in Table 2:

[0040] Table 2

[0041]

[0042] The KS test is performed on the standardized data set sample of steel ingot rate obtained in Table 2 above. The results of the KS test are shown in Table 3:

[0043] Table 3

[0044] feature pvalue feature pvalue Ingot Specifications 0.000000739 Surface Other 0.061171093 Scrap input 0.000023648 Mo 0.015705672 Return to Alloy Steel 0.071811607 Ni 0.008786475 Carbon Steel 0.000208725 Cr 0.000050450 Scarring 0.005652619 Mn 0.001324739 short ruler 0.000014218 C 0.000001375 White Spot 0.000030607 Si 0.000052072

[0045] Features with pvalue greater than 0.05 satisfy the normal distribution. Features that satisfy the normal distribution include alloy steel and other surface features. Features that do not satisfy the normal distribution include ingot specifications, scrap steel input, carbon steel, scars, short lengths, white spots, Mo, Ni, Cr, Mn, C, and Si, a total of 12. For features with normal distribution results, outliers are detected according to the 3σ principle: the standard deviation of the sample set on feature j is σ, and the mean is μ. Since feature j conforms to the normal distribution, the probability that the value of feature j is distributed in (μ-3σ, μ+3σ) is 0.9974. Clear the values ​​that exceed (μ-3σ, μ+3σ). After inspection, there is no abnormal data in the corresponding data of alloy steel content and other surface features that conform to the normal distribution, and there is no need to clear them.

[0046] The quartile method is used to detect outliers for non-normal distribution characteristics. The box plot of the characteristic data of non-normal distribution of steel ingot rate under the quartile method is as follows: Figure 2 As shown. Clear the abnormal values: let the upper quartile of the sample set on feature j be Q1, and the lower quartile be Q2, then there are the following maximum and minimum value boundaries:

[0047]

[0048] Among them, k can be 1.5 or 3. In this embodiment, it is 1.5. Based on this solution, the upper and lower boundaries of feature j are obtained, and the values ​​beyond the boundaries are cleared. The results after clearing are shown in Table 4:

[0049] Table 4

[0050]

[0051] Among them, NAN is the cleared value, indicating that the white spot characteristic value of sample 14, the Si characteristic value of sample 16, and the carbon steel and white spot characteristic values ​​of sample 17 are abnormal and need to be cleared.

[0052] Then, all cleared values ​​are interpolated by the mean value, and the missing values ​​are interpolated by the average value of the valid value of the attribute. Taking the samples listed in Table 4 as an example, the results after the mean value interpolation are shown in Table 5:

[0053] Table 5

[0054]

[0055] The feature data in Table 5 are normalized. The sample data of the 14 features obtained are scaled to

[0056] [0,1] range. Use formula (2):

[0057]

[0058] Among them, f i (j) is the eigenvalue of the jth sample in the i-th eigenvector, f i (j)′ is the eigenvalue of the jth sample of the i-th eigenvector after normalization, min(f i ) is the minimum value of the elements in the i-th eigenvector, max(f i ) is the maximum value of the elements in the i-th eigenvector. Taking the samples in Table 5 as an example, the normalized data is shown in Table 6:

[0059] Table 6

[0060]

[0061] Step 3. Build the WOA-LightGBM regression model

[0062] Step 3.1 Build a decision tree based on the GBDT algorithm.

[0063] Assume that the data set D = {(x i ,y i )|i=1,2,...,n}, where n represents the number of samples, x i represents the i-th sample, each sample has i features, y i Represents the sample target value.

[0064] First, initialize the weak learner: Estimate the constant value that minimizes the loss function according to formula (3):

[0065]

[0066] Where L is the loss function; C is the output value of the subinterval, generally the mean, which minimizes the estimated loss function; y i is the target value of sample i, i = 1, 2, …, n.

[0067] Secondly, model iteration: let m represent the number of iterations, and the maximum value is M. According to formula (4), the residual r is calculated mi , update the target value according to the residual, L is the loss function, f m-1 (x) is the output value of the previous iteration:

[0068]

[0069] According to the residual r mi Fit the regression tree and get the leaf node domain R of the mth tree mj, j = 1, 2, 3, ..., J. Where j represents the number of leaf nodes. The output value C of each region is calculated according to formula (5) to minimize the loss function:

[0070]

[0071] Then update the regression tree according to formula (6):

[0072]

[0073] Where I is the indicator function, when the regression tree determines x∈m j , its value is 1, otherwise it is 0.

[0074] Finally, the final model is output after M iterations.

[0075] Step 3.2 On this basis, the Light GBM model is constructed using the histogram algorithm and the leaf growth strategy with depth limit.

[0076] (1) Histogram algorithm

[0077] Discretize continuous floating-point features to obtain L integers, and construct a histogram with a width of L. When the model traverses the data, it uses the discretized values ​​as indexes to accumulate statistics in the histogram. After completing a traversal, find the optimal split point among the discretized L integers.

[0078] (2) Leaf growth strategy with depth restriction

[0079] Each time a split occurs, the leaf node with the largest split gain is selected from all current leaf nodes for splitting, and a maximum depth max_depth is set. This ensures the efficiency of the model and prevents overfitting.

[0080] Step 3.3. On this basis, use WOA to optimize the LightGBM model

[0081] The WOA whale algorithm is used to optimize LightGBM, obtain the optimal hyperparameters, and improve the accuracy of the model. The main steps include:

[0082] Step 3.3.1 Initialize the position and parameters of the whale population; calculate the parameter values ​​of the initial position of the whale population and initialize the current optimal solution;

[0083] Initialize the whale position and parameters: Set the number of whale populations SearchAgents_no to 30; use the maximum depth max_depth, the number of leaves num_leaves, and the minimum number of data that can be on a leaf mindatain_leaf in the LightGBM model as the objects of WOA optimization, and set the individual dimension dim to 3; randomly initialize or manually set the initial position of the whale population

[0084] SearchAgentsPosition[SearchAgents_no, dim]; Each individual in the whale population represents an alternative solution. The initial fitness of each individual in the current whale population is calculated according to formula (7):

[0085]

[0086] Fitness represents the fitness of the individual, y represents the predicted value, represents the mean of the true values, Represents the true value. 2 Indicates the goodness of fit of the model

[0087] Select the optimal solution among all current individuals as the initial optimal solution, and initialize the optimal fitness with the corresponding fitness.

[0088] Step 3.3.2 Hyperparameter optimization: Iteratively update the spatial position of the whale population, calculate the fitness based on the updated whale group position in each round, and determine the position of the next generation of whales based on the change in fitness; determine whether the end condition is met, if so, output the position corresponding to the optimal whale individual and the corresponding parameter value, if not, continue to iterate until the preset conditions are met. The preset conditions set in this model are that the number of iterations reaches the maximum iteration generation k or the fitness reaches 0.8, where k is 500.

[0089] At each iteration, a number p is randomly selected between [0,1].

[0090] If p is less than 0.5, WOA searches for prey according to the shrinking and surrounding mechanism, and updates the position of individual whales in different ways according to the size of ||A||. The calculation formula of A is shown in formula (8): If ||A|| is less than 1, WOA updates the position of individual whales according to formula (10); if ||A|| is greater than or equal to 1, an individual in the current whale population is randomly selected and the whale position is updated according to formula (11):

[0091] represents the coefficient vector, The step length calculation formula (8) describes the whale's swimming:

[0092]

[0093] in, and is a random vector in [0,1]; During the search process, it decreases linearly from 2 to 0, as shown in formula (9):

[0094]

[0095] Among them, t represents the current number of iterations, T max is the maximum number of iterations.

[0096]

[0097] Where t represents the number of current iterations; represents a solution randomly selected from all individuals in the whale population; It describes the distance vector between the whale individual and the randomly selected whale individual at the tth iteration. represents the position vector of the individual whale at the tth generation of iteration, Represents the position vector of the individual whale at the iteration t+1 generation.

[0098] Update the spatial position of the whale population according to formula (11), the effect is as follows Figure 3 As shown:

[0099]

[0100] Among them, t represents the number of current iterations, It describes the distance vector between the current optimal solution and a solution in the t-th generation population. represents the position vector of the individual whale at the tth generation of iteration, Represents the position vector of the individual whale at the iteration t+1 generation.

[0101] represents the best whale position vector d so far, Represents the current whale position vector.

[0102] Since the global convergence of the WOA algorithm is affected by the parameters of the shrinkage and encirclement mechanism, the parameters A and C are controlled:

[0103] CX * (t)-X i When (t)>0, the coefficients A and C are required to satisfy formula (12). If not, a new random vector is taken. and Among them, X * (t) represents the current optimal solution, Xi (t) represents the i-th solution:

[0104] and

[0105] CX * (t)-X i When (t) < 0, the coefficients A and C are required to satisfy formula (13). If not, a new random vector is taken. and

[0106]

[0107] The probability of other situations occurring is small, so they are no longer considered.

[0108] If p is greater than or equal to 0.5, WOA performs hunting behavior:

[0109] Update the position of individual whales according to formula (14), such as Figure 4 As shown:

[0110]

[0111] in, represents the distance from the position of the i-th searching whale individual to the target optimal solution, b is a constant used to define the shape of the logarithmic spiral, and l is a random vector uniformly distributed on [-1,1].

[0112] In particular, if the updated feature exceeds the reasonable range of the feature, the update of the feature is canceled.

[0113] According to formula (7), the fitness of all individuals in the population is calculated.

[0114] The solution with the largest current fitness is selected as the current optimal solution, and its fitness is taken as the current optimal fitness.

[0115] Determine whether the end condition is met. If so, output the position corresponding to the optimal whale individual and the corresponding parameter value. If not, continue to iterate until the preset condition is met.

[0116] In the present invention, the constraint condition is that the fitness is greater than 0.8 or the number of iterations reaches a maximum k, where k is 500.

[0117] Step 3.4 Model training

[0118] Step 3.4.1: Divide the cleaned data into 70% as training set and 30% as test set. Some training sets and test sets are shown below. The first 5 samples of the training set are shown in Table 7, and the first 5 samples of the test set are shown in Table 8.

[0119] Table 7

[0120]

[0121] Table 8

[0122]

[0123] Step 3.4.2 uses K-fold cross-validation to divide the original data into K groups, use each subset of data as a validation set, and use the remaining K-1 groups of subset data as training sets. K rounds of training are performed in this way. In this embodiment, K=10.

[0124] Step 3.4.3 Input the data containing 14 features in the test set to obtain the predicted value of steel ingot rate. The results of the above samples 83, 64, 199, 77, and 70 predicted by the model are shown in Table 9:

[0125] Table 9

[0126] Sample No. True value Predicted value 83 0.935 0.936 64 0.936 0.941 199 0.928 0.937 77 0.946 0.944 70 0.899 0.922

[0127] Step 4. Model Evaluation

[0128] After completing the model training, it is necessary to evaluate the prediction effect of the model. This embodiment mainly uses the root mean square error (RMSE), mean absolute error (MAE) and mean absolute percentage error (MAPE). The formula is as follows:

[0129]

[0130] Among them, y is the actual value of steel ingot rate, is the predicted value, is the mean of the actual value, and m is the number of samples.

[0131] Figure 5 This is the prediction effect diagram of WOA-LightGBM. In order to better evaluate the model effect, the prediction results of the model are compared with the existing methods. The comparison results are shown in Table 10.

[0132] Table 10

[0133] Prediction Model RMSE MAE MSE LightGBM 0.0491 0.0372 0.0026 Gradient Boosting Regressor 0.0455 0.0333 0.0022 Bayesian Ridge 0.0459 0.0362 0.0024 K Neighbors Regressor 0.0437 0.0332 0.0021 AdaBoost Regressor 0.0444 0.0328 0.0022 PSO-LightGBM 0.0448 0.0340 0.0020 WOA-LightGBM 0.0442 0.0323 0.0020

[0134] It can be seen that according to the root mean square error, mean absolute error and mean absolute percentage error, the WOA-LightGBM model of the present invention is superior to the original LightGBM, Bayesian Ridge regression, AdaBoost and other models. The steel ingot rate calculated by the WOA-LightGBM regression fusion method has better accuracy, such as Figure 6-8As shown. The present invention uses the LightGBM model as the basic model, which can process data without time series relationship; the WOA algorithm is used to control the parameters of the shrinkage and encirclement mechanism to ensure the probability convergence of the WOA algorithm, which can more reliably find the global optimal solution than the original PSO algorithm. In addition, the present invention eliminates abnormal data through a variety of abnormality detection methods, thereby reducing the interference of noise on the LightGBM model. The present invention can improve the prediction accuracy of steel ingot rate.

[0135] The WOA-LightGBM model of the present invention can process ordinary data without time-series relationship, and WOA converges globally according to probability. The convergence of the algorithm can be controlled by controlling the parameters of the shrinkage and encirclement mechanism of the WOA algorithm. Therefore, WOA-LightGBM can find the optimal solution better than PSO-LSTM.

Claims

1. A steel ingot rate prediction method based on WOA optimized LightGBM, characterized in that: The steps include: Step 1: Data collection; a number of sample steels with ingot yield values ​​are collected, and the characteristic data of these samples are used as a sample set to extract the characteristic data of the steel with ingot yield to be predicted; the sample set and the characteristic data of the steel with ingot yield to be predicted in step 1 both include: ingot specifications, scrap steel input, returned alloy steel, carbon steel, scars, short length, white spots, other surface features, Mo, Ni, Cr, Mn, C and Si, a total of 14 characteristics; Step 2: Data cleaning: standardizing, performing KS test, outlier detection, missing value filling and normalization on the sample data in step 1; Step 3: Build the WOA-LightGBM regression model. Build a decision tree based on the GBDT algorithm, and use the histogram algorithm and the leaf growth strategy with depth restriction to build the LightGBM model. Use the WOA whale algorithm to optimize LightGBM, and limit the value range of parameters A and C in the WOA algorithm so that the model can automatically obtain the optimal hyperparameters during training. Input the sample data cleaned in step 2 into the WOA-LightGBM model for training. Use the K-fold cross-validation method during training to obtain the trained WOA-LightGBM model. Step 4: Input the characteristic data of the steel whose ingot yield is to be predicted into the WOA-LightGBM model trained in step 3 to obtain the predicted value of the ingot yield of the steel.

2. According to claim 1, a method for predicting steel ingot rate based on WOA optimized LightGBM is characterized in that: Step 2 specifically includes: Step 2.1: Delete the samples with a success rate of 1 and standardize the data to obtain standardized data set samples; Step 2.2: Perform KS test on the standardized data set samples; Step 2.3 performs outlier detection on the data set that has passed the KS test and clears the outlier values; Step 2.4 fills the missing values ​​for the cleared values; Step 2.5 normalizes the padded data set to obtain the cleaned data set.

3. According to claim 2, a method for predicting steel ingot yield based on WOA optimized LightGBM is characterized in that: In step 2.3, if the KS test results meet the characteristics of normal distribution, outliers are detected according to the 3σ principle and the outliers are cleared; if the characteristics of non-normal distribution are met, outliers are detected using the quartile method and the outliers are cleared.

4. According to claim 3, a method for predicting steel ingot rate based on WOA optimized LightGBM is characterized in that: The approach of step 2.4 is as follows: use the mean interpolation method to process the missing values ​​of the feature data of the data set in step 2.3 and fill in the values ​​cleared in step 2.

3.

5. The method for predicting steel ingot rate based on WOA optimized LightGBM according to claim 1, characterized in that: The method for limiting the value range of parameters A and C in the WOA algorithm in step 3 is: first calculate the distance vector between the current optimal position and a solution in the current whale population, and according to the size relationship between the distance vector and the zero vector, require the values ​​of parameters A and C to meet different restrictions.

Images