A data-driven-based copper smelting copper matte grade prediction method

Abstract

The present application relates to a kind of based on data-driven copper smelting copper grade prediction method, belong to electric digital data processing technical field.Based on data-driven copper smelting copper grade prediction method, its steps include: relevant feature data selection;Data pre-processing;Data expansion;Random forest is constructed the copper grade prediction model;XGBoost is constructed the copper grade prediction model;With the copper grade prediction model of random forest and the copper grade prediction model of XGBoost as base learner, with linear regression model as meta-learner, construct the copper grade prediction model of Stacking ensemble learning, the copper grade is predicted.The present application constructs the copper grade prediction model based on Stacking ensemble learning, realizes the high-quality prediction of copper grade, and the prediction accuracy is higher than single prediction model, also solves the problem of single copper grade prediction model in industrial production.

Description

Technical Field

[0001] This invention relates to a data-driven method for predicting the grade of matte in copper smelting, belonging to the field of electronic digital data processing technology. Background Technology

[0002] Pyrometallurgical copper smelting refers to the process of removing impurities such as sulfur, iron, lead, zinc, nickel, arsenic, antimony, tin, magnesium, and oxygen by oxidizing impurities through oxidation slag formation or volatilization under high-temperature molten conditions, utilizing the properties that some impurities have a greater affinity for oxygen than copper, while their oxides are insoluble in molten copper. The matte smelting step is the most crucial in pyrometallurgical copper production. Its purpose is to oxidize some of the iron in the copper concentrate or roasted ore and remove it through slag formation with gangue and flux, producing matte with a high copper content (28%–78%). The matte grade at this stage directly affects the amount of cold feed added in subsequent processes, thus impacting the subsequent blowing, refining, and even the overall production cost. Therefore, controlling the copper content within the target range is of paramount importance.

[0003] Given a fixed total amount and composition of materials fed into the furnace, changes in the amount of flux, blast volume, and energy inputs such as coal will inevitably lead to changes in the grade of matte, and consequently, production costs. In actual industrial production, the amounts of flux, blast volume, and energy inputs such as coal fluctuate constantly, and these factors also change with actual production conditions, resulting in corresponding changes in the grade and cost of matte during the smelting process.

[0004] In actual enterprise production, finding the lowest-cost operating conditions for the target matte grade, given a fixed total amount and composition of materials fed into the furnace, will help enterprises achieve energy conservation, emission reduction, and cost savings, and is of great significance for guiding production. Driving the copper smelting industry towards intelligent development with artificial intelligence is of great significance to the intelligent manufacturing of my country's copper smelting industry.

[0005] With constantly changing operating conditions, determining whether the grade of the produced copper matte is within the target range through offline testing not only requires a large amount of manpower and resources, increasing production costs, but also results in a long delay in feedback, which cannot effectively guide production.

[0006] With the iterative updates and advancements in computer science and technology, artificial intelligence, as a product of our time, is increasingly becoming the engine of the new industrial revolution. Machine learning, as one of the most important branches of artificial intelligence, has been widely applied in all aspects of production and daily life, and it also plays a powerful role in the analysis and prediction of matte grade data. Machine learning prediction can determine the optimal operating conditions for matte smelting in real time, not only eliminating cumbersome testing and analysis processes and saving production costs, but also providing real-time and scientific guidance for matte smelting operations, improving production efficiency, and marking a significant step towards the intelligent transformation of metallurgical enterprises.

[0007] To improve prediction accuracy, researchers have successively established machine learning models to predict matte grade. Traditional methods mainly rely on neural network models or a combination of neural network models and fuzzy algorithms for matte grade prediction. Wang Jilin et al.'s published article, "Research on the Model and Key Parameter Prediction of Copper Flash Smelting Process Based on Mass Balance," addresses the difficulty of directly detecting parameters such as matte grade, matte temperature, and iron-silicon ratio online. They established a prediction model for these three parameters based on a BP neural network, utilizing a large amount of historical smelting data to mine implicit information, thus achieving dynamic prediction. However, due to the tendency of BP neural networks to get trapped in local minima, their prediction accuracy is low and the model's prediction stability is weak. Therefore, finding a new prediction method for matte grade prediction to improve the accuracy and stability of the model has significant theoretical and practical value for guiding copper pyrometallurgical processes. Summary of the Invention

[0008] To address the problems and shortcomings of the existing technologies, this invention provides a data-driven method for predicting the grade of matte in copper smelting. This invention constructs a matte grade prediction model based on Stacking ensemble learning, achieving high-quality predictions of matte grade. Its prediction accuracy is higher than that of single prediction models, and it also solves the problem of relying on a single model for matte grade prediction in industrial production.

[0009] To achieve the above-mentioned technical objectives, the present invention is implemented through the following technical solution:

[0010] A data-driven method for predicting the grade of matte in copper smelting, comprising the following steps:

[0011] Step 1: Selection of relevant feature data;

[0012] Step 2: Data preprocessing;

[0013] Step 3: Data augmentation;

[0014] Step 4: Construct a random forest model for predicting copper matte grade;

[0015] Step 5: Construct an XGBoost model for predicting the grade of copper matte;

[0016] Step 6: Using the random forest copper grade prediction model from Step 4 and the XGBoost copper grade prediction model from Step 5 as base learners, and the linear regression model as the meta-learner, construct a stacking ensemble learning copper grade prediction model to predict the copper grade.

[0017] In step 1, the relevant characteristic data selected are factors that are strongly correlated with the grade of matte: oxygen concentration, amount of copper concentrate added, amount of slag-forming agent silica added, air volume, amount of diesel added, amount of coal added, Cu content in copper concentrate, Fe content in copper concentrate, S content in copper concentrate, Si content in copper concentrate, As content in copper concentrate, Ca content in copper concentrate, Zn content in copper concentrate, H2O content in copper concentrate, Si content in flux silica, Fe content in flux silica, H2O content in flux silica, and oxygen-related amounts.

[0018] The data preprocessing in step 2 includes missing data processing, outlier processing, and data normalization.

[0019] The data augmentation in step 3 includes the following steps:

[0020] Step 3.1: Divide the preprocessed data from Step 2 into a training set and a test set;

[0021] Step 3.2: The training set obtained in Step 3.1 is expanded using the SMOTE resampling algorithm to obtain an expanded training set.

[0022] Step 4, which involves constructing a random forest model for predicting the grade of copper enamel, includes the following steps:

[0023] Step 4.1: Construct a random forest copper grade prediction model using the expanded training set obtained in Step 3.2;

[0024] Step 4.2: Use the test set obtained in Step 3.1 to predict the random forest copper grade prediction model in Step 4.1 to obtain the prediction set.

[0025] Step 4.1, which involves expanding the training set to construct a random forest copper grade prediction model, specifically includes:

[0026] Step 4.1.1: Using the bagging algorithm Where K is the number of trees, f k (x) represents the k-th decision tree. A random forest copper grade prediction model I is constructed.

[0027] Step 4.1.2: Determine the parameters and range of the expanded training set in Step 3.2, and train the random forest copper grade prediction model I in Step 4.1.1 using the particle swarm optimization algorithm to obtain the optimal parameters of the output model;

[0028] Step 4.1.3: Construct a random forest copper ore grade prediction model by adjusting the optimal parameters of the model.

[0029] Step 5 involves constructing the XGBoost copper matte grade prediction model, which includes the following steps:

[0030] Step 5.1: Construct an XGBoost copper matte grade prediction model using the expanded training set obtained in Step 3.2;

[0031] Step 5.2: Use the test set obtained in Step 3.1 to predict the XGBoost copper matte grade prediction model in Step 5.1 to obtain the prediction set.

[0032] Step 5.1, which involves expanding the training set to construct an XGBoost model for predicting the grade of copper matte, specifically involves:

[0033] Step 5.1.1: Using the boosting algorithm:

[0034] in Let f be the prediction result for the i-th sample, K be the total number of trees, and f be the prediction result for the i-th sample. k For the k-th decision tree, construct XGBoost copper matte grade prediction model I;

[0035] Step 5.1.2: Determine the parameters and range of the expanded training set in Step 3.2, and train the XGBoost copper matte grade prediction model I in Step 5.1.1 using the particle swarm optimization algorithm to obtain the optimal parameters of the output model;

[0036] Step 5.1.3: By outputting the optimal parameters of the model, according to the algorithm: Where y i Let i be the true value of the i-th sample. For sample x i The training error, Ω(f) k Let ) be the regularization term of the k-th decision tree, defined as:

[0037]

[0038] Where T is the number of leaf nodes, w is the score of the leaf nodes, and γ and λ are two trade-off factors, thus constructing the XGBoost model for predicting the grade of copper matte.

[0039] The specific steps of Stacking ensemble learning in step 6 are as follows:

[0040] Step 6.1: The prediction set obtained by the random forest copper grade prediction model in Step 4.2 and the prediction set obtained by the XGBoost copper grade prediction model in Step 5.2 are used as the new features x1 and x2 of the training set of the second layer model; the average of the prediction sets obtained by the random forest copper grade prediction model in Step 4.2 and the average of the prediction sets obtained by the XGBoost copper grade prediction model in Step 5.2 are used as the new features y1 and y2 of the test set of the second layer model.

[0041] Step 6.2: Use a multiple linear regression learner to train the new features x1 and x2 in the training set from Step 6.1 to obtain the multiple linear regression model;

[0042] Step 6.3: Use the new features y1 and y2 of the test set obtained in Step 6.1 to predict the multiple linear regression model in Step 6.2, and finally obtain the prediction results of the Stacking ensemble learning model for predicting the grade of copper enamel.

[0043] In step 3.1, the preprocessed data is divided into training and test sets using a 5-fold cross-validation method.

[0044] A system based on a data-driven method for predicting the grade of copper matte in copper smelting has been developed, which can be directly used for predicting the grade of copper matte in copper smelting. The overall architecture of the system mainly consists of four layers: a data acquisition layer, a data processing layer, a prediction layer, and an application layer. After inputting the values ​​of each independent variable in the interface, the predicted results of the copper matte grade can be displayed in real time.

[0045] The beneficial effects of this invention are:

[0046] This invention constructs a copper matte grade prediction model based on Stacking ensemble learning, effectively improving the prediction accuracy of copper matte grade. It provides a basis for determining whether the produced copper matte meets the target grade after changing operating conditions, thus finding the lowest-cost operating conditions. This invention can provide important reference for optimizing industrial production processes, contributing to the realization of smart factory construction, energy conservation and emission reduction, and cost savings, and is of great significance for guiding production. Attached Figure Description

[0047] Figure 1 This is a flowchart of the copper matte grade prediction method based on Stacking ensemble learning in this invention.

[0048] Figure 2 This is a flowchart of the random forest model for predicting the grade of copper matte, which is constructed using the random forest machine learning algorithm of this invention.

[0049] Figure 3This is a flowchart of the XGBoost learning algorithm used in this invention to construct an XGBoost copper matte grade prediction model;

[0050] Figure 4 This is a structural diagram of the copper matte grade prediction model based on the Stacking ensemble learning algorithm of this invention;

[0051] Figure 5 This is a detailed diagram of the structure of the copper matte grade prediction model based on the Stacking ensemble learning algorithm of this invention.

[0052] Figure 6 This is a comparison chart of the prediction results and actual results of the copper matte grade prediction model based on the Stacking ensemble learning algorithm of this invention;

[0053] Figure 7 This is a diagram showing the prediction results of the Stacking ensemble learning model for copper matte grade, and the prediction results of the random forest and XGBoost copper matte grade prediction models, respectively. Detailed Implementation

[0054] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0055] In this embodiment, the implementation environment is an Intel Core i7-6700HQ CPU, 3.5GHz, 12GB of memory, Windows 10 operating system, and the algorithm is implemented using Python 3.6.8.

[0056] This experiment collected 1217 data points from a copper smelting enterprise in Southwest my country from December 26, 2018 to April 25, 2022. Due to the limited training data, it was necessary to augment the training data to increase its quantity. The training and test sets were divided at a 9:1 ratio. After obtaining 1096 training data points, the SMOTE resampling algorithm was used to expand the original training set from 1096 data points to 15000 data points. The expanded dataset was then merged with the original training set to form the final training set for training the machine learning model.

[0057] Example 1

[0058] like Figure 1 As shown, the data-driven method for predicting matte grade in copper smelting includes the following steps:

[0059] Step 1: Selection of relevant characteristic data; using literature analysis and Pearson correlation coefficient analysis, factors strongly correlated with matte grade were selected: oxygen concentration, amount of copper concentrate added, amount of slag-forming agent silica added, air volume, amount of diesel added, amount of coal added, Cu content in copper concentrate, Fe content in copper concentrate, S content in copper concentrate, Si content in copper concentrate, As content in copper concentrate, Ca content in copper concentrate, Zn content in copper concentrate, H2O content in copper concentrate, Si content in flux silica, Fe content in flux silica, H2O content in flux silica, and oxygen-related quantities, as shown in Table 1.

[0060] Step 2: Data preprocessing; Data preprocessing includes missing data handling, outlier handling, and data normalization; Missing data handling uses deletion, imputation, or no processing; Outlier handling uses the Isolation Forest algorithm; Data normalization algorithms include min-max normalization and standard deviation normalization.

[0061] Step 3: Data augmentation;

[0062] Step 3.1: Divide the data preprocessed in Step 2 into a 9:1 training set and a test set, and use a 5-fold cross-validation method to divide the data. That is, divide the training set into 5 parts, of which 4 parts are training samples and 1 part is the test set sample.

[0063] Step 3.2: The training set obtained in Step 3.1 is expanded using the SMOTE resampling algorithm to obtain an expanded training set.

[0064] Table 1

[0065]

[0066] Step 4, as follows Figure 2 As shown, a random forest model for predicting the grade of copper enamel is constructed.

[0067] The steps include:

[0068] Step 4.1: Construct a random forest copper grade prediction model using the expanded training set obtained in Step 3.2, specifically as follows:

[0069] Step 4.1.1: Using the bagging algorithm Where K is the number of trees, f k (x) represents the k-th decision tree. A random forest copper grade prediction model I is constructed.

[0070] Step 4.1.2: Determine the parameters and range of the expanded training set in Step 3.2, as shown in Table 2. Train the random forest copper grade prediction model I in Step 4.1.1 using the particle swarm optimization algorithm to obtain the optimal parameters of the output model.

[0071] Table 2

[0072]

[0073] The optimal parameters for the output model are: number of decision trees n_estimators = 600, maximum number of features max_features = 4, maximum depth of decision trees max_depth = 25, and other remaining parameters use the model's default parameters.

[0074] Step 4.1.3: Construct a random forest copper grade prediction model by adjusting the optimal parameters of the model;

[0075] Step 4.2: Use the test set obtained in Step 3.1 to predict the random forest copper grade prediction model in Step 4.1 to obtain the prediction set.

[0076] Step 5, as follows Figure 3 As shown, the steps for constructing an XGBoost model to predict the grade of copper matte include:

[0077] Step 5.1: Construct an XGBoost copper matte grade prediction model using the expanded training set obtained in Step 3.2, specifically as follows:

[0078] Step 5.1.1: Using the boosting algorithm:

[0079] in Let f be the prediction result for the i-th sample, K be the total number of trees, and f be the prediction result for the i-th sample. k For the k-th decision tree, construct XGBoost copper matte grade prediction model I;

[0080] Step 5.1.2: Determine the parameters and range of the expanded training set in Step 3.2, as shown in Table 3. Train the XGBoost copper matte grade prediction model I in Step 5.1.1 using the particle swarm optimization algorithm to obtain the optimal parameters of the output model, as shown in the optimal parameter model.

[0081] Table 3

[0082]

[0083] The optimal parameter combination obtained from the final training is: number of decision trees n_estimators = 463, learning rate = 0.13, maximum depth of decision trees max_depth = 8, and other remaining parameters use the model's default parameters.

[0084] Step 5.1.3: By outputting the optimal parameters of the model, according to the algorithm: Where y i Let i be the true value of the i-th sample. For sample x i The training error, Ω(f) k Let ) be the regularization term of the k-th decision tree, defined as:

[0085]

[0086] Where T is the number of leaf nodes, w is the score of the leaf nodes, and γ and λ are two trade-off factors, thus constructing the XGBoost model for predicting the grade of copper matte.

[0087] Step 5.2: Use the test set obtained in Step 3.1 to predict the XGBoost copper matte grade prediction model in Step 5.1 to obtain the prediction set;

[0088] Step 6, as follows Figure 4 and 5 As shown, using the random forest copper grade prediction model in step 4 and the XGBoost copper grade prediction model in step 5 as base learners, and the linear regression model as the meta learner, a stacking ensemble learning copper grade prediction model is constructed to predict the copper grade.

[0089] The specific steps of Stacking ensemble learning in step 6 are as follows:

[0090] Step 6.1: The prediction set obtained by the random forest copper grade prediction model in Step 4.2 and the prediction set obtained by the XGBoost copper grade prediction model in Step 5.2 are used as the new features x1 and x2 of the training set of the second layer model; the average of the prediction sets obtained by the random forest copper grade prediction model in Step 4.2 and the average of the prediction sets obtained by the XGBoost copper grade prediction model in Step 5.2 are used as the new features y1 and y2 of the test set of the second layer model.

[0091] Step 6.2: Use a multiple linear regression learner to train the new features x1 and x2 in the training set from Step 6.1 to obtain the multiple linear regression model;

[0092] Step 6.3: Use the new features y1 and y2 of the test set obtained in Step 6.1 to predict the multiple linear regression model in Step 6.2, and finally obtain the prediction results of the Stacking ensemble learning model for predicting the grade of copper matte. The prediction results are shown in the figure below. Figure 6 As shown.

[0093] The prediction performance of Stacking ensemble learning versus two single models, Random Forest and XGBoost, under optimal parameters is compared in Table 4.

[0094] Table 4 Comparison of prediction results of the three models

[0095]

[0096] Experimental results show that the Stacking ensemble learning model, with random forest and XGBoost as base learners and linear regression as meta learners, has a relative error between predicted and actual values ​​in the range of 0-1%, which is better than the two single models.

[0097] from Figure 7 As can be seen, compared with the two single models mentioned above, the Stacking ensemble learning model has lower MAE (mean absolute error), RMSE (root mean square error), and MAPE (mean absolute percentage error). The Stacking ensemble learning model's MAE is reduced by 8.53% compared with the best-performing single model XGBoost, the RMSE is reduced by 8.10% compared with the XGBoost model, the MPAE is reduced by 8.23% compared with the XGBoost model, and the R2 value is increased by 0.54% compared with the XGBoost model. From each indicator, it is clear that the Stacking ensemble learning model has a better prediction effect on copper matte grade than the two single models.

[0098] Example 2

[0099] This system, based on a data-driven method for predicting the grade of copper matte in copper smelting, can be directly used for predicting the grade of copper matte. The system's overall architecture consists of four layers: a data acquisition layer, a data processing layer, a prediction layer, and an application layer. After inputting the values ​​of each independent variable into the interface, the predicted matte grade can be displayed in real time.

[0100] The specific embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Within the scope of knowledge possessed by those skilled in the art, various changes can be made without departing from the spirit of the present invention.

Claims

1. A data-driven method for predicting the grade of matte in copper smelting, characterized by the following steps: include: Step 1: Selection of relevant feature data; Step 2: Data preprocessing; Step 3: Data augmentation; Step 4: Construct a random forest model for predicting copper matte grade; Step 5: Construct an XGBoost model for predicting the grade of copper matte; Step 6: Using the random forest copper grade prediction model in Step 4 and the XGBoost copper grade prediction model in Step 5 as base learners, and the linear regression model as the meta learner, construct a stacking ensemble learning copper grade prediction model to predict the copper grade. In step 1, the relevant characteristic data selected are factors that are strongly correlated with the grade of matte: oxygen concentration, amount of copper concentrate added, amount of slag-forming agent silica added, air volume, amount of diesel added, amount of coal added, Cu content in copper concentrate, Fe content in copper concentrate, S content in copper concentrate, Si content in copper concentrate, As content in copper concentrate, Ca content in copper concentrate, Zn content in copper concentrate, H2O content in copper concentrate, Si content in flux silica, Fe content in flux silica, H2O content in flux silica, and oxygen-related amounts; The data augmentation in step 3 includes the following steps: Step 3.1: Divide the preprocessed data from Step 2 into a training set and a test set; Step 3.2: The training set obtained in Step 3.1 is augmented using the SMOTE resampling algorithm to obtain an augmented training set; Step 4, which involves constructing a random forest model for predicting the grade of copper enamel, includes the following steps: Step 4.1: Construct a random forest copper grade prediction model using the expanded training set obtained in Step 3.2; Step 4.2: Use the test set obtained in Step 3.1 to predict the random forest copper grade prediction model in Step 4.1 to obtain the prediction set; Step 5 involves constructing the XGBoost copper matte grade prediction model, which includes the following steps: Step 5.1: Construct an XGBoost copper matte grade prediction model using the expanded training set obtained in Step 3.2; Step 5.2: Use the test set obtained in Step 3.1 to predict the XGBoost copper matte grade prediction model in Step 5.1 to obtain the prediction set; The specific steps of Stacking ensemble learning in step 6 are as follows: Step 6.1: Combine the prediction set obtained from the random forest copper grade prediction model in Step 4.2 with the prediction set obtained from the XGBoost copper grade prediction model in Step 5.2 to form the training set of the second-layer model. , The average of the prediction sets obtained from the random forest copper grade prediction model in step 4.2 and the average of the prediction sets obtained from the XGBoost copper grade prediction model in step 5.2 are used as the new features of the test set for the second-layer model. , ; Step 6.2: Use a multiple linear regression learner to process the new features of the training set from Step 6.

1. , The model is trained to obtain a multiple linear regression model. Step 6.3: The new features of the test set obtained in Step 6.1 , The prediction results of the multivariate linear regression model in step 6.2 are obtained by making predictions on the model and finally obtaining the prediction results of the Stacking ensemble learning model for predicting the grade of copper matte.

2. The data-driven method for predicting matte grade in copper smelting according to claim 1, characterized in that: The data preprocessing in step 2 includes missing data processing, outlier processing, and data normalization.

3. The data-driven method for predicting matte grade in copper smelting according to claim 1, characterized in that: Step 4.1, which involves expanding the training set to construct a random forest copper grade prediction model, specifically includes: Step 4.1.1: Using the bagging algorithm ,in The number of trees. For the first k Using decision trees, a random forest copper enamel grade prediction model I was constructed. Step 4.1.2: Determine the parameters and range of the expanded training set in Step 3.2, and train the random forest copper grade prediction model I in Step 4.1.1 using the particle swarm optimization algorithm to obtain the optimal parameters of the output model; Step 4.1.3: Construct a random forest copper ore grade prediction model by adjusting the optimal parameters of the model.

4. The data-driven method for predicting matte grade in copper smelting according to claim 1, characterized in that: Step 5.1, which involves expanding the training set to construct an XGBoost model for predicting the grade of copper matte, specifically involves: Step 5.1.1: Using the boosting algorithm: ,in For the first i Prediction results for each sample K The total number of trees. For the first k Using decision trees, construct XGBoost copper matte grade prediction model I; Step 5.1.2: Determine the parameters and range of the expanded training set in Step 3.2, and train the XGBoost copper matte grade prediction model I in Step 5.1.1 using the particle swarm optimization algorithm to obtain the optimal parameters of the output model; Step 5.1.3: By outputting the optimal parameters of the model, according to the algorithm: ,in For the first i The true value of each sample For the sample The training error, For the first k The regularization term of a decision tree is defined as follows: in, T The number of leaf nodes. w The score for the leaf nodes. and These are two trade-off factors that form the basis of the XGBoost copper matte grade prediction model.

5. A system based on the data-driven method for predicting the grade of copper smelting matte according to any one of claims 1 to 4, which can be directly used for predicting the grade of copper smelting matte.

Images