An online prediction method and model construction for copper flash smelting matte grade

Abstract

The present invention relates to the technical field of copper flash smelting matte grade prediction, and in particular to a method and model construction for online copper flash smelting matte grade prediction. Starting from the actual correlation between matte products and production inputs, the model is trained, verified and tested using input data and output data of historical matte production. Then, the copper flash smelting matte input data within a certain time period is acquired in real time and substituted into a prediction model as a model input for online calculation. The predicted value of the matte grade released at the corresponding time point is obtained in real time. This solves the problem of mismatch between input data and output data in the prediction model constructed in the prior art, and realizes accurate online prediction of the copper flash smelting matte grade.

Description

Technical Field

[0001] The present invention relates to the technical field of copper flash smelting matte grade prediction, and in particular to an online copper flash smelting matte grade prediction method and model construction. Background Art

[0002] Copper flash smelting is a copper smelting process that combines roasting and smelting in a single device. It is the most widely used sulfide ore copper smelting method in existing technologies. Copper matte is the main product of the copper flash smelting process. The percentage of copper in the matte is the main factor in evaluating the matte grade, and the matte grade will have a significant impact on the downstream blowing process. In production practice, it is necessary to adjust the production process through real-time monitoring of the matte grade to ensure the output quality of subsequent processes. However, the current online detection equipment for copper matte grade is still immature, making it difficult to achieve accurate detection and draw conclusions. Therefore, the use of model prediction has become a feasible solution for matte grade prediction and has been studied accordingly.

[0003] For example, Patent No. 202410489618.2 discloses the collection of actual production data (including mineral unit oxygen consumption, solvent ratio, oxygen enrichment concentration, copper content, sulfur content, iron content, silica content, and magnesium oxide content) during a stable production cycle of a copper smelter. The collected data is normalized, and 80% of the processed data is used as training data and 20% as test data. Based on the existing GA-BP model, the genetic algorithm (GA) is improved to obtain an improved genetic algorithm (GA). The specific improvement is to add an adaptive crossover probability adjustment mechanism and an adaptive mutation probability adjustment mechanism to the genetic algorithm (GA). The adaptive crossover probability adjustment mechanism is used to adjust the population update speed in the early genetic stage and the stability in the late genetic stage of the genetic algorithm (GA), and the adaptive mutation probability adjustment mechanism is used to improve the population diversity; the BP neural network model is optimized by the improved genetic algorithm (GA), and the actual production data (mineral unit oxygen consumption, solvent ratio, oxygen enrichment concentration, copper content, sulfur content, iron content, silica content, and magnesium oxide content) are used as inputs to the BP neural network model to ultimately derive the copper matte grade value; and the prediction accuracy of the improved GA-BP model is verified.

[0004] Another example: Patent application number 202411808546.X discloses an improved genetic algorithm core mechanism to enhance optimization efficiency and model prediction accuracy. The method includes collecting actual production data from a copper smelter during its stable production cycle; improving the existing GA-BP neural network by adding an adaptive crossover probability adjustment mechanism and an adaptive mutation probability adjustment mechanism to the genetic algorithm. The adaptive crossover probability adjustment mechanism not only relies on individual fitness but also combines global and local information to enhance global search capabilities and local optimization capabilities. The adaptive mutation probability adjustment mechanism not only combines individual fitness but also introduces global entropy and dynamic noise effects, while allowing the mutation probability to increase randomly in the later stages of evolution to escape the local optimum. Using the improved GA-BP neural network, actual production data is used as input to predict the grade of oxygen-enriched bottom-blown copper matte and output the copper matte grade value.

[0005] It can be seen that the existing technology adopts statistical and machine learning algorithms, uses the product quality index inspection data collected during the actual production process as the model output annotation value, and uses the historical production input data synchronized with the output product as the model input data to jointly construct the model training and test data set, and obtains the copper matte grade quality index prediction model through model training and testing. However, the copper flash smelting process has the following conditions: (1) The flash smelting pool is long, and the copper discharge port at the end of the flash smelting pool is more than ten meters away from the flash smelting tower molten pool area, resulting in a long time delay from feeding to copper matte outflow; (2) There are complex physical and chemical reactions and heat transfer and mass transfer processes in the copper flash smelting pool, so that the copper matte composition produced at a certain point in time is actually the result of the combined effect of all feeding and production conditions in a long period of time before that point in time. The copper matte grade prediction model in the existing technology can only take into account the influence of the production conditions and feeding synchronized with the copper matte discharge on the copper matte grade, resulting in poor accuracy of the copper matte grade prediction results. Summary of the Invention

[0006] Based on the above technical problems, the present invention provides a method and model construction for online prediction of copper matte grade in copper flash smelting, which takes into account the feeding and production conditions during the period from feeding to residence time in the flash smelting pool before the output (release) of copper matte, so as to construct an online prediction model for copper matte grade, solve the technical problem of mismatch between input data and output data in the existing technology, and thus improve the accuracy of copper matte grade prediction in copper flash smelting.

[0007] The specific technical solutions are:

[0008] One of the purposes of the present invention is to provide a method for online prediction of copper matte grade in copper flash smelting, comprising the following steps:

[0009] S1: Set the data sampling time points of the copper flash smelting production process at equal intervals n;

[0010] S2: Collect input data according to the data sampling time, and collect and calculate the average residence time A from feeding to releasing copper matte according to the following formula:

[0011] A = (S × H) ÷ (Q × C)

[0012] Where S is the effective area of ​​the flash smelting pool, in m2; H is the average thickness of the copper matte layer in the flash smelting pool, in m; Q is the average amount of copper concentrate fed per unit time, in m 3 / h; C is the copper matte yield, unit is %;

[0013] S3: collecting matte output data at the data sampling time point, and constructing an input time period by tracing back A from the data sampling time point, and intercepting the input data within the input time period; marking and matching the input data with the matte output data one by one into a data set;

[0014] S4: Divide the data set into a ratio of 7:2:1 for training data set: validation data set: test data set, and construct a primary model using an artificial intelligence algorithm. Use the data set to train, validate, and test the model, and select the best model T based on the test results to obtain the copper matte grade prediction model.

[0015] S5: Real-time acquisition of input data and matte discharge time point to input into matte grade prediction model T, and the predicted value of matte grade at the matte discharge time point is obtained through model calculation.

[0016] The input data created by the present invention includes but is not limited to feeding data and equipment operation data; the feeding data includes but is not limited to feeding copper content, feeding sulfur content, feeding iron content and feeding silica content; the equipment operation data includes but is not limited to feeding amount per unit time, wind flow rate and oxygen flow rate; the copper matte output data is copper matte grade detection data.

[0017] The feed rate per unit time of the present invention includes but is not limited to the feed rate of copper concentrate, the feed rate of slag forming agent and the feed rate of fly ash; the air flow rate includes but is not limited to the process air flow rate and the dispersed air flow rate; the oxygen flow rate includes but is not limited to the central oxygen flow rate, the regulating oxygen flow rate and the secondary oxygen lance oxygen flow rate.

[0018] The present invention creates the step S5, deploying the copper matte grade prediction model T in a computer, which is connected to the batching and control system of the copper flash smelting pool and can read the feeding data of the copper flash smelting process.

[0019] The artificial intelligence algorithm created by the present invention is capable of supporting vector machines, logistic regression, random forests and / or deep learning.

[0020] A second object of the present invention is to provide a model construction method for online prediction of copper flash smelting matte grade, comprising the following steps:

[0021] (1) Determine the prediction target: copper flash smelting matte grade;

[0022] (2) Setting model evaluation criteria: using mean square error (MSE), root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R2) data to evaluate model prediction accuracy;

[0023] (3) Data collection and cleaning: Collect input data and copper matte output data; process missing data, abnormal data, and data standardization;

[0024] (4) Through feature selection or feature extraction, the original data is converted into effective input and the corresponding prediction target data is constructed into a data set;

[0025] (5) Select several artificial intelligence regression algorithms to construct primary models respectively, and then select the best one based on the actual prediction effect of each primary model;

[0026] (6) Divide the data set into training set, validation set and test set. Use the training set to train the model, use the validation set to adjust and optimize the model parameters, and then use the test set to test the model prediction effect.

[0027] The present invention creates a method for converting the original data into an effective input:

[0028] a. Set the time point for copper matte sampling as t, then trace back an average residence time A from time t, construct an input time period from tA to time t, and intercept the input data within the input time period to form a set of input data; at the set sampling time point t, sample the discharged copper matte, number it, and send it for inspection, and record the copper matte grade of the sample to construct a copper matte output data;

[0029] b. Assemble the model input and output data, set the number of input data to E, and calculate the number of input data collection times L corresponding to a copper matte output data according to L=A÷n; suppose that the copper matte product is sampled for the kth time at sampling time t, and the copper matte output data of the copper matte grade obtained by chemical analysis is Q k , which corresponds to a set of input data I k i,j (i=1,2,3…E; j=1,2,3…L); the input data collection times corresponding to j=1,2,3…L are tA, t-A+n, t-A+2n, …, t respectively;

[0030] c. Record the input data and copper matte output data collected in step b and organize them into input data I k 1,1 , I k 1,2 , I k 1,3 ...I k 1,E , I k 2,1 , I k 2,2 , I k 2,3 ...I k 2,E , I k 3,1 , I k 3,2 , I k 3,3 ...I k 3,E ,……,I k L,1 , I k L,2 , I k L,3 ...I k L,E and copper matte output data Q k One-to-one corresponding data sets.

[0031] Compared with the prior art, the technical effects created by the present invention are embodied in:

[0032] The present invention creates a prediction model using an artificial intelligence algorithm, uses historical data from the copper flash smelting production process to train, verify and test the model, and obtains a copper flash smelting matte grade prediction model. The prediction model is deployed in a computer, and the computer is used to obtain input data within a certain period of time in real time. The data is substituted into the copper flash smelting matte grade prediction model as model input for online calculation. Based on the actual correlation between the copper matte product and the production input, the predicted value of the copper matte grade released at the corresponding time point is predicted in real time, thereby solving the problem of mismatch between input data and output data in the prediction model constructed in the prior art, and realizing online and accurate prediction of the copper flash smelting matte grade. DETAILED DESCRIPTION

[0033] In order to facilitate those skilled in the art to correctly understand the present invention and enable those skilled in the art to fully understand the technical content of the present invention, the technical solution of the present invention is further explained below in conjunction with specific implementation methods, but this explanation does not limit the scope of protection required for the present invention. The scope of protection of the present invention by those skilled in the art cannot be limited to the following explanations. Any equivalent replacement or change made by any those skilled in the art or persons familiar with the technology in this field, on the basis of the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

[0034] Example

[0035] The online method for predicting copper matte grade from copper flash smelting includes:

[0036] The copper concentrate samples were taken from the feed belt every 5 minutes and sent to the testing laboratory for analysis of feed composition. The copper content, sulfur content, iron content, silica content and other impurity contents in the copper concentrate fed in during this period were recorded to obtain the feeding data (as shown in Table 1).

[0037] Table 1 Sample table for collecting and recording feeding data

[0038]

[0039] At the same time period as the sampling of the feed copper concentrate, equipment operating data such as the feed amount per unit time (total amount of copper concentrate, slagging agent and fly ash), air flow (total flow of process air and dispersed air) and oxygen flow (total flow of central oxygen, regulating oxygen and secondary oxygen lance oxygen) are read from the copper flash smelting pool control system (as shown in the sample table in Table 2).

[0040] Table 2 Sample table for recording data of equipment operation

[0041]

[0042] The feeding data and equipment operation data collected in Tables 1 and 2 are sorted according to the sampling time to obtain the input data (as shown in the sample table in Table 3).

[0043] Table 3

[0044]

[0045] When releasing copper matte, samples are taken, numbered and sent for inspection every 5 minutes to obtain the copper content in the copper matte, and the copper flash smelting copper matte output data is obtained (as shown in Table 4).

[0046] Table 4 Sample table of copper matte output data records

[0047]

[0048] The effective area S of the copper flash smelting pool was 150 m2, the average thickness H of the copper matte in the pool was 1.2 m, and the average amount of material fed per unit time Q was 30 m3. 3 / h, the copper matte yield C is 35%. According to the formula of average residence time A= (S×H) / (Q×C), the average residence time of the feed components in the copper flash smelting pool from feeding to releasing the copper matte is ≈17h.

[0049] For each matte discharge batch, the starting time is calculated by tracing back the average residence time of the input materials (the total amount of copper concentrate, slag forming agent, and fly ash) in the copper flash smelting pool from the time of matte discharge and sampling. From the starting time to the time of matte discharge and sampling, 204 data records (A ÷ n) = 17 h ÷ 5 min = (17 × 60) ÷ 5 are selected from the input data (as shown in the sample table in Table 3) to construct a set of input data. From the copper flash smelting matte output data (as shown in the sample table in Table 4), the matte grade data corresponding to this matte discharge batch is selected as a matte output data item. The input data are mapped one-to-one with the matte output data to construct Dataset 1 for copper flash smelting matte, which is recorded as shown in the sample table in Table 5.

[0050] Table 5. Data set for copper flash smelting matte - sample record

[0051]

[0052] The copper flash smelting matte dataset was divided into three sub-datasets: training set, validation set and test set according to the ratio of 7:2:1; artificial intelligence machines such as support vector, random forest and deep learning were used for modeling algorithms respectively. The training set was used for model training, the validation set was used to verify the trained model and adjust the parameters, and the test set was used to test the model after verification and parameter adjustment. According to the test results, the random forest model was selected as the copper flash smelting matte grade prediction model.

[0053] The copper flash smelting matte grade prediction model is deployed in a computer at the copper flash smelting matte production site, and the computer is connected to the copper flash smelting pool batching system and control system through an intranet. The computer can read the copper concentrate in the batching system, including but not limited to copper content, sulfur content, iron content, silicon dioxide content and other impurity contents; at the same time, the computer can read the control system data, including but not limited to the feed amount per unit time (copper concentrate feed amount, slagging agent feed amount and fly ash feed amount), air flow (process air flow rate and dispersed air flow rate), oxygen flow rate (central oxygen flow rate, regulating oxygen flow rate and secondary oxygen lance oxygen flow rate); and can trace back an average residence time A (17 hours) from the current time point, and intercept all sampled input data in the time period from t-17 hours to t, as model input and substituted into the copper flash smelting matte grade prediction model. The model operation outputs the result, that is, the predicted value of the copper content in the matte released at the time point t.

[0054] This prediction method starts from the actual correlation between copper matte products and production inputs, uses the input data and output data of historical copper matte production to conduct model training, verification and testing, and then obtains the copper matte input data of copper flash smelting within a certain time period in real time, substitutes it into the prediction model as the model input for online calculation, and obtains the predicted value of the copper matte grade released at the corresponding time point in real time. This solves the problem of mismatch between input data and output data of the prediction model constructed by the existing technology, and realizes the online and accurate prediction of copper matte grade from copper flash smelting.

[0055] Tables 1 to 5 in the present invention are sample tables, which are intended to accurately express the process of constructing the prediction model and processing the data to form a data set in the present invention, so that those skilled in the art can make an accurate understanding. Moreover, the actual model training requires a large amount of data, and the data set is divided by random sampling. It is neither necessary nor meaningful to list all of them in the table. Therefore, ellipsis is used in the table instead.

[0056] In addition, the data set division ratio involved in the invention is a reference. Those skilled in the art may adopt other data set division ratios, which may be specifically adjusted according to the common sense basic requirements of those skilled in the art and artificial intelligence algorithm model training.

Claims

1. A method for online prediction of copper matte grade in copper flash smelting, characterized in that: The following steps are involved: S1: Set the data sampling time points of the copper flash smelting production process at equal intervals n; S2: Collect input data at the data sampling time points, and collect and calculate the average residence time A from the time of feeding to the time of releasing the copper matte according to the following formula: A=(S×H)÷(Q×C); Where S is the effective area of ​​the flash smelting pool, in m2; H is the average thickness of the copper matte layer in the flash smelting pool, in m; Q is the average amount of copper concentrate fed per unit time, in m 3 / h; C is the copper matte yield, unit is %; S3: Copper matte output data is collected at the data sampling time point, and the input time period is constructed by tracing back A from the data sampling time point, and the input data within the input time period is intercepted; the average residence time A is used to construct the input time period that matches the copper matte output data, and the average residence time is dynamically adjusted with the changes in the effective area S of the flash smelting pool, the average thickness H of the copper matte layer in the flash smelting pool, the average copper concentrate feed amount Q per unit time, and the average copper concentrate feed amount C per unit time; The input data and copper matte output data are marked and matched one by one to form a data set; S4: Divide the data set into a ratio of 7:2:1 for training data set: validation data set: test data set, and construct a primary model using an artificial intelligence algorithm. Use the data set to train, validate, and test the model, and select the best model T based on the test results to obtain the copper matte grade prediction model. S5: Real-time acquisition of input data and matte discharge time point to input into matte grade prediction model T, and the predicted value of matte grade at the matte discharge time point is obtained through model calculation.

2. The method for online prediction of copper flash smelting matte grade according to claim 1, characterized in that: The input data includes feeding data and equipment operation data; the feeding data includes feeding copper content, feeding sulfur content, feeding iron content and feeding silica content; the equipment operation data includes feeding amount per unit time, wind flow rate and oxygen flow rate; the copper matte output data is copper matte grade detection data.

3. The method for online prediction of copper flash smelting matte grade according to claim 2, characterized in that: The feed rate per unit time includes the copper concentrate feed rate, the slag forming agent feed rate and the fly ash feed rate; the air flow rate includes the process air flow rate and the dispersed air flow rate; the oxygen flow rate includes the central oxygen flow rate, the regulating oxygen flow rate and the secondary oxygen lance oxygen flow rate.

4. The method for online prediction of copper flash smelting matte grade according to claim 1, wherein: In step S5, the copper matte grade prediction model T is deployed in a computer, which is connected to the batching and control system of the copper flash smelting pool and can read the feeding data of the copper flash smelting process.

5. The method for online prediction of copper flash smelting matte grade according to claim 1, characterized in that: The artificial intelligence algorithm can be support vector machine, random forest and / or deep learning.

6. The method for online prediction of copper flash smelting matte grade according to claim 1, characterized in that: The step S4, the steps of constructing a model using an artificial intelligence logistic regression algorithm are as follows: (1) Determine the prediction target: copper flash smelting matte grade; (2) Setting model evaluation criteria; (3) Data collection and cleaning: Collect input data and copper matte output data; process missing data, abnormal data, and data standardization; (4) Through feature selection or feature extraction, the original data is converted into effective input and the corresponding prediction target data is constructed into a data set; (5) Select several artificial intelligence logistic regression algorithms to construct primary models respectively, and then select the best one based on the actual prediction effect of each primary model; (6) Divide the data set into training set, validation set and test set. Use the training set to train the model, use the validation set to adjust and optimize the model parameters, and then use the test set to test the model prediction effect.

7. The method for online prediction of copper flash smelting matte grade according to claim 6, characterized in that: The original data is converted into a valid input method: a. Set the time point for copper matte sampling as t, then trace back an average residence time A from time t, construct an input time period from tA to time t, and intercept the input data within the input time period to form a set of input data; at the set sampling time point t, sample the discharged copper matte, number it, and send it for inspection, and record the copper matte grade of the sample to construct a copper matte output data; b. Assemble the model input and output data, set the number of input data to E, and calculate the number of input data collection times L corresponding to a copper matte output data according to L=A÷n; suppose that the copper matte product is sampled for the kth time at sampling time t, and the copper matte output data of the copper matte grade obtained by chemical analysis is Q k , which corresponds to a set of input data I k i,j (i=1,2,3...E; j=1,2,3...L); the input data collection times corresponding to j=1,2,3...L are tA, t-A+n, t-A+2n,..., t respectively; c. Record the input data and copper matte output data collected in step b and organize them into input data I k 1,1 , I k 1,2 , I k 1,3 ...I k 1,E , I k 2,1 , I k 2,2 , I k 2,3 ...I k 2,E , I k 3,1 , I k 3,2 , I k 3,3 ...I k 3,E ,……,I k L,1 , I k L,2 , I k L,3 ...I k L,E and copper matte output data Q k One-to-one corresponding data sets.

8. The method for online prediction of copper flash smelting matte grade according to claim 6, characterized in that: In step (2), mean square error (MSE), root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R2) data are used to evaluate the model prediction accuracy.

9. A method for constructing a model for online prediction of copper matte grade in copper flash smelting, characterized in that: The following steps are involved: (1) Determine the prediction target: copper flash smelting matte grade; (2) Setting model evaluation criteria: using mean square error (MSE), root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R2) data to evaluate model prediction accuracy; (3) Data collection and cleaning: Collect input data and copper matte output data; process missing data, abnormal data, and data standardization; (4) by feature selection or feature extraction, the raw data is converted into effective input and a data set is constructed corresponding to the determined prediction target data; the raw data is converted into effective input in the following manner: the time point of copper matte sampling is set as t, then an average residence time A is traced back from time t, an input time period is constructed from tA to time t, and the input data within the input time period is intercepted to form a set of input data; the average residence time A is used to construct an input time period that matches the copper matte output data, and the average residence time is dynamically adjusted with the changes in the effective area S of the flash smelting pool, the average thickness H of the copper matte layer in the flash smelting pool, the average copper concentrate feed amount Q per unit time, and the average copper concentrate feed amount C per unit time; (5) Select several artificial intelligence logistic regression algorithms to construct primary models respectively, and then select the best one based on the actual prediction effect of each primary model; (6) Divide the data set into training set, validation set and test set. Use the training set to train the model, use the validation set to adjust and optimize the model parameters, and then use the test set to test the model prediction effect.

10. The model construction method for online prediction of copper flash smelting matte grade according to claim 9, characterized in that: The original data is converted into a valid input method: a. Set the time point for copper matte sampling as t, then trace back an average residence time A from time t, construct an input time period from tA to time t, and intercept the input data within the input time period to form a set of input data; at the set sampling time point t, sample the discharged copper matte, number it, and send it for inspection, and record the copper matte grade of the sample to construct a copper matte output data; b. Assemble the model input and output data, set the number of input data to E, and calculate the number of input data collection times L corresponding to a copper matte output data according to L=A÷n; suppose that the copper matte product is sampled for the kth time at sampling time t, and the copper matte output data of the copper matte grade obtained by chemical analysis is Q k , which corresponds to a set of input data I k i,j (i=1,2,3...E; j=1,2,3...L); the input data collection times corresponding to j=1,2,3...L are tA, t-A+n, t-A+2n,..., t respectively; c. Record the input data and copper matte output data collected in step b and organize them into input data I k 1,1 , I k 1,2 , I k 1,3 ...I k 1,E , I k 2,1 , I k 2,2 , I k 2,3 ...I k 2,E , I k 3,1 , I k 3,2 , I k 3,3 ...I k 3,E ,……,I k L,1 , I k L,2 , I k L,3 ...I k L,E and copper matte output data Q k One-to-one corresponding data sets.

Images