Modelación del porcentaje de insolubles en concentrado final de cobre en función de variables operacionales del proceso en Compañía Minera Doña Inés de Collahuasi.
Loading...
Date
2025
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Universidad de Concepción
Abstract
Este informe presenta un estudio de la variación de porcentaje de insolubles presentes en el concentrado final de cobre a través del uso de modelos predictivos supervisados con la información recopilada de 3 años y de 6 meses, cada uno de distintos periodos. Para esto se seleccionan datos de variables que tengan significancia en la variación de insolubles, esto a través de una limpieza de datos, o sea eliminando datos anormales, atípicos y con valor 0, y luego se analizan estadísticamente, optando por variables que tienen un valor p menor a 0,05 y algunas variables que sean mayor a 0,05 y que impacten de gran manera con los insolubles, con el fin de trabajar con las variables escogidas y que beneficien a los modelos tanto teórica y estadísticamente.
Los modelos predictivos que se usaron fueron supervisados, los cuales trabajan a partir de una variable objetivo (en este caso, porcentaje de insolubles) respecto a las variables independientes, siendo estas obtenidas anteriormente. Se trabajan con modelo de tipo árboles (XGBoost, Gradient Boosting y Random Forest), modelo de redes neuronales (Redes Neuronales Profundas, DNN, y Redes Neuronales Artificiales, ANN) y con modelos que usan máquinas de vectores de soporte, que en este caso se usó la Regresión de Vectores de Soporte (SVR). Se construye una tabla de métricas de desempeño, en donde se comparan los modelos y se selecciona el modelo que entregue valores ideales en dicha tabla.
Los resultados obtenidos en el análisis de datos operacionales muestran que los algoritmos predictivos basados en árboles, especialmente XGBoost, presentan un mejor desempeño al trabajar con información histórica de tres años, debido a su capacidad para manejar grandes volúmenes de datos y relaciones complejas. En cambio, al utilizar un subconjunto reducido de seis meses, las Redes Neuronales Artificiales entregan mejores resultados en términos de ajuste y error promedio respecto a las otras técnicas evaluadas. Esto sugiere que, mientras los métodos tipo árbol se adaptan mejor a series extensas, las redes neuronales responden con mayor eficacia cuando se dispone de un conjunto de datos más limitado y reciente.
This report examines the variation in the percentage of insoluble material in the final copper concentrate using supervised predictive models, based on data collected over two distinct periods: three years and six months. To this end, variables significantly influencing insoluble content were selected through data preprocessing, which included the removal of abnormal, outlier, and zero-value data, followed by statistical analysis. Variables with a p-value below 0.05 were prioritized, while some with higher values were retained due to their practical relevance, ensuring the inclusion of inputs with both statistical and operational significance. The predictive models applied are supervised learning approaches, relying on a target variable (percentage of insoluble) and a set of independent variables derived from the processed dataset. These include tree-based algorithms (XGBoost, Gradient Boosting, and Random Forest), neural network architectures (Artificial Neural Networks, ANN, and Deep Neural Networks, DNN), and a support vector machine model, specifically Support Vector Regression (SVR). A comparative performance metrics table was constructed to identify the model achieving the most accurate and reliable results. The findings indicate that tree-based models, particularly XGBoost, perform optimally with three years of historical data due to their capacity to manage large datasets and capture complex relationships. Conversely, with a smaller six-month dataset, the ANN model demonstrated superior performance in terms of fit and error reduction compared to other approaches. This suggests that tree based algorithms are more suitable for extensive datasets, while neural networks offer greater adaptability and accuracy when working with shorter and more recent data series.
This report examines the variation in the percentage of insoluble material in the final copper concentrate using supervised predictive models, based on data collected over two distinct periods: three years and six months. To this end, variables significantly influencing insoluble content were selected through data preprocessing, which included the removal of abnormal, outlier, and zero-value data, followed by statistical analysis. Variables with a p-value below 0.05 were prioritized, while some with higher values were retained due to their practical relevance, ensuring the inclusion of inputs with both statistical and operational significance. The predictive models applied are supervised learning approaches, relying on a target variable (percentage of insoluble) and a set of independent variables derived from the processed dataset. These include tree-based algorithms (XGBoost, Gradient Boosting, and Random Forest), neural network architectures (Artificial Neural Networks, ANN, and Deep Neural Networks, DNN), and a support vector machine model, specifically Support Vector Regression (SVR). A comparative performance metrics table was constructed to identify the model achieving the most accurate and reliable results. The findings indicate that tree-based models, particularly XGBoost, perform optimally with three years of historical data due to their capacity to manage large datasets and capture complex relationships. Conversely, with a smaller six-month dataset, the ANN model demonstrated superior performance in terms of fit and error reduction compared to other approaches. This suggests that tree based algorithms are more suitable for extensive datasets, while neural networks offer greater adaptability and accuracy when working with shorter and more recent data series.
Description
Tesis presentada para optar al título de Ingeniero/a Civil Metalúrgico/a.
Keywords
Cobre, Industria minera, Algoritmos Procesamiento de datos