Optimización en la selección de sistemas de levantamiento artificial aplicando ciencia de datos: Campo Lago Agrio

Abstract

This project addresses the optimization of artificial lift system selection, with the aim of improving operational efficiency in wells from the Oriente Basin in Ecuador. The hypothesis established was that the application of machine learning models allows accurate identification of the most suitable lift system according to the production characteristics of each well. The justification lies in the need to reduce uncertainty in decision-making and to make more efficient use of resources within the oil industry. For the development of the study, historical production records were collected, processed through cleaning, normalization, and winsorization techniques, and class imbalance methodologies were applied. Subsequently, multiclass classification models were trained, including Decision Tree, Random Forest, XGBoost, and Stacking, which were validated using accuracy, precision, recall, and F1-score metrics. The results showed agreement between model predictions and installed systems in 83.3 % of the analyzed cases. XGBoost and Stacking achieved the best performances, with accuracy values above 99 %, while Random Forest and Decision Tree presented intermediate results. It is concluded that the use of machine learning models constitutes a reliable and applicable tool in the oil industry for predicting and recommending artificial lift systems in wells with similar conditions. Keywords: Machine learning, data mining, artificial lift, prediction, optimization