A Machine Learning Approach for Intervention Recommendation and Production Forecasting in Mature Ecuadorian Oil Fields

A machine learning (ML) approach is presented to optimize decision-making in well intervention planning for mature oil fields in Ecuador. Supervised classification and regression are used to automate two critical tasks: (1) recommending the most appropriate intervention type, and (2) forecasting the expected fluid production increment (triangle BFPD) postintervention. Drawing on four integrated historical databases encompassing over 76,000 records of production, pressure, and reservoir attributes, a structured pipeline was implemented, including pre-processing, feature selection, and class balancing using SMOTE + TomekLinks. For activity classification, Random Forest paired with SelectKBest yielded the best performance, with an F1-macro score of 0. 7 9. For production forecasting, XGBoost achieved an R2 of 0.728 and MAE of 219.11 for REPERF interventions. Because of the poor predictive performance of MATRIX STIMULATION (R2<0), clustering (K-means + PCA) was explored, revealing internal heterogeneity, but insufficient to improve accuracy, highlighting the need for enhanced data granularity. This dual-model approach reduces the manual candidate screening time by over 2,000 hours annually, introduces data-driven prioritization in campaign planning, and lays the foundation for scalable, real-time recommendation systems in high-density oil field operations.