Perpustakaan Digital ITB

Well intervention remains one of the most important production enhancement strategies in mature oil fields, particularly for shut-in wells affected by reservoir depletion and artificial lift performance deterioration. However, intervention activities involve significant operational costs and uncertainties, while candidate selection is often based on engineering judgment and manual data evaluation. The complex interaction between reservoir characteristics, production performance, and Electric Submersible Pump (ESP) operating conditions further complicates the identification of wells with the highest probability of intervention success. Therefore, a more objective and data-driven screening approach is required to support intervention planning and reduce decision uncertainty. This study develops a machine learning-based screening framework for evaluating shut-in offshore well intervention candidates in Field-X by integrating reservoir, production, and ESP-related parameters. Historical intervention records from 2019 to 2025 were collected, resulting in approximately 300 validated intervention cases after data quality filtering. Intervention success was defined using a minimum 5% post intervention oil production increase relative to pre-intervention conditions. Multiple supervised classification algorithms, including Random Forest, XGBoost, Gradient Boosting, Extra Trees, Logistic Regression, LightGBM, and CatBoost, were evaluated using K-Fold cross-validation, hyperparameter tuning, and threshold optimization. To improve prediction robustness, a consensus-based approach using majority voting was implemented to generate the final GO/NO-GO recommendation. The results demonstrate that tree-based ensemble models consistently outperformed Logistic Regression, indicating the presence of complex non-linear relationships among reservoir, production, and ESP parameters. The optimal configuration was obtained using a 15% test split, achieving F1-score and ROC AUC values exceeding 0.85. Furthermore, blind test validation using ten unseen wells showed that the consensus-based framework successfully classified all evaluated intervention cases according to their actual outcomes. Feature importance and SHapley Additive exPlanations (SHAP) analyses identified Oil rate, Pump Efficiency, Pump Intake Pressure (PIP), ESP Current, and Gas-Oil Ratio (GOR) as the most influential parameters affecting intervention feasibility. The engineering interpretation indicates that intervention success is governed by the combined interaction between reservoir deliverability and ESP operating conditions rather than a single controlling parameter. The proposed framework provides a practical decision-support tool for objective well intervention screening while maintaining engineering interpretability through explainable artificial intelligence. In addition, the workflow was implemented within a graphical user interface (GUI) to facilitate practical application by both engineering and data science users.