Summary Accurate assessment of reservoir health is essential for efficient and sustainable oilfield development. In this study, we develop a data-driven approach to evaluate and classify health status in conventional sandstone reservoirs using machine learning and interpretable analysis. The categorical boosting (CatBoost) ensemble model predicts cumulative oil production based on 15 key development indicators, trained and validated on a data set comprising 42 real blocks and 158 simulated cases from a mature oil field. K-fold cross-validation yields high predictive accuracy coefficient of determination (R2) = 0.964; root mean squared error (RMSE) = 0.3096. Shapely additive explanations (SHAP) analysis quantifies feature importance, identifying remaining recoverable reserves per well as the dominant positive driver and water-cut increase rate as the least influential. An enhanced radar chart, integrated with a quantified risk coefficient, effectively visualizes reservoir health status and categorizes warning levels as green (safe), yellow (light warning), and red (heavy warning) in accordance with established industry standards for conventional sandstone reservoirs. The framework effectively detects anomalous development trends, enabling proactive adjustment of injection/production strategies. This interpretable data-driven framework supports rapid decision-making and near-real-time health assessment based on periodically updated field data.
Wang et al. (Sun,) studied this question.