Quantitative Evaluation and Machine Learning–Based Diagnostic Warning Method for Multilayer Heterogeneous Reservoir Development: A Case Study from Z Oilfield, Dongpu Depression, China

Abstract Rational evaluation and warning-oriented diagnosis of development performance are essential for efficient management of multilayer heterogeneous reservoirs because they support timely identification of underperforming development units. In this study, a quantitative fuzzy comprehensive evaluation method and a machine learning–based diagnostic warning framework are proposed. First, 10 evaluation indicators integrating geological, development, and economic factors were selected from 68 candidate parameters across eight key aspects. Second, the index weight set was calculated by combining the criteria importance through intercriteria correlation (CRITIC) method, coefficient of variation (COV) method, and analytic hierarchy process (AHP), and the trichotomy method was used to determine the classification boundaries of good, medium, and poor. Based on the established quantitative evaluation method, 52 development units of the Z Oilfield, a mature multilayer heterogeneous fault-block waterflooding oilfield in the Dongpu Depression, China, were quantitatively evaluated. Third, a diagnostic warning model was developed by comparing multiple regression models, including linear regression, ridge regression, random forest, gradient boosting, and back-propagation (BP) neural network. The 52 development units were divided into training and independent testing subsets at a ratio of 70:30, and fivefold cross-validation was performed within the training subset for model assessment. The results show that gradient boosting achieved the best testing performance among the tested models, with a test coefficient of determination ( R 2 ) of 0.737, root-mean square error (RMSE) of 5.940, and mean absolute error (MAE) of 4.884. Finally, the selected model was applied to 10 representative development units, and the results indicated that M82 requires particular attention, and several medium-level units should be continuously monitored. The proposed method provides a practical tool for quantitative evaluation, diagnostic warning, and refined management of multilayer heterogeneous reservoirs.