AI-Driven Re-Fracturing Optimization in Shale Oil Wells: A Data-Centric Approach from the Ordos Basin

Abstract This paper aims to develop a data-driven framework for optimizing re-fracturing decisions in low-yield shale oil wells within the Ordos Basin. The objective is to improve production forecasting and design efficiency through the integration of geological, engineering, and production data using machine learning and multi-objective optimization. The study collected data from 124 horizontal and 249 directional re-fractured wells. After preprocessing with K-nearest neighbors imputation and statistical anomaly detection, a hybrid feature selection strategy was used, combining Pearson and Spearman correlations, mutual information, SHapley Additive exPlanations, Recursive Feature Elimination, and Boruta algorithms. For horizontal wells, a multi-branch neural network combining multilayer perceptron and Transformer architectures was developed to predict 24-month production. Directional wells were modeled using eXtreme Gradient Boosting. A Non-dominated Sorting Genetic Algorithm II was applied to optimize design parameters such as fluid and proppant volume, additive dosage, and injection rate. Model validation involved Tree-structured Parzen Estimator optimization and five-fold cross-validation. Data libraries were successfully constructed for 110 horizontal and 249 directional wells with an imputation accuracy of 87 percent. Fourteen key variables were identified as dominant re-fracturing performance drivers, including proppant volume, fluid volume, porosity, and remaining reserves. The multilayer perceptron–Transformer model achieved a 0.92 coefficient of determination for horizontal wells, and the eXtreme Gradient Boosting model achieved 0.94 for directional wells. In field testing at the Guping Block, the optimized refracturing design improved two-year cumulative production by 2.3 percent, increasing yield from 6,426 to 6,579 tonnes, while maintaining cost control. These results confirm the reliability of the framework in enabling precise production forecasting and economically optimized re-fracturing strategies for low-productivity shale oil wells. This paper introduces a replicable workflow that combines explainable artificial intelligence and multi-objective optimization for refracturing design. It marks the first application of a multilayer perceptron–Transformer model in shale oil refracturing prediction and provides a data-centric decision-support tool tailored to unconventional reservoir redevelopment.