A Novel Approach for Production Allocation in Multi-Layer Oil Reservoirs Based on Machine Learning Combining Game Theory

Accurate oil production allocation in multi-layered oil reservoirs is crucial for residual oil characterization and enhanced oil recovery (EOR). Traditional methods are often time-consuming and fail to integrate dynamic and static features effectively, leading to low precision. This study introduces an innovative approach that integrates Sequential Model-Based Optimization (SMBO), Ensemble Learning (EnL) models, and Cooperative Game Theory (CGT) to address these challenges. The SMBO algorithm was effectively applied to determine the optimal hyperparameters for EnL models. Moreover, a strategy was proposed to effectively fuse dynamic production data with static reservoir features, providing a comprehensive data foundation. In single-well applications, our proposed model demonstrated high predictive accuracy (adjusted R2) above 0.95, which establishes a robust basis for allocation. Compared with the traditional KH method, the innovative allocation approach could improve perdition accuracy significantly and reduce mean absolute error () by 83% at least which can support allocation in different layers timely. In block application, the proposed approach showed promising performance, with over 85% of wells achieving adjusted R2 > 0.7. Additionally, the ratio of wells with a prediction error of more than 3.0 m3/d to all wells could be reduced from 74.19% to 9.68% in terms of the proposed model compared with the traditional KH method. It shows the proposed model performs better on production prediction than traditional KH method. In conclusion, the proposed approach provides a reliable and accurate tool for analyzing production dynamics, characterizing remaining oil, and making informed decisions in development adjustments. The systematic methodology and successful field application of our models demonstrate their potential to improve the efficiency of EOR.