新疆石油天然气 ›› 2026, Vol. 22 ›› Issue (1): 114-124.DOI: 10.12388/j.issn.1673-2677.2026.01.013

• 新能源 • 上一篇    下一篇

机器学习驱动的二氧化碳气水交替驱开发策略优化方法

苏斌,李俊超,朱晨,李纪新   

  1. 西安石油大学,陕西西安  710065

  • 收稿日期:2025-06-13 修回日期:2025-12-20 接受日期:2025-12-29 出版日期:2026-03-10 发布日期:2026-03-10
  • 通讯作者: 李俊超(1987-),2015年毕业于北京大学力学专业,博士,教授,目前从事油气藏数值模拟及相关软件研发和教学工作。(E-mail)lijunchao@xsyu.edu.cn
  • 作者简介:苏斌(1997-),西安石油大学机械工程学院在读硕士,目前从事油气藏数值模拟和地应力模拟研究。(E-mail)subinsjl@163.com
  • 基金资助:
    国家自然科学基金青年基金项目“陆相中低熟页岩油原位转化开发裂缝动态演化机理研究”(52204047);国家自然科学基金项目“井下多化学反应生烃放热协同汽大幅度提高稠油采收率基础问题研究”(U23B6003)。

An Optimization Method for CO2 Water-Alternating-Gas (WAG) Development Strategy Driven by Machine Learning

SU Bin,LI Junchao,ZHU Chen,LI Jixin   

  1. Xi'an Shiyou University,Xi'an 710065,Shaanxi,China

  • Received:2025-06-13 Revised:2025-12-20 Accepted:2025-12-29 Online:2026-03-10 Published:2026-03-10

摘要: 数值模拟方法在CO₂气水交替驱(CO₂-WAG)开发策略优化中虽然具有较高的预测精度,但在多参数组合优化和全局优化中因计算量巨大使其效率受限,难以满足复杂油藏开发方案的高效优化需求。为了解决这一问题,提出了一种基于随机森林模型的数值模拟替代方法,并结合全局优化框架显著提升了开发策略优化的计算效率与预测精度。通过改进的K-Fold交叉验证方法优化模型超参数,构建了能精准捕捉输入参数与开发效果之间复杂非线性关系的替代模型。以随机森林模型的测试集拟合优度(R²)和均方根误差(RMSE)为主要评价指标,验证了模型的预测可靠性,并通过与数值模拟结果对比进一步论证了方法的可行性和精确性。优化中结合循环遍历法探索增油量与CO₂埋存率之间的权衡,提出了多目标下的最优注采策略。优化结果表明,与传统开发方案相比,基于随机森林模型优化后的开发方案能够在显著提高累计产油量的同时增加CO₂埋存量并延迟气窜时间。在某致密油藏的实际模型中,相比样本最优方案,优化方案使累计产油量提升8.6%,CO₂埋存量增加14.5%,气窜时间延迟6年,显著改善了驱替效果和提高了开发效益。

关键词: CO2-WAG, 致密油藏, 随机森林, 主控因素, 机器学习

Abstract: Although numerical simulation exhibits high predictive accuracy in optimizing development strategies for CO₂ water-alternating-gas (CO2-WAG) flooding,its limited efficiency cannot facilitate high-efficiency optimization of development plans for complex reservoirs due to the enormous computational burden involved in multi-parameter combinational and global optimization. To address this issue,a surrogate modeling approach based on the random forest algorithm is proposed and integrated with a global optimization framework to greatly enhance computational efficiency and predictive accuracy. By employing an improved K-Fold cross-validation method for hyperparameter optimization,a surrogate model capable of accurately capturing the complex nonlinear relationships between input parameters and development outcomes is constructed. The reliability of the model is validated using the test set of goodness of fit(R²) and root mean square error (RMSE),and its feasibility and precision are further demonstrated through comparisons with numerical simulation results. During the optimization process,an iterative traversal method is employed to explore the trade-off between incremental oil recovery and CO₂ storage efficiency,leading to the proposal of an optimal injection-production strategy meeting multiple objectives. The optimization results reveal that compared with conventional development schemes,the random forest algorithm optimized strategy significantly increases cumulative oil production,enhances CO₂ storage efficiency,and delays gas breakthrough. For a real tight reservoir model,the optimized scheme delivers an increase of 8.6% in cumulative oil production,an improvement of 14.5% in CO₂ storage,and a six-year delay in gas breakthrough,demonstrating significant enhancements in displacement performance and development benefits.

Key words: CO2-WAG, tight reservoir, random forest, main controlling factor, machine learning

中图分类号: