Abstract: In petroleum exploration, CO₂ flooding demonstrates significant advantages in enhancing oil recovery; however, issues such as severe gas channeling and limited sweep efficiency remain prominent. To address these challenges, a novel multi-component gas EOR technology utilizing flue gas components-primarily CH₄, N₂, and CO₂ is proposed. For multi-component gas flooding in low-permeability reservoirs, accurate prediction of gas breakthrough time is crucial for optimizing production strategies. In this study, a low-permeability reservoir in Xinjiang is selected as the research subject, and a representative reservoir model is constructed using CMG numerical simulation software. Key geological and fluid parameters, including permeability, heterogeneity, crude oil viscosity, and multi-component gas composition, are incorporated into the model to ensure a realistic simulation of reservoir conditions. By analyzing the dynamic evolution of oil-gas components during multi-component gas flooding, and changes in saturation profiles, a methodology for identifying multi-component displacement fronts is established. Six types of displacement fronts are defined—compositional sweep front, CH₄-rich front, effective compositional front, N₂-rich front, CO₂-rich front, and gas slug front—and their respective characteristics are systematically described. Furthermore, the breakthrough times of these different fronts are determined. Subsequently, grey relational analysis was employed to investigate the influence of multiple factors on breakthrough time, with the permeability, reservoir heterogeneity, and multi-component gas composition identified as dominant factors. Finally, a comparative analysis of linear regression, random forest, and neural network algorithms was conducted. After implementing hyperparameter optimization with feature selection, ensemble learning, and learning-curve-guided hyperparameter tuning with cross-validation, the ensemble-optimized Random Forest model demonstrated superior predictive performance. The optimized model achieved a maximum R² of 0.837 7 and a minimum RMSE of 0.046 4, enabling intelligent prediction of multi-component gas front breakthrough time in multi-component gas flooding processes.

Key words: multi-component gas flooding, gas breakthrough time prediction, low-permeability reservoir, numerical simulation, optimized random forest model

摘要： 在石油开采领域,注CO₂在提高采收率方面具有显著优势,但存在气体窜流严重、波及效率有限等问题,为此基于注烟道气提高采收率理论提出了以CH₄、N₂、CO₂为主的多元气体提高采收率新技术。在低渗透油藏多元气驱技术中,准确预测不同组分气体突破时间对优化开采策略意义重大。以新疆地区低渗透油藏为研究对象,构建了典型油藏模型,考虑渗透率、非均质性、原油黏度、多元气组成等地质、流体和工艺参数,通过分析多元气驱过程中油、气组分动态及饱和度剖面变化,建立了多组分前缘识别方法,确定了组分波及前缘、富CH₄前缘、有效组分前缘、富N₂前缘、富CO₂前缘和气段塞前缘6种前缘类型及其特征,并得到了不同前缘突破时间。其次,运用灰色关联度分析,研究了多因素对突破时间的影响,明确了渗透率、非均质性、多元气组成等主控因素。最后,对比线性回归、随机森林、神经网络等算法,经超参数优化+特征选择方法、集成学习、超参数优化+交叉验证和学习曲线+超参数优化随机森林模型后,基于集成学习优化的随机森林模型预测效果最佳,判定系数最大为0.837 7,均方根误差最小为0.046 4,实现了多元气驱多组分前缘突破时间的智能预测。

关键词: 多元气驱, 气体突破时间预测, 低渗透油藏, 数值模拟, 随机森林优化模型