论文详情
非常规油气储层矿物成分复杂,流体赋存形式多样,孔隙度、渗透率和饱和度等储层参数与测井响应呈非线性关系,难以构建理论模型或经验公式。近年发展迅速的机器学习为测井解释提供了新思路。综述了机器学习在测井解释中的应用现状,概述了集成机器学习的概念、框架和工作机制,指出机器学习包括同质集成与异质集成(委员会机器)两种,简述了两者的差异。重点介绍了委员会机器的基本原理和工作机制。委员会机器是一种结合神经网络和决策树等多种智能算法构建的集成机器学习系统,采用特定的组合策略实现多专家共同决策,对改善训练模型和预测结果具有显著优势。在测井解释中,针对分类、回归问题分别发展了分类委员会机器和回归委员会机器。测井流体识别和储层参数预测的应用表明,委员会机器预测结果比单个智能算法具有更好的精度和鲁棒性,尤其适用于测井解释中的小样本问题。针对有机页岩生烃能力测井评价问题,引入门网络预学习技术构建动态委员会机器,实现了总有机碳含量的智能预测,其预测精度高于传统委员会机器。此外,为了进一步提升预测结果的准确性和可靠性,在委员会机器训练中又引入了地球物理模型约束项,提出了物理模型与委员会机器联合驱动的思路,使致密砂岩孔隙度、渗透率和饱和度的预测精度得到进一步提高。可以看出,基于多元测井数据及其它多源的录井、测试和岩石物理实验数据,利用委员会机器学习算法可以有效地实现储层特征定性判别和定量评价,是测井解释发展的必然趋势。
Unconventional oil and gas reservoirs feature complex mineral compositions,fluids occurring in various forms,and reservoir parameters such as porosity,permeability,and saturation having a nonlinear correlation with the logging response.Therefore,constructing theoretical models or empirical formulas can be challenging.In recent years,the rapid development of machine learning has provided new ideas for logging interpretation.In this paper,the progress in machine learning applications to logging interpretation was discussed.Integrated machine learning methods can be classified as homogenous and heterogeneous.In the latter,a committee machine is constructed using multiple intelligent algorithms such as neural networks and decision trees,and a specific combination strategy is used to achieve joint decision-making.This offers notable advantages in improving training models and prediction results.For logging interpretation,two types of committee machines exist,namely,the classification committee machine and the regress committee machine.Results of the application of the committee machine in fluid identification and reservoir parameter prediction demonstrated that it produced more accurate and robust results than the individual intelligent algorithms.For the logging-based evaluation of the organic shale hydrocarbon generation capacity,the gate network pre-learning technology was introduced to construct a dynamic committee machine and achieve an intelligent prediction of the total organic carbon content.The accuracy of the prediction results was higher than that achieved by any single intelligent algorithm or traditional committee machine.To further improve the accuracy and reliability of the intelligent prediction results,geophysical model constraints were introduced in the training.This allowed to further improve the prediction results of porosity,permeability,and saturation in tight sandstone reservoirs.Through integrated machine learning algorithms for multi-source logging data and other testing,and petrophysical experimental data,a qualitative identification and quantitative evaluation of reservoir characteristics can be achieved.We are convinced that machine learning will be a protagonist in the development of logging interpretation methods.
国家自然科学基金项目(42174149,41774144)、国家科技重大专项(2016ZX05050)及中国石油塔里木油田分公司委托课题(041018120073)共同资助。