基于层次专家委员会机器模型的致密储层裂缝开度预测方法

裂缝开度是表征致密储层品质及评价油气产能的关键参数。受沉积、成岩和构造作用的影响,致密储层具有较强的非均质性,导致测井响应特征复杂、无规律,利用常规测井解释方法或者单一机器学习模型很难准确预测储层裂缝开度。为了解决这一问题,提出一种基于层次专家委员会机器模型的致密储层裂缝开度预测方法。首先,从岩心和成像测井资料中获取储层裂缝开度参数,选取相同深度敏感的测井数据作为特征变量构建样本集;然后,采用核岭回归、支持向量回归、BP神经网络作为基础专家网络单元训练、学习样本集;再借助递阶层次结构模型和门神经网络模型构建层次网络模块,自适应生成各个基础专家网络单元的初始权重;最后,综合考虑各个基础专家网络单元的预测性能,利用条件交替期望变换确定各个基础专家网络在最终输出中的贡献,准确预测储层的裂缝开度。实际资料应用表明,该方法能有效地定量表征井中储层裂缝开度,可为致密储层评价提供可靠的地球物理技术支撑。

A prediction method of fracture aperture based on hierarchical expert committee machine model for tight reservoirs

Fracture aperture is a key parameter for cha-racterizing the quality of tight reservoirs and evalua-ting oil and gas productivity. Tight reservoirs have strong heterogeneity due to the influences of sedimentation, diagenesis, and tectonism, resulting in complex and irregular logging response characteristics. Consequently, it is difficult to accurately predict reservoir fracture aperture by the conventional logging interpretation method or with a single machine learning model. To solve this problem, this paper proposes a prediction method of fracture aperture based on the hierarchical expert committee machine model for tight reservoirs. Firstly, the parameters of reservoir fracture aperture are obtained from core and imaging logging data, and sensitive logging data at same depth are selected as characteristic variables to construct a sample set. Secondly, the kernel ridge regression, support vector regression, and BP neural network are used as the basic expert network units to train and learn the sample set. Thirdly, the initial weight of each basic expert network unit is adaptively generated by hierarchical network modules built with the hierarchical structure model and the gated neural network model. Lastly, the prediction performance of each basic expert network unit is comprehensively considered. The contribution of each basic expert network to the final output is determined by alternating conditional expectation transform, and the fracture aperture of the reservoir is thereby accurately predicted. The practical application shows that this method can effectively quantitatively characterize the reservoir fracture aperture in a well and provide reliable geophysical information for the evaluation of tight reservoirs.