裂缝是超深层致密油气运移的主要渗流通道,对超深层油气勘探开发至关重要。高温高压环境下的岩石物理性质更为复杂,裂缝测井响应弱且多解性强。针对这一难题,本文提出一种基于深度核方法(DKM)的超深层致密砂岩裂缝测井识别方法,该方法通过核主成分分析提取裂缝非线性特征,通过深度学习级联结构深度挖掘用于裂缝识别的不同尺度测井响应特征,通过无梯度优化算法自动确定最优模型结构及参数,避免了深度学习需要进行的超参数调整的问题。以塔里木盆地克深气田下白垩统巴什基奇克组的超深层致密砂岩储层为例,对所提方法进行了实例应用和验证,在测井裂缝响应敏感性分析的基础上,优选了6种测井曲线用于裂缝识别,前3种DEN、RD和RM为实测测井数据,后3种RSD、nT1和nT2是为了获取更多裂缝信息而重构的曲线,并厘清了裂缝段与无裂缝段在测井参数方面的差异。裂缝识别结果与岩心裂缝描述对比表明,深度核方法可以较为准确地识别超深层致密砂岩裂缝,相比常规多核方法,精度可以提升5%以上,在实际单井裂缝识别工作中具有较强的适用性。

Fractures are the main seepage channels for oil and gas migration in ultra-deep tight reservoirs, and are crucial for ultra-deep oil and gas exploration and development. Ultra-deep tight reservoirs have highly complex petrophysical characteristics under high-temperature, high-pressure environments, resulting in ambiguous and multi-solution well log responses pertaining to fractures. To solve this problem, we proposes a deep kernel method (DKM) for fracture identification in ultra-deep tight sandstones. This method employs kernel principal component analysis to extract non-linear log features associated with fractures. It utilizes a deep learning cascade structure to extensively explore the log response characteristics across various scales for accurate fracture identification. Furthermore, it employs gradient-free optimization algorithms to automatically determine the optimal model structure and parameters. We conducted a case study in the ultra-deep tight sandstone reservoirs of the Lower Cretaceous Bashijiqike Formation in the Keshen gas field, Tarim Basin, and the proposed method was applied and verified. Through sensitivity analysis of logging responses to fractures, six specific logging curves were chosen for fracture identification. The first three variables, DEN, RD, and RM, correspond to direct measurements from well logging, whereas the latter three, RSD, nT1, and nT2, are reconstructed curves specifically developed to enhance the detection of fracture-related information. This distinction effectively clarifies the differences in logging parameters between fractured and non-fractured zones. A comparative analysis between the fracture identification results and the core fracture descriptions demonstrated the accuracy of the deep kernel method in identifying fractures within ultra-deep tight sandstone formations. This method achieved an accuracy improvement of over 5% compared to the conventional multi-kernel support vector machine method, thus exhibiting robust applicability for single-well fracture identification.