页岩气优质储层预测与评价

Shale Gas Quality Reservoir Prediction and Evaluation

摘要: 现今页岩气勘探深度已经超过4000 m，标志着非常规油气勘探越来越成为焦点。但对LM地区勘探实践表明，不同构造位置不同埋深的水平井产能差异较大，并与裂缝发育程度、页岩含气性及岩石物理参数切相关。因此通过对原始测井纵波速度、横波速度和密度等参数进行预处理，基于常规测井资料对孔隙度、脆性指数、TOC等属性计算，并采用基于深度学习的组合预测模型对LM勘探区不同埋深带的优质页岩储层进行评价。研究结果表明：1) 低频和高频信息约束下，结合地震数据空间分布特征的混合深度学习网络结构，能实现高分辨率波阻抗反演，为页岩孔隙度预测打下基础；2) 在体曲率属性、蚂蚁追踪属性及相干属性提取的基础上，结合竞争神经网络的优势进行属性融合对减少构造解释的多解性具有重要作用；3) 建立了可靠的深度学习组合预测模型，完成研究区地质“甜点”精细化预测，克服了单一指标进行预测的多解性，并且本文将预测结果与实际测井解释的分类结果进行对比，匹配度较高，证明了该方法的准确性和有效性。

Abstract: Nowadays, the depth of shale gas exploration has exceeded 4000m, marking that unconventional oil and gas exploration is becoming more and more focal. However, the exploration practice in LM area shows that the production capacity of horizontal wells with different burial depths in different tectonic locations varies greatly and is related to the degree of fracture development, gas content of shale and petrophysical parameters. Therefore, by pre-processing parameters such as longitudinal wave velocity, transverse wave velocity and density of original logs, calculating porosity, brittleness index, TOC and other attributes based on conventional logging data, and using a combined prediction model based on depth learning to evaluate high-quality shale reservoirs in different burial depth zones of LM exploration area. The results show that: 1) a hybrid deep learning network structure combining both low-frequency and high-frequency information constraints with the spatial distribution characteristics of seismic data can achieve high-resolution wave impedance inversion and lay the foundation for shale porosity prediction; 2) on the basis of body curvature attributes, ant-tracking attributes and 3D ground stress prediction, combining the advantages of competitive neural networks for attribute fusion is important to reduce the multi-solution of tectonic interpretation; 3) a reliable deep learning combination prediction model is established to complete the geological “sweet spot” refinement prediction in the study area, which overcomes the multi-solution nature of single-indicator prediction, and the prediction results are compared with the classification results of actual logging interpretation in this paper, and the matching degree is high, which proves the accuracy and effectiveness of this method.