利用先验信息约束的深度学习方法定量预测致密砂岩“甜点”

鄂尔多斯盆地东缘LX区块二叠系石盒子组发育河流相致密砂岩储层，高产气层具有孔隙度大于12%、渗透率大于1 mD、含气饱和度大于50%等特征，寻找高产“甜点”亟需开展储层参数定量评价。基于传统地震反演间接预测孔隙度等参数的方法精度低。LX区块地震数据与测井曲线对应关系不一致，存在大量矛盾样本，导致常规卷积神经网络难以应用。为此，在常规卷积神经网络模型的基础上增加一个全连接网络结构。地震数据与测井数据之间通过托布里兹局部网络结构连接，用于解决储层参数与地震数据不直接相关问题。全连接网络结构通过引入线(道)号、层位、地震相等先验信息，可以解决矛盾样本问题。通过引入地层格架、地震相等先验约束信息，构建了适用于致密储层的深度学习网络模型，形成了地质导向的样本井优选方法，从而定量预测储层参数、刻画了高产气层“甜点”平面展布。实际应用结果表明，孔隙度、渗透率、含气饱和度预测结果与井数据吻合度高，新部署的5口钻井测试无阻流量均超过1万m³/d,有力推动了致密气高效开发。

Quantitative prediction of tight sandstone sweet spots based on deep learning method with prior information constraints

The Permian Shihezi Formation is located at the LX block at the eastern margin of the Ordos Basin, and it develops tight sandstone reservoirs with fluvial facies. Reservoirs with high gas production feature a porosity of larger than 12%, a permeability of higher than 1 mD, and a gas saturation of more than 50%, and the quantitative evaluation of reservoir parameters shall be urgently carried out to find sweet spots with high production. However, the accuracy of indirectly predicting porosity and other parameters by traditional seismic inversion is low. In addition, the seismic data and well-logging curves of the LX block have inconsistent corresponding relations, and a lot of conflict samples exist, which makes conventional convolutional neural networks difficult to be applied. Therefore, a fully connected network architecture is added to the conventional convolutional neural network, and the seismic data and well-logging data are connected through local Toeplitz network architecture, so as to deal with the indirect correlation between reservoir parameters and seismic data. The fully connected network architecture can address the conflict samples by introducing prior information including the line/channel number, horizon, and seismic facies. Furthermore, a deep learning network model suitable for tight reservoirs is established by introducing prior constraint information such as stratigraphic framework and seismic facies, and a geo-oriented method for selecting the best sample well is developed, so as to quantitatively predict reservoir parameters and describe the plane distribution of the sweet spots in reservoirs with high gas production. The actual application results show that the predicted results of porosity, permeability, and gas saturation are in good agreement with the well-logging data, and the newly deployed five wells are tested and achieve an open-flow capacity of more than 10,000 m³/d during drilling, which effectively promotes the efficient development of tight gas.