论文详情
致密储层具有地层薄、孔隙度低、横向非均质性强等特点。现有储层预测技术在解决此类问题时,主要依靠人工从反演属性体中寻找可能的甜点区域。由于地层的砂岩含量、孔隙度值与地震反射特征并无直接关系,导致甜点识别准确率低。为此,根据测井数据和地震数据的空间分布特征和数据分布特征,将全局和局部连接网络相结合,有针对性地创建了适用于致密储层甜点预测的混合深度学习网络结构,其中局部连接网络负责学习数据分布特征,全局连接网络负责学习空间分布特征。在甜点预测时,先预测砂岩储层,在此基础上预测孔隙度值。为解决孔隙度数据分布不均匀、有效值与背景值比例不均衡的问题,以砂岩含量曲线为约束条件设置阈值,筛选高于阈值的对应层段的孔隙度值,建立了砂岩含量遮挡的孔隙度训练样本集构建方法。鄂尔多斯盆地东北部的致密砂岩甜点识别结果表明,孔隙度预测结果准确度高,能有效识别本区的致密储层甜点发育区。
Tight reservoirs occur in the form of thin beds with low porosity and strong lateral heterogeneity.Usually,possible sweet spots are manually determined using several inverted seismic attributes.As there is no direct causal relationship between the sand volume,porosity,and seismic data,the recognition accuracy of sweet spots is usually low.Based on the spatial and data distribution characteristics of well logging and seismic data,a hybrid deep-learning network suitable for tight reservoir prediction was established herein.This network is a combination of global and local connected networks,the local connected network is used for learning data distribution characteristics,and the global connected one is used for learning spatial distribution characteristics.To perform sweet spot prediction,a sandstone reservoir was identified,and its porosity was evaluated.To avoid an uneven distribution of porosity data as well as unbalanced ration between the effective and background value,a training-sample selection strategy using a sand-volume shadow was introduced.In this strategy,a threshold value was set according to the sand-volume log,and the porosity values of the corresponding layers higher than this threshold were selected.The method,applied to the northeastern part of the Ordos Basin,achieved a high accuracy in terms of porosity prediction and sweet spot identification.
中海石油(中国)有限公司科技项目(YXKY-2019-ZY-04)资助。