基于半监督学习的井震联合储层横向孔隙度预测方法

Seismic and well logs integration for reservoir lateral porosity prediction based on semi-supervised learning

传统地震储层预测技术一般基于弹性参数反演和岩石物理建模的级联流程实现储层孔隙度预测,其预测精度受到波动理论和岩石物理理论的近似假设、初始模型和二次反演累积误差等因素的影响.为缓解这些问题,本文提出了一种基于双向门控递归单元神经网络的半监督学习井震联合孔隙度预测方法,实现从地震数据直接预测储层横向孔隙度.通过少量的地震测井样本标签对和多目标函数约束建立智能化多尺度多信息融合孔隙度预测模型,实现地震数据到孔隙度,孔隙度再到生成地震数据的闭环映射.此外,在网络模型每次迭代更新的过程中随机引入非井旁地震道参与网络训练,非井旁地震道的波形匹配能在一定程度上保证井间孔隙度的预测精度.模型数据和实际数据测试结果表明,本文提出的方法相比于有监督学习孔隙度预测方法能进一步提高储层孔隙度的预测准确性和横向连续性,获得较为可靠的储层物性参数的空间分布.

To realize reservoir porosity prediction, traditional seismic reservoir prediction methods are historically implemented by a cascade workflow coupled with elastic parameter inversion and rock-physic modeling. The accuracy of predicted results is affected by many elements, such as the approximate assumptions of the wave and rock physics theory, initial models, and accumulated errors caused by secondary inversion. To alleviate these problems, a semi-supervised learning seismic and well logs integration porosity estimation method based on Bidirectional Gated Recursive Units(Bi-GRUs) neural networks is proposed in this paper, which can be used to directly predict lateral reservoir porosity from seismic data. An intelligent multi-scale and multi-information fusion porosity prediction model based on multi-objective function constraints and a small number of seismic and well pairs is established, which can realize the closed-loop mapping of seismic data to porosity, and porosity to generated seismic data. In addition, non-well seismic traces are randomly introduced to participate in the network training process in each iteration. The prediction accuracy of inter-well porosity is guaranteed by the seismic waveform fitting degree of non-well seismic traces to some extent. The test results of synthetic data and field data demonstrate that the proposed method can further improve the prediction accuracy and lateral continuity of the reservoir porosity section compared with the supervised learning method, and obtain more reliable spatial distribution of petrophysical property.