地球物理学进展 ›› 2019, Vol. 34 ›› Issue (4): 1548-1555.doi: 10.6038/pg2019CC0528

• 应用地球物理学Ⅰ(油气及金属矿产地球物理勘探) • 上一篇    下一篇

利用卷积神经网络对储层孔隙度的预测研究与应用

杨柳青1,陈伟2,3,*(),查蓓1   

  1. 1. 长江大学地球物理与石油资源学院,武汉 430100
    2. 油气资源与勘探技术教育部重点实验室(长江大学),武汉 430100
    3. 非常规油气湖北省协同创新中心,武汉 430100;
  • 收稿日期:2018-12-03 修回日期:2019-03-05 出版日期:2019-08-20 发布日期:2019-08-30
  • 通讯作者: 陈伟 E-mail:chenwei2014@yangtzeu.edu.cn
  • 作者简介:杨柳青,男,1993年生,硕士在读,主要从事深度学习的地震资料解释研究.(E-mail: yangliuqingqin@163.com)
  • 基金资助:
    基于经验模态分解的自由表面多次波衰减方法研究(41804140)

Prediction and application of reservoir porosity by convolutional neural network

YANG Liu-qing1,CHEN Wei2,3,*(),ZHA Bei1   

  1. 1. College of Geophysics and Petroleum resources, Yangtze University,Wuhan 430100, China
    2. Key Laboratory of Exploration Technology for Oil and Gas Resources of Ministry of Education, Yangtze University, Wuhan 430100, China
    3. Hubei Cooperative Innovation Center of Unconventional Oil and Gas, Wuhan 430100, China
  • Received:2018-12-03 Revised:2019-03-05 Online:2019-08-20 Published:2019-08-30
  • Contact: Wei CHEN E-mail:chenwei2014@yangtzeu.edu.cn

摘要:

在测井资料与岩心数据较少,储层参数在地质综合解释方面又非常重要的情况下,如何提高储层参数的预测精准度显得尤为关键.本文提出使用深度学习方法根据已有的测井数据预测岩心孔隙度,构建出基于Adam优化算法、dropout技术、ReLU激励函数等技术融合的卷积神经网络模型.首先,分析选定测井参数与孔隙度的相关性.然后,通过交叉验证实验优选模型参数,如学习率、网络深度等.最后,运用卷积神经网络建立自然伽马等四种测井参数与孔隙度的非线性映射关系.通过对实际资料的训练测试,结果显示,基于测井资料的卷积神经网络模型相对于深度神经网络等模型而言,总样本预测数值相关性提高了2%左右,模型预测数值偏差下降了1.5%,通过实验验证卷积神经网络模型具有预测优越性与鲁棒性,在储层参数预测方向有良好的应用前景.

关键词: 测井资料, 深度学习, 卷积神经网络, 储层参数

Abstract:

In the case that the logging data and core data are few, and the reservoir parameters are very important in terms of comprehensive geological interpretation, how to improve the prediction accuracy of reservoir parameters is particularly critical. In this paper, a deep learning method is proposed to predict core porosity based on existing logging data, and a convolutional neural network model based on Adam optimization algorithm, dropout technique and ReLU excitation function is constructed. First, the correlation between selected logging parameters and porosity is analyzed. Then, model parameters such as learning rate, network depth, etc. are preferred through cross-validation experiments. Finally, a convolutional neural network is used to establish a nonlinear mapping relationship between four logging parameters such as natural gamma and porosity. Through the training test of the actual data, the results show that the convolutional neural network model based on logging data is more than 2% of the total sample prediction relative to the deep neural network model, and the model prediction numerical deviation is reduced by 1.5%, the experimental verification of the convolutional neural network model has the superiority and robustness of prediction, and has a good application prospect in the prediction direction of reservoir parameters.

Key words: Logging data, Deep learning, Convolutional neural network, Reservoir parameters

中图分类号: