A multiple-input deep residual convolutional neural network for reservoir permeability prediction

被引:17
作者
Masroor, Milad [1 ]
Niri, Mohammad Emami [1 ]
Sharifinasab, Mohammad Hassan [1 ]
机构
[1] Univ Tehran, Inst Petr Engn, Coll Engn, Sch Chem Engn, Tehran, Iran
来源
GEOENERGY SCIENCE AND ENGINEERING | 2023年 / 222卷
关键词
Permeability; Conventional well logs; Deep learning; Multiple; -input; Convolutional neural network; Graphical feature images; STRINGTOWN OIL-FIELD; WEST-VIRGINIA; MACHINE; OPTIMIZATION;
D O I
10.1016/j.geoen.2023.211420
中图分类号
TE [石油、天然气工业]; TK [能源与动力工程];
学科分类号
0807 ; 0820 ;
摘要
Permeability plays an essential role in reservoir-related studies, including fluid flow characterization, reservoir modeling/simulation, and management. However, operational constraints and high costs limit the wide access to the direct measurements of reservoir permeability. Over the years, various machine learning (ML) techniques have been utilized to predict reservoir permeability from widely-available reservoir-related datasets (such as well logs). Nevertheless, experience in real case studies reveals that there is still room to improve the process of permeability prediction in both well and field scales. In this study, an innovative Deep Learning (DL)-based approach is developed for fast and accurate reservoir permeability prediction from conventional well logs. By integrating a multiple-input module, convolutional network, deeper bottleneck, and residual structures, we developed a novel Multiple-Input deep Residual Convolutional Neural Network (named MIRes CNN) for this purpose. It is simultaneously fed by two different types of datasets, Numerical Well Logs (NWLs) and Graphical Feature Images (GFIs). To construct this MIRes CNN model, we integrated a Single-Input deep Residual one Dimensional CNN (SIRes 1D-CNN) to handle the NWL dataset and a Single-Input deep Residual two Dimensional CNN (SIRes 2D-CNN) to treat with GFIs. The significant advantages of the proposed architecture are its capability to efficiently exploit mixed data, reduce both overfitting problem and computational cost, and add more flexibility to the process compared to the classical single-input deep neural networks. The other novelty of this work lies in its generation of an image from geophysical well logs, which offers some physical insight. We compared the proposed MIRes CNN model with two Single-Input deep Residual CNN methods (i.e., SIRes 1D-CNN and SIRes 2D-CNN) and two baseline methods (i.e., Random Forest (RF) and Group Method of Data Handling (GMDH)). The statistical results concluded that the proposed multiple-input method outperforms singlemodality-based techniques and the baseline methods for all employed performance indicators. The study also demonstrates the importance of using physically meaningful log-derived images for CNN model training and the crucial role played by the modified architectures (e.g., residual and deeper bottleneck architectures) in enhancing the accuracy and performance of the model.
引用
收藏
页数:21
相关论文
共 50 条
  • [21] Inter-patient arrhythmia classification with improved deep residual convolutional neural network
    Li, Yuanlu
    Qian, Renfei
    Li, Kun
    [J]. COMPUTER METHODS AND PROGRAMS IN BIOMEDICINE, 2022, 214
  • [22] Neural Network Model for Permeability Prediction from Reservoir Well Logs
    Abdel Azim, Reda
    Aljehani, Abdulrahman
    [J]. PROCESSES, 2022, 10 (12)
  • [23] Research on dynamic prediction of tool life based on deep convolutional neural network
    Guo H.
    Ren B.-C.
    Yan X.-G.
    Tian Q.
    Ren D.-Y.
    [J]. Kongzhi yu Juece/Control and Decision, 2022, 37 (08): : 2119 - 2126
  • [24] A deep convolutional neural network model for rapid prediction of fluvial flood inundation
    Kabir, Syed
    Patidar, Sandhya
    Xia, Xilin
    Liang, Qiuhua
    Neal, Jeffrey
    Pender, Gareth
    [J]. JOURNAL OF HYDROLOGY, 2020, 590 (590)
  • [25] A deep learning method for prediction of cardiovascular disease using convolutional neural network
    Sajja T.K.
    Kalluri H.K.
    [J]. Revue d'Intelligence Artificielle, 2020, 34 (05) : 601 - 606
  • [26] Residual Deep Convolutional Neural Network Predicts MGMT Methylation Status
    Panagiotis Korfiatis
    Timothy L. Kline
    Daniel H. Lachance
    Ian F. Parney
    Jan C. Buckner
    Bradley J. Erickson
    [J]. Journal of Digital Imaging, 2017, 30 : 622 - 628
  • [27] Deep Residual Separable Convolutional Neural Network for lung tumor segmentation
    Dutande, Prasad
    Baid, Ujjwal
    Talbar, Sanjay
    [J]. COMPUTERS IN BIOLOGY AND MEDICINE, 2022, 141
  • [28] Residual Deep Convolutional Neural Network Predicts MGMT Methylation Status
    Korfiatis, Panagiotis
    Kline, Timothy L.
    Lachance, Daniel H.
    Parney, Ian F.
    Buckner, Jan C.
    Erickson, Bradley J.
    [J]. JOURNAL OF DIGITAL IMAGING, 2017, 30 (05) : 622 - 628
  • [29] Deep Convolutional Neural Network for Accurate Prediction of Seismic Events
    Turarbek, Assem
    Bektemesov, Maktagali
    Ongarbayeva, Aliya
    Orazbayeva, Assel
    Koishybekova, Aizhan
    Adetbekov, Yeldos
    [J]. INTERNATIONAL JOURNAL OF ADVANCED COMPUTER SCIENCE AND APPLICATIONS, 2023, 14 (10) : 604 - 613
  • [30] Prediction of Wave Power Generation Using a Convolutional Neural Network with Multiple Inputs
    Ni, Chenhua
    Ma, Xiandong
    [J]. ENERGIES, 2018, 11 (08)