Scholarlink

<p>Accurate lithofacies classification is crucial for reservoir evaluation in tight sandstone reservoirs, where geological heterogeneity and scarce core data pose significant challenges. This study proposes a hybrid deep learning framework that integrates Convolutional Neural Networks (CNN) for spatial feature extraction from well log sequences with Deep Q-Networks (DQN) for reinforcement-based adaptive decision-making. Unlike conventional supervised learning, the DQN component explicitly optimizes for geological consistency through a reward mechanism that considers sequential dependencies along the wellbore. To address data quality issues, Thermal Kernel Interpolation (TKI) is employed for missing value recovery using stratigraphic similarity, while Borderline Synthetic Minority Over-sampling Technique (BSMOTE) mitigates class imbalance by targeting boundary samples. Comparative experiments demonstrate that the proposed CNN-DQN model substantially outperforms traditional machine learning approaches, particularly in identifying transitional lithofacies and maintaining stratigraphic coherence. This workflow provides a robust and interpretable solution for lithofacies characterization in tight sandstone reservoirs with complex geological conditions.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Enhancing Tight Sandstone Reservoir Classification with CNN-DQN and BSMOTE

  • 期刊论文

  • 发表日期：

    2026年5月12日

查看全文
  • Pan Li,
  • Jun Li,
  • Jia-bing Meng

摘要

Accurate lithofacies classification is crucial for reservoir evaluation in tight sandstone reservoirs, where geological heterogeneity and scarce core data pose significant challenges. This study proposes a hybrid deep learning framework that integrates Convolutional Neural Networks (CNN) for spatial feature extraction from well log sequences with Deep Q-Networks (DQN) for reinforcement-based adaptive decision-making. Unlike conventional supervised learning, the DQN component explicitly optimizes for geological consistency through a reward mechanism that considers sequential dependencies along the wellbore. To address data quality issues, Thermal Kernel Interpolation (TKI) is employed for missing value recovery using stratigraphic similarity, while Borderline Synthetic Minority Over-sampling Technique (BSMOTE) mitigates class imbalance by targeting boundary samples. Comparative experiments demonstrate that the proposed CNN-DQN model substantially outperforms traditional machine learning approaches, particularly in identifying transitional lithofacies and maintaining stratigraphic coherence. This workflow provides a robust and interpretable solution for lithofacies characterization in tight sandstone reservoirs with complex geological conditions.