Scholarlink

<p>Severe freshwater mud invasion in the sandy-shaly reservoirs of the A’nan Fault Block in the Erlian Oilfield has led to the pronounced development of low-contrast oil layers within the study area. This phenomenon significantly increases the difficulty of fluid identification and constrains efficient exploration and development. Targeting the A’nan 31 Block, this study integrates well-logging and well-testing data to propose an intelligent fluid identification method based on SMOTE-SABO-BiLSTM, aiming to improve the identification accuracy of low-contrast oil layers in complex sandy-shaly reservoirs. Eight key well-logging curves, including 90-inch array induction resistivity (AT90), acoustic compressional slowness, 30-inch array induction resistivity (AT30), photoelectric factor, microgradient resistivity, compensated neutron log, microelectrode resistivity, and spontaneous potential. They were selected as input features through feature importance analysis. To address the issue of imbalanced reservoir types, the SMOTE (synthetic minority oversampling technique) algorithm was employed to oversample minority-class samples (e.g., water layers, oil–water layers). This approach was combined with a bidirectional long short-term memory network (BiLSTM), leveraging its strength in processing sequential data, and the subtraction-average-based optimizer (SABO) was introduced to automatically optimize model hyperparameters. Experimental results demonstrate that when the SABO population size is 50, the model achieves an optimal hyperparameter combination (learning rate: 0.01, Dropout rate: 0.10, LSTM units: 64/127). This configuration yields a fluid identification accuracy of 92.24% on the test set, significantly outperforming the unoptimized BiLSTM (87.59%), SABO-LSTM (84.83%), and basic LSTM (76.72%) models. The proposed SMOTE-SABO-BiLSTM hybrid model effectively addresses the dual challenges of imbalanced reservoir types and low-resistivity oil layer identification in sandy mudstone reservoirs. By utilizing SMOTE to handle data imbalance and incorporating SABO-optimized BiLSTM, it substantially enhances the identification accuracy for complex sandy mudstone reservoir fluids. This method provides a valuable reference for identifying low-contrast oil layers with similar characteristics.</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

Application of SMOTE-SABO-BiLSTM hybrid model in intelligent identification of low-contrast oil layer in sandy-shaly reservoirs

  • 期刊论文

  • 发表日期：

    2026年1月27日

查看全文
  • Yuxuan Zhang,
  • Diren Liu,
  • Jingdong Yang,
  • Huan Feng,
  • Pang Wu,
  • Yu Gong

摘要

Severe freshwater mud invasion in the sandy-shaly reservoirs of the A’nan Fault Block in the Erlian Oilfield has led to the pronounced development of low-contrast oil layers within the study area. This phenomenon significantly increases the difficulty of fluid identification and constrains efficient exploration and development. Targeting the A’nan 31 Block, this study integrates well-logging and well-testing data to propose an intelligent fluid identification method based on SMOTE-SABO-BiLSTM, aiming to improve the identification accuracy of low-contrast oil layers in complex sandy-shaly reservoirs. Eight key well-logging curves, including 90-inch array induction resistivity (AT90), acoustic compressional slowness, 30-inch array induction resistivity (AT30), photoelectric factor, microgradient resistivity, compensated neutron log, microelectrode resistivity, and spontaneous potential. They were selected as input features through feature importance analysis. To address the issue of imbalanced reservoir types, the SMOTE (synthetic minority oversampling technique) algorithm was employed to oversample minority-class samples (e.g., water layers, oil–water layers). This approach was combined with a bidirectional long short-term memory network (BiLSTM), leveraging its strength in processing sequential data, and the subtraction-average-based optimizer (SABO) was introduced to automatically optimize model hyperparameters. Experimental results demonstrate that when the SABO population size is 50, the model achieves an optimal hyperparameter combination (learning rate: 0.01, Dropout rate: 0.10, LSTM units: 64/127). This configuration yields a fluid identification accuracy of 92.24% on the test set, significantly outperforming the unoptimized BiLSTM (87.59%), SABO-LSTM (84.83%), and basic LSTM (76.72%) models. The proposed SMOTE-SABO-BiLSTM hybrid model effectively addresses the dual challenges of imbalanced reservoir types and low-resistivity oil layer identification in sandy mudstone reservoirs. By utilizing SMOTE to handle data imbalance and incorporating SABO-optimized BiLSTM, it substantially enhances the identification accuracy for complex sandy mudstone reservoir fluids. This method provides a valuable reference for identifying low-contrast oil layers with similar characteristics.