Scholarlink

<p>Subsurface geological formations not only contain hydrocarbon reserves but also have significant potential for CO<sub>2</sub> sequestration. This study investigated the predictive capabilities of a novel 1D convolutional neural network (CNN1D) model integrated with multi-head attention mechanisms, which are essential for effective CO<sub>2</sub> sequestration strategies. The model uses continuous depth-based well log measurements of various subsurface properties as input features, including the gamma ray, spontaneous potential, acoustic log, thorium, potassium, caliper, and compensated neutron log. The proposed model effectively captures complex relationships within the data, yielding significant improvements in lithology classification accuracy. Moreover, the particle swarm optimization algorithm plays a crucial role in hyperparameter tuning, dynamically adapting model parameters across iterations to enhance predictive performance. The results showed significant improvements in lithology prediction accuracy, with gray–black silty mudstone achieving the highest classification accuracy of 91% compared with other lithofacies. The model outperforms traditional manual lithology identification methods by offering greater consistency, faster processing, and improved handling of complex, high-dimensional datasets. Evaluation metrics such as the precision, recall, and F1-score highlight the model's effectiveness, with an overall accuracy of 66%, supporting its potential as a reliable tool for assessing CO<sub>2</sub> storage potential and informing decision-making in sustainable energy projects.</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

Improved Reservoir Lithology Estimation on Hybrid PSO-Transformer-CNN1D Modeling for CO2 Sequestration Using Well Log Data

  • 期刊论文

  • 发表日期：

    2026年4月9日

查看全文
  • Elieneza Nicodemus Abelly,
  • Feng Yang,
  • Allou Koffi Franck Kouassi,
  • He Zheng,
  • Erasto E. Kasala,
  • Grant Charles Mwakipunda,
  • Wakeel Hussain,
  • Christopher Nyangi Mkono

摘要

Subsurface geological formations not only contain hydrocarbon reserves but also have significant potential for CO2 sequestration. This study investigated the predictive capabilities of a novel 1D convolutional neural network (CNN1D) model integrated with multi-head attention mechanisms, which are essential for effective CO2 sequestration strategies. The model uses continuous depth-based well log measurements of various subsurface properties as input features, including the gamma ray, spontaneous potential, acoustic log, thorium, potassium, caliper, and compensated neutron log. The proposed model effectively captures complex relationships within the data, yielding significant improvements in lithology classification accuracy. Moreover, the particle swarm optimization algorithm plays a crucial role in hyperparameter tuning, dynamically adapting model parameters across iterations to enhance predictive performance. The results showed significant improvements in lithology prediction accuracy, with gray–black silty mudstone achieving the highest classification accuracy of 91% compared with other lithofacies. The model outperforms traditional manual lithology identification methods by offering greater consistency, faster processing, and improved handling of complex, high-dimensional datasets. Evaluation metrics such as the precision, recall, and F1-score highlight the model's effectiveness, with an overall accuracy of 66%, supporting its potential as a reliable tool for assessing CO2 storage potential and informing decision-making in sustainable energy projects.