Shale content prediction of well logs based on CNN-BiGRU-VAE neural network

• Published in: Journal of Earth System Science Vol. 132; no. 3; p. 139
• Main Authors: Zhang, Haoyu, Wu, Wensheng
• Format: Journal Article
• Language: English
• Published: New Delhi Springer India 01-09-2023; Springer Nature B.V
• Subjects: Artificial neural networks; Data logging; Deep learning; Earth and Environmental Science; Earth Sciences; Gamma rays; Geothermal studies; Lithology; Logging; Machine learning; Mathematical models; Mudstone; Neural networks; Physical properties; Prediction models; Root-mean-square errors; Sedimentary rocks; Sedimentation & deposition; Shale; Shales; Space Exploration and Astronautics; Space Sciences (including Extraterrestrial Physics; Well logs; Deep learning; CNN; CNN-BiGRU-VAE; shale content prediction; CNN-BiGRU
• ISSN: 0973-774X; 0253-4126; 0973-774X
• DOI: 10.1007/s12040-023-02164-4

Shale content (
V
sh
) is related to reservoir lithology and physical properties, and accurate shale content prediction models can improve lithology identification efficiency and reduce cost. However, there has been little work on intelligent prediction of shale content. This study focused on mudstone from the geothermal study area of western Gansu Province. CNN-BiGRU-VAE models were selected to predict the shale content using gamma-ray (GR), spontaneous potential (SP), compensated neutron (CNL), resistivity (RESIS), and acoustic time difference (DT) logs. The predicted shale content of the CNN-BiGRU-VAE model was compared with those of traditional Larionov, Clavier and Stieber models. To fairly evaluate these models, the root mean square error (RMSE), mean absolute error (MAE) and mean absolute percentage error (MAPE) were introduced. The results show that Larionov, Clavier and Stieber methods have high RMSE, MAE and MAPE values, indicating these methods poorly predicted shale content. These models significantly underestimate the shale content, resulting in evident deviations between the predicted shale content and their real value. Conversely, CNN-BiGRU-VAE provided accurate shale content prediction, with values of RMSE, MAE and MAPE of 0.032, 0.069 and 0.081. The CNN-BiGRU-VAE model can greatly outperform other models by extracting complex relationships between well-log data and shale content values. These findings demonstrate the effectiveness of the new method in shale content prediction when only logging data are available.