Background Coal-bearing basins contain primary and tectonically deformed coals due to multistage tectonic deformations. However, the gas-bearing properties of coal seams differ significantly due to varying pore and fracture densities, permeability, and mechanical properties. This makes coal structure assessment critical to coalbed methane (CBM) exploration and production.
Objective and Method To enhance the accuracy and intelligence of coal structure identification, this study constructed a CNN-BiLSTM-Attention hybrid model that integrated a Bayesian optimization strategy. This model allowed for efficient fusion and representation of multi-scale log data by combining the local feature extraction capability of the convolutional neural network (CNN), the temporal sequence modeling strength of the bidirectional long short-term memory (BiLSTM), and the feature focusing ability of the Attention mechanism. Moreover, this model showed elevated stability and high training efficiency thanks to automatic parameter tuning through Bayesian optimization. Focusing on coal seams in the Shanxi and Benxi formations within the Ordos Basin, this study constructed a dataset of primary, primary-cataclastic, and cataclastic coals based on conventional log data, subjected to normalization, outlier removal, and interpolation of missing values, as well as data from cores. Then, the hybrid model was trained and assessed using the cross-entropy loss function.
Results and Conclusions The CNN-BiLSTM-Attention hybrid model yielded an accuracy of 95.12%, outperforming isolated BiLSTM and CNN models. This hybrid model yielded precision and recall rates above 93% for various coal structures. Furthermore, it yielded a uniform error distribution, as indicated by the confusion matrices. This model was applied to well X2, demonstrating high consistency and discriminative ability for transition zones between varying coal structures. This significantly reduces misclassification between primary-cataclastic and cataclastic coals. Additionally, the hybrid model exhibits strong robust performance in processing noise in log data. This study offers a reliable and effective approach for fine-scale CBM assessment.
DownLoad: