Coal Seam Roof and Floor Lithology Prediction for Underground Coal Gasification Based on Deep Residual Shrinkage Network

Abstract

Lithology identification is a crucial task in coal underground gasification projects, serving as a prerequisite for ensuring the safe operation of these endeavors. The inherent complexity in the relationship between logging parameters and lithological compositions creates ambiguity, leading to biases in traditional logging interpretation methodologies. We introduce a lithological prediction model, the deep residual shrinkage network (DRSN), which integrates residual networks, attention mechanisms, and soft-thresholding strategies. This network mitigates the gradient vanishing issue common in traditional neural networks and enhances the model's focus on essential features, thereby improving its ability to capture critical information. Acoustic, bulk density, neutron, gamma, and deep resistivity logs are used as inputs, with lithology as the output. A comparative analysis between the DRSN and other newer lithological prediction models is conducted. Blind well testing results demonstrate the superior performance of the DSRN, with higher Accuracy, Precision, Recall, and F1 Scores of 0.8221, 0.7198, 0.8004, and 0.7465, respectively. This study provides a novel and rapid method for lithology evaluation of strata in the early stages of underground coal gasification.