Prediction and Evaluation of Coalbed Gas Content Based on Machine Learning

ABSTRACT Under the “dual carbon” goals, coalbed methane, as a clean energy source, is of significant importance to energy security and carbon emission reduction. The gas content of coal is a key parameter for evaluating coal reservoirs. Traditional experimental methods are limited by core integrity and data constraints, making widespread application difficult. Logging data has the advantages of continuity and low cost, and combined with machine learning, it can effectively characterize the nonlinear relationship between logging responses and gas content. This study focuses on the No. 5 coal seam of the Shanxi Formation in the Yanchang Block on the southeastern edge of the Ordos Basin. Five logging parameters were selected: sonic transit time, natural gamma, deep lateral resistivity, shallow lateral resistivity, and density. Three prediction models were constructed: BP neural network, random forest, and convolutional neural network (CNN). Comparisons show that the CNN model performs best, with a coefficient of determination (R 2 ) of 0.93079 between predicted and measured gas content, and a mean absolute error of 0.36 m 3 /t. Using the CNN model to predict gas content distribution, the results show that the No. 5 coal seam gas content ranges from 2.16 to 25 m 3 /t, with an average of 21.34 m 3 /t, generally exhibiting higher values in the northwest and lower values in the central area. This study provides machine learning support for efficient prediction of coalbed methane gas content and has reference value for coalbed methane exploration and development.