Improved estimation of coalbed methane content using the revised estimate of depth and CatBoost algorithm: A case study from southern Sichuan Basin, China

Abstract

The coalbed methane blocks are always structurally and topographically complex, and there are no models to accurately predict the coalbed methane content in the southern Sichuan Basin, China. This study proposes a feasible machine learning model to achieve a more accurate estimation of coalbed methane content with a small data set. A revised estimate of depth (Z) based on hydrogeological grounds was used as a feature of the prediction model, which can resolve the errors introduced by using the commonly used measured depth as a feature in areas with drastic topographical and structural changes. The revised estimate of depth (Z) and the measured depth were applied for data analysis along with three machine learning algorithms (support vector regression, gradient boosting decision tree, and CatBoost) to obtain the optimized model for coalbed methane content estimation. Modeling results show that the coefficient of determination (R2), mean absolute error, and mean square error of all algorithms were improved after replacing the measured depth with the revised estimate of depth (Z). The Catboost algorithm performed better than the other two algorithms in this study. The prediction results based on CatBoost using the revised estimate of depth (Z) showed the highest accuracy among all the comparison models. The analysis results show that the R2 between Z and coalbed methane content was 150% higher than that between measured depth and coalbed methane content. This vast difference between measured depth and revised estimate of depth (Z) outputs was caused by drastic changes in topography and structure.