A coalbed methane content prediction method based on an improved Stacking ensemble model

The traditional desorption method for determining coalbed methane (CBM) gas content suffers from high costs, long cycles, and limited regional coverage, making it unsuitable for rapid economic recoverability assessments, cost reduction, and efficiency improvement. To address this, a CBM gas content prediction method based on an improved Stacking ensemble algorithm was proposed. This method optimizes the hyperparameters of five heterogeneous base models—support vector machine (SVM), random forest (RF), XGBoost, LightGBM, and multi-layer perceptron (MLP)—through grid search. It constructed a new meta-training set by innovatively weighting and averaging the raw training-set features and prediction results from each base model, which was then input into a linear regression meta-model for secondary learning. This achieved dual integration of raw features and base-model prediction information. The model was applied to CBM blocks in the southern Qinshui Basin. The results demonstrated that the improved Stacking ensemble model achieved a coefficient of determination of 0.94 on the test set, representing improvements of 10.64% and 6.38% over the optimal single base model and standard Stacking model, respectively. Compared with the standard Stacking model, the mean absolute error, root mean square error, and mean squared error are reduced by 14.3%, 25.2%, and 43.7%, respectively. SHAP (SHapley Additive exPlanations, a method for interpreting machine learning model predictions) interpretability analysis indicated that coal seam slant depth, spontaneous potential, and natural gamma ray are the primary features influencing gas content prediction. Field application demonstrated good agreement between predicted and measured values for eight wells, with a mean absolute error below 0.84 m³/t and a mean relative error controlled within 6%. This model significantly enhances the accuracy and reliability of CBM gas content prediction, providing a new technical approach for CBM resource area selection and efficient development.