Intelligent Prediction Method of Deep Coalbed Methane Content Based on Bayesian Optimization Optimized Random Forest

Abstract

Coalbed methane (CBM) content is a critical indicator for CBM resource evaluation and development. Traditional prediction methods rely mainly on core experiments or empirical formulas, which suffer from high costs, long cycles, limited generalization ability, and difficulties in capturing the complex nonlinear relationships between geological/geophysical parameters and CBM content. To address these shortcomings, this paper proposes an intelligent prediction method for deep CBM content based on random forest regression and professional geological feature engineering. The method integrates well-logging data (dual-lateral resistivity LL, long-spaced gamma ray GRLS, natural gamma ray GR, spontaneous potential SP, caliper CAL) and geological parameters (coal seam thickness, bedrock thickness) to construct composite features such as pressure coefficient, resistivity gradient, and organic matter indicator. A robust standardization method based on median and interquartile range (IQR) is adopted to eliminate the interference of outliers. Bayesian optimization is used to adaptively optimize the hyperparameters of the random forest model, avoiding the subjectivity of manual parameter tuning. Experimental results on deep coal seam samples demonstrate that the proposed method achieves excellent fitting performance with a coefficient of determination (R²) exceeding 0.85, showing strong generalization ability, robustness to abnormal data, and engineering practical value. It can provide reliable technical support for the efficient exploration and development of deep CBM resources.