This study aims to quantitatively characterize the pore-structure heterogeneity of deep coal reservoirs using a multifractal approach and to classify reservoir types by combining multifractal parameters with unsupervised learning. Twelve deep coal samples from the Linxing–Shenfu block in the eastern Ordos Basin were selected, and a series of experiments were conducted, including organic matter maturity, mineral composition analysis, coal porosity and permeability measurements, scanning electron microscopy (SEM), nuclear magnetic resonance (NMR), and CO2 adsorption. Based on multifractal theory, the heterogeneity and connectivity of pore structures were characterized, and a partial least-squares regression (PLSR) model was constructed to investigate the influence of mineral composition on multifractal parameters. The singularity spectrum width (Δα), Hurst exponent (H), porosity, and permeability were combined to build a four-dimensional feature matrix, which was then reduced by principal component analysis (PCA) and clustered using the K-means++ algorithm to achieve unsupervised reservoir classification. NMR and CO2 adsorption exhibit complementary sensitivity to pore-size ranges and, when integrated, allow the characterization of full-scale pore-structure heterogeneity. The Δα values derived from NMR are generally higher than those from CO2 adsorption, indicating a stronger apparent heterogeneity in the NMR response, which is interpreted to be related to an overestimation of heterogeneity in the micropore range. The reservoirs in the study area are divided into four types with markedly different flow capacities, in which Class I and Class III reservoirs have average permeabilities of 2.40 mD and 27.55 mD, respectively. The proposed workflow, which couples multiscale experimental data with clustering algorithms, shows good reproducibility and provides a theoretical basis for refined zoning and development of deep coalbed methane reservoirs.
Li et al. (Mon,) studied this question.