Multifractal Analysis and Neural Network Prediction of Pore Structures in Coal Reservoirs Based on NMR T2 Spectra

Low-field nuclear magnetic resonance (NMR) is widely used for accurate characterization of coal pore structure. The NMR T-2 spectrum represents the pore size distribution. Also, the T-2 cutoff (T-2c) value is a key parameter, reflecting the free/bound fluid proportion of coal. To characterize the pore structure of coal more comprehensively, the NMR T-2 spectra of coals with different pore structures were characterized by multifractal, and then, the T-2c values were predicted by a BP neural network model. The main conclusions are as follows: the T-2 spectra of three ranks of coals (anthracite, bituminous coal, and lignite) showed typical unimodal, bimodal, and trimodal distributions, respectively. The porosity had a weak negative correlation with T-2c, whereas the proportion of free fluid had a strong negative correlation with T-2c. The quality indices tau(q) of the three coals changed monotonously, which conformed to the multifractal characteristics. The generalized fractal dimension spectra decreased in an inverse S-shape, decreasing while the multifractal singular spectra were hook-shaped. D-min, Delta D, alpha(max), alpha(0), and Delta a showed strong positive correlations with T-2c, which indicated that with an increase in T-2c, the proportion of large-size pores decreased and the local pore size distribution became more concentrated and inhomogeneous. The predicted T-2c values of the training, verification, and test sets of the BP neural network model fitted the measured T-2c values well, and the mean square error was only 0.17%. The trained BP neural network model was reliable and can be used for the T-2c prediction of more similar coal samples.

Automated summary

Low-field nuclear magnetic resonance (NMR) is widely used to characterize the pore structure of coal. The T-2 spectra of different coal samples show typical distributions, and the pore structure has correlations with the T-2c values. Multifractal analysis and a BP neural network model can predict the T-2c values accurately. The trained BP neural network model is reliable for predicting T-2c values of similar coal samples.