Multifractal Analysis and Neural Network Prediction of Pore Structures in Coal Reservoirs Based on NMR T 2 Spectra
Yong Sun,Cheng Zhai,Yang Zhao,Jizhao Xu,Yuzhou Cong,Yangfeng Zheng,Wei Tang
DOI: https://doi.org/10.1021/acs.energyfuels.1c01409
2021-07-06
Abstract:Low-field nuclear magnetic resonance (NMR) is widely used for accurate characterization of coal pore structure. The NMR <i>T</i><sub>2</sub> spectrum represents the pore size distribution. Also, the <i>T</i><sub>2</sub> cutoff (<i>T</i><sub>2c</sub>) value is a key parameter, reflecting the free/bound fluid proportion of coal. To characterize the pore structure of coal more comprehensively, the NMR <i>T</i><sub>2</sub> spectra of coals with different pore structures were characterized by multifractal, and then, the <i>T</i><sub>2c</sub> values were predicted by a BP neural network model. The main conclusions are as follows: the <i>T</i><sub>2</sub> 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 <i>T</i><sub>2c</sub>, whereas the proportion of free fluid had a strong negative correlation with <i>T</i><sub>2c</sub>. The quality indices τ(<i>q</i>) 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. <i>D</i><sub>min</sub>, Δ<i>D</i>, α<sub>max</sub>, α<sub>0</sub>, and Δα showed strong positive correlations with <i>T</i><sub>2c</sub>, which indicated that with an increase in <i>T</i><sub>2c</sub>, the proportion of large-size pores decreased and the local pore size distribution became more concentrated and inhomogeneous. The predicted <i>T</i><sub>2c</sub> values of the training, verification, and test sets of the BP neural network model fitted the measured <i>T</i><sub>2c</sub> 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 <i>T</i><sub>2c</sub> prediction of more similar coal samples.This article has not yet been cited by other publications.
energy & fuels,engineering, chemical