In order to reduce the errors of near-infrared spectral acquisition, analytical models of coal spectra with different particle sizes, 0.2, 1, 3 and 13 mm, were studied in this paper. The feature information of spectra was extracted by PCA method, then two quantitative analytical models were established based on GA-BP and GA-Elman neural network algorithms. Through spectral preprocessing with data normalization and multiplicative scatter correction methods, the results showed that with the 0.2 mm size, the correlations between spectra and the standard value were the strongest, and the analytical precision of models were the best. But for smoothed spectra, the models, under 1 mm size, were better than others. Smoothing method was not suitable for the spectra with less obvious wave crest characteristics, while multiplicative scatter correction method was better. According to original spectra, particle size of 0.2 mm had the highest accuracy, followed by 1 and 3 mm and the worst was under 13 mm. Overall, the larger the size for coal particle, the more the unstable factors for spectra, increasing negative influences on analytical models.