Calorific value is an essential fuel parameter during the process of energy utilization. Forty-four coal samples with different calorific values were quantitatively analyzed by laser-induced breakdown spectroscopy (LIBS) in this paper. The influences of different spectral preprocessing methods such as smoothing, standard normal variate transformation (SNV), multiplicative scatter correction (MSC), mean centering (MC), and derivation by convolution (Savitzky-Golay) on the quantitative model have been analyzed and compared. The results showed that the quantitative model obtained the best comprehensive performance with the application of the 11 points smoothing combined with the second-order derivation. The correlation coefficient of calibration set, correlation coefficient of validation set, RMSECV, and RMSEP were 0.9909, 0.9972, 0.467, and 0.276 MJ/kg, respectively, which were optimized compared with the traditional partial least-squares (PLS) model by 15.9%, 37.7%, 75.1%, and 88.3%, respectively. It indicated the stability and prediction accuracy of the model were greatly improved. This research showed that the spectral smoothing reduced the differences of matrix effects among different samples effectively, and the derivative convolution (Savitzky-Golay) could further eliminate the interelement influences.