As a major indicator of coal property, the high accuracy and fast calorific value quantitative analysis plays a significant role in combustion efficiency and economic operation. Laser-induced breakdown spectroscopy (LIBS) was proposed to calibrate the calorific value of 35 coal samples combined with BP neural networks and cluster analysis. The calibration curve established on a certain type of coal samples can not be applied directly to the quantitative analysis for different types of coal considering the influence of matrix effect on the LIBS spectral data. Three different groups were classified with K-means clustering method according to the calorific value, ash content and volatile matter. The training set and the prediction set were optimized. Based on the correlation analysis of analytical line intensity and calorific value, taking into account the physical meaning of the analytical line, the peak intensity of 12 elements was taken as the input parameter of BP neural network model which was established for calorific value analysis of coal samples. The performance of the BP network and the comparison result of 3 different groups samples were studied. The result indicate that the R2 value of calibration curve is 0.996, relative error (RE) and the relative standard deviation (RSD) of calorific value is less than 3.42% and 4.23%, respectively, performing good results of repeated measurement. The calibration curve has different prediction ability for three group coal samples. The influence of experimental parameter fluctuation and matrix effect was reduced in a certain degree using peak intensity as the input parameter. The repeatability and accuracy of quantitative analysis results can be further improved by establishing BP neutral network for various type of coal samples specifically. The analytical results of calorific value based on LIBS technology combined with BP neural network were well predicted, which was potentially proven as a promising technologyy for fast on-line analysis.