Abstract: To overcome the shortcomings of traditional linear regression model in coal calorific value prediction, such as low prediction accuracy and narrow applicability, the linear correlation between the index of proximate analysis and elemental analysis and the calorific value of coal was comprehensively investigated. the characteristic variables of influencing the calorific value were selected by Mean Impact Value (MIV) method. Combining Particle Swarm Optimization(PSO) method to optimize the traditional BP neural network, a non-linear prediction method of coal calorific value is proposed based on MIV-PSO-BP neural network. The results show that only ash and carbon content show a certain linear correlation with calorific value in the indexes of proximate analysis and elemental analysis of coal, the other indexes show a poor linear correlation with calorific value. The indexes of proximate analysis of coal have great influence on calorific value, while only carbon content and hydrogen content have significant influence on calorific value from elemental analysis, and the effect of other indicators on calorific value is negligible. Compared with other calorific prediction models proposed by other researchers, the average relative error and root mean square error of MIV-PSO-BP neural network model are lower, and the overall prediction effect is better. The correlation coefficient can reach 98.60%.