Abstract: Low temperature retorting tar yield is a key index to evaluate oil-rich coal. Accurate and efficient prediction of tar yield is of great significance for the evaluation and clean, efficient development and utilization of oil-rich coal resources.. In order to improve the accuracy and efficiency of prediction of tar yield, in view of the multivariate nonlinear complex relationship between coal, rock and coal quality indexes, the characteristic parameters affecting tar yield are divided into three combinations of strong, strong + medium, strong + medium + weak by screening 129 groups of borehole data with complete coal, rock and coal quality indexes in Lianghuai coalfield through correlation analysis. The correlation analysis was conducted to determine the hydrogen-to-carbon ratio, hydrogen content, vitrinite reflectance, volatile matter, and tar yield were most strongly correlated, while the oxides of calcium, magnesium, iron, silicon, aluminum, and fixed carbon were moderately correlated with tar yield. Other indicators had weaker correlations through Pearson correlation coefficient method. The characteristic parameters affect tar yield were divided into three parameter combinations: strong, strong+medium, and strong+medium+weak, and a combination optimization algorithm prediction model based on PSO-GA-BP. Different parameter combinations were trained using machine learning, and the actual application effects of different prediction models were compared and analyzed.The results show that the best fitness of the sample data of the strong + medium feature parameter combination is the largest in the training process, and the absolute coefficient R2, root mean square error RMSE and mean absolute error MAE are better than other feature parameter combinations with better performance and training status. The PSO-GA-BP combined optimization algorithm has the smallest error, and has more advantages in improving the prediction accuracy and data fitting effect of tar yield compared with BP, GA-BP and PSO-BP algorithm. The PSO-GA-BP is applied to the prediction of tar yield of extended sample collected in roadway, and the prediction model shows good generalization ability combined algorithm model established by drilling. The establishment of a complete and comprehensive database of oil-rich coal rock and coal quality is the basis and prerequisite for further improving the model learning ability and forecasting effect through advanced intelligent algorithms.