Abstract: The dry deep screening of wet viscous fine-grained raw coal is one of the key technologies to realize the efficient and clean utilization of coal. Cross-type fine-grained roller screen (cross screen) is a new type of dry deep screening equipment, which effectively solves the problems of screen surface plugging and other problems that are easy to occur in traditional dry screening equipment. The mathematical model of the screening process and the DEM (Discrete Element Method) model are difficult to accurately predict the actual screening performance. Based on the machine learning method, the intelligent prediction model of the screening rate of the cross screen is studied. The Spearman correlation coefficient matrix heat map was used to analyze the correlation between the four characteristic variables of feed rate, external water content, sieve surface inclination and sieve shaft speed and the screening rate and the correlation between the characteristics. Based on linear regression (LR), support vector machine (SVM), decision tree (DT) and random forest (RF) algorithms, four intelligent prediction models of cross screening rate were established. Combined with particle swarm optimization (PSO), the hyper-parameter combination optimization of support vector machine, decision tree and random forest models is carried out to obtain the optimal parameter combination of the model and improve the prediction performance and generalization ability of the model. The prediction performance of each model was compared by using three evaluation indexes coefficient of determination (R²), mean square error (E_MS) and mean absolute error (E_MA). Among them, the PSO-SVM prediction model has the best performance and the strongest fitting ability to the data. Its evaluation index R² reaches 0.976 1, and the error between the predicted result and the actual value is the smallest. The corresponding evaluation indexes E_MS and E_MA are 3.110×10⁻⁴ and 1.353×10⁻². The prediction performance of the LR model is the worst, and its evaluation index R² is only 0.722 2, and the error between the predicted result and the actual value is the largest, E_MS and E_MA are 1.320×10⁻³ and 3.137×10⁻². In addition, compared with the LR model, the prediction accuracy of the model obtained by adding L₁ and L₂ regularization is increased by 20.26 % and 4.43 %, respectively. The research results provide a reference for the establishment of the machine learning intelligent prediction model of the screening rate of the cross screen. It provides a new method for analyzing the influence mechanism of the characteristic variables of the cross screen on the screening rate, and provides a theoretical basis for realizing the intelligent control and structural optimization of the cross screen.