Abstract: Fully mechanized caving is the main mining method for extra-thick coal seams in my country. The control accuracy of the caving actuator depends largely on the data feedback of the tail beam posture. In order to improve the control accuracy of the caving actuator, a method for predicting the inclination of the support tail beam based on the long short-term memory neural network (LSTM) was proposed. The absolute coordinates of the support bottom plate, the inclination of the tail beam, the relative height of the tail beam, the frame shifting rate and the column pressure related to the tail beam caving action were used as the input variables of the RNN convolutional network and the LSTM neural network. The historical data of coal caving in a fully mechanized caving working face of a coal mine were used to train and verify the support tail beam posture prediction model, and the support tail beam posture prediction model was established. The tail beam inclination was predicted for 16 consecutive hours. The fitting degree of the predicted tail beam inclination curve and the actual tail beam inclination curve reached 98.7%. In the fully mechanized top coal caving face, 3～4 production shifts of hydraulic support tail beam inclination prediction tests were carried out. After comparing and analyzing the predicted tail beam inclination curve with the actual tail beam inclination curve, when the confidence interval was (0.98, 1.02), the prediction accuracy for 16 hours of continuous production was 98.40%. The LSTM-based support tail beam inclination posture prediction method solved the problem of tail beam inclination control in the electro-hydraulic control system's adaptive coal caving operation, laying the foundation for unmanned coal caving in the fully mechanized top coal caving face.