Abstract: The normal and continuous operation of coal cutting cycles is determined by the advancing and alignment of hundreds of hydraulic supports on fully-mechanized coal mining faces, which serves as a critical factor. At present, most automatic follow-up advancing control technologies for hydraulic supports are based on the fixed logic of coal mining processes and mathematical reasoning decision-making models, cannot adapt to complex and variable working face environments, and thus their control performance fails to meet the alignment requirements of hydraulic support clusters for normal production. Consequently, a mode combining automatic control for initial advancing and manual adjustment for re-advancing is still mostly adopted for the on-site follow-up operation of hydraulic supports. In view of this, for the problem of re-advancing regulation and decision-making, artificial intelligence modeling technology is employed, on-site data is used to learn manual operation experience, and a method for learning advancing regulation behaviors and intelligent decision-making of fully-mechanized hydraulic supports is researched and proposed. Firstly, a mathematical characterization method for the advancing distance of hydraulic supports based on stroke data is proposed, a hydraulic support advancing distance matrix and its corresponding mathematical equations set are constructed. Mathematical conversion from stroke data to advancing distance under the two modes of initial follow-up advancing and re-regulation is realized, and the characteristic distribution and spatial trajectory of the advancing distance cycle of hydraulic support clusters are revealed. Subsequently, a modeling method for learning the advancing regulation behaviors of hydraulic supports is proposed, the structure of a hybrid deep learning model integrating CNN with LSTM, Spatial Attention and Transformer (CNN-(LSTM+Spatial Attention+Transformer)) is designed, and a decision-making model for hydraulic support advancing regulation is established through training on nearly ten months of real on-site data samples. For the classification task of determining whether re-advancing regulation is required for hydraulic supports, a CNN model is trained, with the accuracy on the test set reaching 86.05%. For the regression task of predicting re-advancing regulation distance values of hydraulic supports, a hybrid model integrating LSTM, Spatial_attention and Transformer is trained, with the mean absolute error (MAE) on the test set reaching 24.500 8 mm and the accuracy within 50 mm accounting for 85.55%. Finally, industrial application tests of the hydraulic support advancing regulation decision-making model are carried out for 20 process cycles. The test results indicate that the consistency rate between the actual and predicted strategies reaches 88.4%, and the model exhibits favorable regression prediction performance for the re-advancing regulation distance values of hydraulic supports, which proves that the generalization ability of the model basically meets the standards for industrial on-site application. The decision-making model is operated in coordination with existing automatic follow-up advancing control systems, and a new mode integrating initial automatic advancing control and adaptive re-advancing regulation of hydraulic supports is thus formed. In this way, a feasible practical path is provided for the technological development of digital empowerment for unmanned intelligent fully-mechanized coal mining.