Abstract: High-intensity mining and extended working faces have significantly modified strata pressure manifestation patterns and triggered spatiotemporal evolution of mining-induced pressure. Achieving intelligent prediction of roof pressure is of great significance for ensuring the safe production of mines. Based on the evolution trend and classification prediction of mine pressure in the extended working face of the medium-thick coal seam under high-intensity mining in Yuandatan Coal Mine, Yuheng Mining Area, Shaanxi Province, the distribution characteristics of the support resistance in the extended working face and the strata pressure behavior law were analyzed. The mine pressure data of the working face was dynamically mapped to the spatial grid cells with topological relationships. The spatio-temporal correlation characteristics of the support in the working face were extracted by using the unsupervised clustering algorithm, forming a spatio-temporal linkage support resistance analysis method. A Transformer (Patch Time Series Transformer, PatchTST) mine pressure prediction model based on the patch mechanism was constructed. Multiple prediction models were horizontally compared and tested based on the field measured data to verify the accuracy of PatchTST and its applicability to the prediction of long mine pressure sequences. Finally, the engineering application performance test and prediction error analysis were carried out. The results show that the pressure distribution in the strike direction of the 11207 extended working face of Yuandatan Coal Mine presents an “M”-shaped feature of “double peaks-valley oscillation”. With the advancement and time progression, the “M”-shaped pressure field generally shows a periodic evolution law of “formation-stabilization-recurrence”. After the mine pressure data is dynamically mapped and clustered in the spatial grid units, the accumulation area of the working face pressure can be accurately identified, and the automatic solution of the pressure intensity classification can be achieved. The PatchTST model has the best prediction accuracy when the look-back window is 240 and the prediction step is 3, with the evaluation indicators M_SE and M_AE being 0.095 and 0.240, respectively. Compared with multiple models based on the attention mechanism, the PatchTST model can achieve the lowest prediction error. The engineering application performance test shows that the method accurately identifies the relatively strong pressure observed on site, and the error analysis also indicates that the prediction accuracy of the model is high, with an accuracy rate of up to 92.8%. The research can provide reference and guidance for the strata pressure behavior law in extended working faces and the intelligent prediction and early warning of working face pressure.