Aiming at the prediction problem caused by the strong nonlinearity and long distance dependence of mine pressure time series data, this paper proposed an xLSTM-Infomer prediction method that combined extended long short term memory network ( xLSTM ) and long sequence prediction model ( Informer ). Compared with the traditional single model, this model used the ability of xLSTM to capture local dynamic features finely, and combined the characteristics of Informer 's global long distance dependence and efficient modeling to realize the long-term prediction of the evolution law of mine pressure. In order to verify the prediction ability of the model, this paper took the inclined thick coal seam working face of Liuhuanggou Coal Mine in Xinjiang as the background to predict the mine pressure data of the working face. The experimental results showed that compared with the benchmark models such as LSTM and Informer, the prediction performance of the model established in this paper was higher. The determination coefficient (R2) of the prediction results of different parts was above 93%, and the highest R2 reached 98.21%. Compared with the comparison model, the indexes of MAE and RMSE were also at the lowest level, and the model could also show higher prediction accuracy under complex working conditions. The study provided a reliable technical support for the realization of intelligent mine pressure monitoring and disaster forewarning.