Abstract：The rock pressure prediction of the fully mechanized mining face plays an important role in ensuring the safe and efficient
mining of coal mines. Long-term and short-term memory networks (LSTM) in the field of deep learning have been proven to im-
prove the accuracy of mine pressure prediction at fully mechanized mining faces. However, the time window setting (including his-
torical data length and predicted data length) directly affects the prediction performance of LSTM. Therefore, the influence of histori-
cal data length and predicted data length on the prediction accuracy of the mine pressure is studied. First, this paper uses deep learn-
ing method long-term and short-term memory network (LSTM) to analyze the historical data of the mine pressure of the 14160 com-
prehensive mining face of the eighth Mine in Pingmei, and predict the future data of the mine pressure. Secondly, the influence of
historical data length applied to the prediction on the prediction performance of the model is studied, and the optimal window width
is determined when the prediction accuracy is the highest. Furthermore, the prediction data length that the model can be used to pre-
dict the mine pressure data within the allowable range of mine pressure prediction accuracy is studied. Finally, using the optimal win-
dow width, the mine pressure prediction of the fully mechanized mining face is realized within the model prediction range. The ex-
perimental results show that the time window setting has a significant impact on the prediction results. The accuracy of the mine
pressure prediction of the fully mechanized mining face can be improved by optimizing the time window setting.
