Abstract: This paper takes the characteristics of roof pressure behavior during mining at the fully mechanized mining face of Shaanxi Coal Group Shenmu Hongliulin Mining Co., Ltd. as the research background. By analyzing the movement and failure laws of overlying rocks in the fully mechanized mining face with large mining height, combined with the interaction relationship between supports and roof, mathematical statistics and deep learning algorithms are used to analyze the collected time-series mine pressure data, and the forecast and early warning of roof pressure in the fully mechanized mining face are realized. Firstly, by analyzing the support pressure data of the fully mechanized working face, the support pressure working stage is divided into the resistance reduction stage, the slow resistance increase stage, and the rapid resistance increase stage; secondly, the pressure determination index of the working face "pressure arch" is proposed; then based on the LSTM network, a mine pressure prediction model is designed to predict the future pressure change of the working face; The periodic pressure prediction results are also divided into sections. The concentrated trend and frequency of prediction errors are analyzed through on-site application cases, and the prediction results of different algorithms are compared. The results show that the pressure prediction model based on deep learning has a high accuracy, but the prediction time is shorter. Compared with the historical pressure prediction model, it is less affected by factors such as geological conditions. The research results can provide certain reference for the intelligent production and management of fully mechanized working faces.