Abstract: The roof pressure is usually measured by the hydraulic support working resistance, and the effect of the roof pressure prediction method based on depth learning is greatly affected by the training sample set. The construction of the training sample set depends on the selection of the time window and the identification of the closely related hydraulic support group. However, the existing methods rely on manual experience to determine the time window, and ignore the correlation between different hydraulic supports, which seriously hinders the improvement of the roof pressure prediction precision. In order to solve the above problems, a roof pressure prediction method of coal working face based on spatiotemporal correlation analysis is proposed. Firstly, the optimal time window is selected by calculating the grey correlation degree of working resistance series of the same hydraulic support in the time dimension. Secondly, the optimal auxiliary matrix is obtained by calculating the grey correlation degree of working resistance sequences of different hydraulic support in spatial dimension, and the closely related hydraulic support group is identified. Finally, based on the optimal time window and the optimal auxiliary matrix, the label and corresponding characteristics of each training sample are determined, and the training sample set is constructed to train the long short time memory (LSTM) model to predict the roof pressure. The experimental results show that the proposed method can reduce the prediction error of roof pressure effectively compared with the method which relies on manual experience to construct training sample sets to complete LSTM model training.