Abstract: At present, roof weighting prediction models mainly rely on data collected by hydraulic support leg pressure sensors and usually adopt time series analysis methods for prediction with time as the horizontal coordinate. However, such methods can hardly effectively characterize the spatial location characteristics of roof weighting, which limits the accuracy and applicability of prediction results in engineering applications. To solve this problem, a prediction model for roof weighting location and intensity in fully mechanized mining faces based on the combination of temporal convolutional network (TCN) and long short-term memory (LSTM) network is proposed. Firstly, based on the hydraulic support leg pressure data, the cycle-end resistance at the end of each coal cutting cycle is extracted, and the support displacement information at the corresponding moment is further determined combined with hydraulic support displacement data. In this way, the cycle-end resistance sequence varying with mining distance can be obtained, so that spatial location features are introduced in addition to the time dimension to enhance the ability to predict the spatial distribution of roof weighting. Subsequently, the TCN-LSTM hybrid neural network model is used to conduct feature extraction and modeling on the processed data, so as to realize accurate prediction of roof weighting location and intensity. The experimental results show that compared with the particle swarm optimization back propagation (PSO-BP) neural network, the LSTM model and the Transformer model, the proposed TCN-LSTM model has significant advantages in prediction accuracy. The root mean square errors of roof weighting location and intensity prediction are reduced by 37%, 30% and 42% respectively, and the mean absolute errors are reduced by 37%, 31% and 40% respectively. The research indicates that the proposed method can effectively improve the accuracy of roof weighting prediction in fully mechanized mining faces.