Abstract: In fully mechanized top coal caving(FMTCC) mining of extra-thick coal seams, traditional strata pressure prediction models often fail to account for the dynamic correlation of support working resistances, making it difficult to accurately characterize the strata pressure evolution law.This frequently leads to inaccurate support suitability criteria and ineffective weighting forecasts.Therefore, there is an urgent need to develop a weighting prediction model based on the dynamic correlation of support working resistances to achieve visual expression of the strong heterogeneity, high dynamism, and significant disturbance characteristics of strata behavior in FMTCC of extra-thick coal seams, thereby ensuring accurate and efficient weighting prediction.This study, based on the FMTCC mining of the extra-thick coal seam in Mengtai Bulian Gou Mine, proposes a dynamically Adaptive Graph Convolutional Recurrent Network(AGCRN) model for strata pressure prediction using deep learning and mathematical calculation methods.Through the collection, screening, and analysis of support resistance data from the No.6 coal seam, along with comparative experiments involving three types of traditional weighting prediction models, the results show that the model achieves optimal performance with an embedding dimension of 10 and a time window of 16 after hyperparameter optimization.The model's Mean Absolute Error(MAE) and Mean Absolute Percentage Error(MAPE) are as low as 551.45～601.68 and 4.49%～6.19% respectively, representing the smallest errors compared to other benchmark models and exhibiting higher accuracy and stability, which verifies the precision and superiority of the AGCRN weighting prediction model for strata pressure prediction in the extra-thick coal seam of Bulian Gou Mine.