Abstract: To address the pronounced spatial asynchrony of ground-pressure responses and strong noise interference in ground-pressure signals in super-long working faces, an adaptive zoning and prediction method integrating spatiotemporal topological constraints and modal-feature reconstruction is proposed. First, considering the differences in the temporal evolution and spatial adjacency of ground-pressure responses at different support locations, Dynamic Time Warping (DTW) is employed to quantify the temporal similarity between support-pressure sequences. A spatiotemporal joint distance model is then established by incorporating spatial topological constraints, and K-Medoids clustering is used to achieve adaptive zoning of the working face, thereby providing a spatial-topological basis for regionalized ground-pressure modeling. Subsequently, Frozen-Frequency Variational Mode Decomposition (FF-VMD) is adopted, in which the frequency priors and decomposition parameters of the ground-pressure signals are determined exclusively from the training set and kept fixed during testing, enabling leakage-free modal decomposition of end-of-cycle support-resistance sequences while suppressing high-frequency noise. On this basis, independent regional Attention-based Spatiotemporal Graph Convolutional Networks (AST-GCN) are constructed. Graph convolution is used to extract spatial correlations among supports within each region, while temporal convolution and an attention mechanism are employed to learn the temporal dependencies of ground-pressure evolution. The modal components are jointly predicted and subsequently reconstructed to obtain the end-of-cycle support resistance. A dynamic regional-block discrimination method is further introduced to achieve the spatiotemporal reconstruction and identification of periodic weighting regions. The results show that, for reconstructed end-of-cycle support-resistance prediction, AST-GCN achieves a Mean Absolute Error (MAE) of 0.59 MPa, a Root Mean Square Error (RMSE) of 0.73 MPa, a Mean Absolute Percentage Error (MAPE) of 1.47%, and a coefficient of determination (R²) of 0.959 8, with lower prediction errors than Long Short-Term Memory (LSTM), Spatio-Temporal Long Short-Term Memory (ST-LSTM), and Spatial-Temporal Graph Convolutional Network (ST-GCN) models. For periodic weighting region discrimination based on modal reconstruction, the F1-score between the predicted and measured-data-based discrimination results reaches 0.719 4, indicating that the proposed method can identify the spatiotemporal distribution characteristics of periodic weighting regions under complex ground-pressure conditions and demonstrates good applicability when re-established and validated on a working face with different geological conditions. The results demonstrate that the proposed method can characterize the regional differences, spatial correlations, and temporal evolution of ground pressure in super-long working faces, enabling spatiotemporal prediction and identification of periodic weighting regions.