Abstract: To address the challenge of accurately and efficiently solving the stress state of main structural components of hydraulic supports in the complex and harsh underground coal mine environment, a stress state determination method based on the fusion analysis of support posture and load is proposed. An orthogonal experimental approach is employed to construct a dataset for hydraulic support stress state determination. To overcome the issues of irregular data sample topology, parameter explosion, and difficulty in multi-scale feature extraction, a multi-scale graph convolutional network model incorporating physical constraints is developed. A dual-branch feature encoding mechanism is adopted to independently encode and deeply fuse the operational pressure parameters representing the load state and the nodal spatial coordinates representing the geometric posture. A dynamic KNN graph construction strategy based on spatial partitioning is designed to break through the memory bottleneck of large-scale node graph construction by block-wise computation. A multi-scale feature aggregation module is introduced to extract multi-resolution features via parallel branches with different downsampling rates, enabling simultaneous capture of local stress concentrations and global deformation trends. A composite loss function integrating spatial consistency and smoothness constraints is formulated to incorporate continuum mechanics priors into the network training process. Comparative experimental analysis shows that the proposed model significantly outperforms MLP, PointNet, and GCN models in terms of mean squared error, mean absolute error, coefficient of determination, and mean relative error, and achieves an order-of-magnitude improvement in inference efficiency compared with conventional numerical simulation. Ablation experiments validate the synergistic contributions of the multi-scale feature fusion module and the physically constrained loss function to model performance. Data efficiency experiments confirm that the physical constraints effectively suppress overfitting when sufficient data are available.