To handle more challenging operation security problems in today’s power system, pre-fault transient stability assessment (TSA) is essentially required to promote the awareness of the system stability risks. Fast and analyzable data-driven methods draw much attention in intelligent TSA schemes, but most of them either lack generalization to various operation topologies and fault locations, or fail to operate on developing systems with changeable scales. Some schemes based on graph convolutional networks (GCNs) enjoy promising topology learning but suffer from poor scale reduction, which affects the robustness against system-scale changes. With this in mind, we propose a novel Attention-based Hierarchical Dynamic grAph Pooling nEtwork (AH-DAPE), where a graph-based hierarchical pooling strategy is initiated for effective scale reduction in power systems. The expressive power of hierarchical pooling is enhanced by a spectral unsupervised loss related to power system simplification, while the temporal learning across dynamic coarsened graphs are enabled by integration of inter-graph convolution and mean/maximum operations. Test results on small IEEE 39 Bus system and large IEEE 300 Bus system validate our scheme’s superiority over existing TSA models and robustness against various operation scenarios, especially when applied to new system scales.