Accurate localization of rock fracture events is critical for timely warning the rockburst events in microseismic monitoring systems. Conventional physics-based inversion techniques and data-driven neural networks rely heavily on P- and S-wave arrival-time, and their performance often degrades substantially when suffer from abnormal arrival-time. To address this limitation, we propose an Attention-Enhanced Hybrid Graph Neural Network (AE-HGNN) for three-dimensional (3d) event localization. In this framework, a graph representation is constructed by embedding both sensor coordinates and arrival-time information, thereby enabling the effective modeling of spatial—temporal dependencies. Subsequently, the architecture incorporates a Graph Attention Network (GAT) module for temporal feature extraction and an attention-enhanced Graph Isomorphism Network (GIN) module for spatial feature learning. Finally, three independent output heads are employed to predict the 3d coordinates of fracture events along each axis, achieving precise localization. Experimental evaluations on localization datasets demonstrate that AE-HGNN consistently outperforms both traditional inversion-based approaches and purely learning-based models in terms of accuracy. Moreover, case studies from Tunnel Boring Machine (TBM) excavation projects highlight the practical applicability and engineering significance of the proposed method in real-world scenarios.