Design of a tunnel water inrush risk prediction model incorporating attention mechanisms

The coupling of geological conditions, hydrodynamic responses, structural displacement, and construction disturbance makes the temporal evolution of tunnel water-inflow risk and corresponding actions difficult to predict using traditional techniques. To address this problem, a multi-source monitoring feature system was established, including pore water pressure, inflow volume, surrounding rock deformation, support strain, geological permeability, and excavation disturbance. Sliding windows were then used to construct time-series samples for a prediction framework containing dilated convolution, bidirectional GRU, and multi-head attention. The methodological novelty does not arise from combining these established modules alone. It lies in a warning-oriented temporal representation that jointly encodes raw states and first-order changes, aligns abrupt seepage mutations with delayed hydraulic–structural responses through parallel branches, and links interpretable precursor evidence to four-level construction actions. Under chronological validation across six monitoring sections of a single tunnel, the model achieved an accuracy of 0.949, an F1-SCORE of 0.946, and an AUC of 0.978. These results represent within-project temporal validation and do not establish transferability to other tunnels or hydrogeological conditions. Within this validation scope, the framework supports risk classification, precursor interpretation, and warning-response linkage at the project scale, thereby extending generic category prediction into response closure.