Abstract: Accurate prediction of water inrush risk levels in karst tunnels in southwest China is essential to prevent casualties, construction delays, and economic losses caused by such disasters. Based on prior research, the factors influencing water inrush in karst tunnels are identified. A comprehensive evaluation system is constructed using nine indices: formation lithology, topography, rock formation, rock mass quality index, rock formation permeability coefficient, sudden water inflow, groundwater level, annual average precipitation, and tunnel burial depth. A hybrid deep learning model integrating a convolutional neural network (CNN) and a long short-term memory (LSTM) network is proposed. Several karst tunnels in southwest China are used for model training and validation, and the performance is compared with standalone CNN and LSTM models. The results show that: (1) The prediction accuracies of the hybrid model for water inrush risk levels Ⅰ, Ⅱ, Ⅲ, and Ⅳ reach 95%, 85%, 87%, and 91%, respectively. (2) The hybrid model outperforms traditional CNN and LSTM models in terms of both prediction accuracy and generalization capability.

Key words: karst tunnel, water inrush prediction, convolutional neural network, long short-term neural network, deep learning