Application of hybrid-optimized and stacking-ensemble labeled neural networks to predict water inflow in drill-and-blast tunnels

Abstract

The precise estimation of water inflow (WI) into the tunnel during the construction phase, as one of the engineering geological hazards, is one of the most critical factors for project advancement and utilization, especially in the early design stages. To address this, the current study developed several predictor networks, including hybrid-optimized supervised learning models such as AdaDelta-recurrent neural network (AdaD-RNN), AdaGrad-long short-term memory (AdaG-LSTM), AdaGrad-gated recurrent unit (AdaG-GRU), Adam optimization-back propagation neural network (AO-BPNN), automatic linear forward stepwise information criterion (ALFS-IC), and a novel stacking-ensemble model. These models were trained and validated using a collected database from 13 drill-and-blast road tunnels in Iran. A new empirical model for predicting tunnel WI was introduced using ALFS-IC with high accuracy (R² = 0.95). The models were trained on a dataset with five features and 600 data points (85 % training, 15 % testing), including physical factors of tunnels (tunnel depth, groundwater level), geomechanical characteristics of materials (rock quality designation), and water inrush feature (water yield property). The importance ranking and multi-task sensitivity analysis revealed that groundwater level and water yield property are the most influential parameters on the road tunnel WI. The analysis indicated strong correlations between predicted and observed values, with the stacking-ensemble and AdaG-GRU models exhibiting superior accuracy in predicting WI into the tunnel with R² = 0.97 and 0.95 and NRMSE = 0.0017 and 0.0019, respectively. The stacking-ensemble algorithm had the highest accuracy rate of 90 % and AUC-ROC value of 98 %.