Prediction and Application Research on Deformation Characteristics of Horizontal Sandstone and Siltstone Formation Tunnels
Year: 2026
Type: article
Abstract: Investigating the relationship between tunnel burial depth and deformation in horizontal sandstone and mudstone strata provides significant guidance for tunnel support design. This study proposes a unified global modeling strategy that employs improved LSTM (Long Short-Term Memory) and RNN (Recurrent Neural Network) algorithms to analyze data from 146,645 monitoring data points at varying burial depths, systematically examining the correlation patterns between burial depth and deformation magnitude. The results indicate that the correlation coefficients between burial depth and arch settlement, cumulative convergence values of the upper platform, and cumulative deformation values of the lower platform range from −0.184 to −0.194 (p < 0.01): the relationship is statistically significant but weak, accounting for only about 3.4–3.8% of the variance in deformation (r2 = 0.034–0.038), so that burial depth alone is of limited practical value for prediction. The spread of the measurements nevertheless varies systematically with depth: in the present dataset, arch settlement and cumulative upper convergence do not exceed 50 mm at burial depths ≤ 150 m and do not exceed 30 mm at depths > 150 m, while the cumulative lower convergence is mostly within 42 mm at depths ≤ 110 m and within 15 mm at depths > 110 m. These transition depths (150 m and 110 m) are ranges observed in the present dataset rather than established critical thresholds. The hybrid LSTM-CNN and RNN-CNN models achieve 4.5%/4.4% reductions in MAE/RMSE compared to pure time-series models, demonstrating stronger robustness in multi-step prediction, and only this dual LSTM-CNN structure converts a longer input window into a genuine gain: its 20-day MAE and RMSE (0.4870 and 0.5704) are the lowest obtained in this study and even fall below the best 5-day result. Notably, the LSTM-CNN maintains high accuracy across a 20-day time horizon and full-range fluctuations (−2 to 15 mm), further highlighting the effectiveness of the convolution–time-series synergy mechanism in extracting local temporal features and suppressing errors. This study provides a theoretically rigorous and practically applicable technical framework for predicting deformation in complex geological tunnels and enabling intelligent early warning systems.
Source: Applied Sciences
DOI: 10.3390/app162111082
Institution: Chengdu University
Language: English
Topic: Tunneling and Rock Mechanics
Subfield: Civil and Structural Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $2,896
OpenAlex ID: W7225629162