Dynamic Risk Forecasting Based on Deep Learning and Collapse Risk Comprehensive Evaluation of Mountain Tunnel Portal Construction

Abstract

In this paper, a comprehensive risk assessment system is proposed to evaluate the risk of collapse in mountain tunnels. This system integrates risk source identification, dynamic and static risk classification, deep learning prediction, and engineering risk evaluation. Firstly, risk events and sources are identified, and a risk evaluation method combines the fuzzy analytic hierarchy process (FAHP) and interval technique for order preference by similarity to ideal solution (TOPSIS). FAHP is used to calculate weights, and a risk classification table based on five classical values is derived using traditional TOPSIS. The actual project’s risk value is then calculated using Interval TOPSIS to determine the risk level. Secondly, six models (BP, SVM, CNN, LSTM, PSO-SLTM, and EPL) are trained and tested to predict surface settlement at the tunnel portal and using RMSE, MAE, and maximum (minimum and average) error values for comparison; the best model is determined. The study concludes that a two-stage model, which uses ensemble empirical mode decomposition to process raw data and particle swarm optimization to optimize long short-term memory hyperparameters, provides the best predictive results. Finally, static and dynamic risks are combined for a comprehensive risk evaluation. The Aktepe Tunnel Project in Xinjiang, China, serves as a case study to successfully and accurately forecast surface settlement and evaluate the safety of the tunnel portal. This assessment confirms that this section of the tunnel is at average risk and that the current building conditions ensure the safety of the tunnel, the case study validates the rationality of the comprehensive evaluation system, offering a reference for tunnel portal risk evaluation.