Multi-objective optimization for limiting tunnel-induced damages considering uncertainties

Abstract

Due to the rapid development of the urban metro system, the situation of new excavation work being conducted adjacent to existing tunnels is quite common and becomes prime hazards in the tunnel design stage, together with uncertainties from the ground condition. To solve this problem, this paper develops a hybrid approach that integrates ensemble learning and non-dominant sorting genetic algorithm-II (NSGA-II) to mitigate the limit support pressure (LSP) and the ground surface deformation (GSD) during the tunnel excavation for improved design. The extreme gradient boosting (XGBoost) algorithm is used to establish ensemble learning models predicting LSP and GSD, where the new tunnel is constructed in parallel to an existing tunnel. NSGA-II is further used to optimize the two targets (i.e., LSP and GSD), considering the uncertainties from geotechnical conditions and errors from the meta-model. With the Monte-Carlo simulation, probability constraints are established to conduct the multi-objective optimization (MOO). Finally, the Pareto front is generated to obtain the best location of the new tunnel, and a comparison is made between MOO with and without considering uncertainties. The best solution is selected by the criterion of the point with the shortest distance from the ideal point. It is found that after considering uncertainties: (1) The improvement percentage of LSP is increased from 9.67% to 11.03%, and that of GSD drops from 2.39% to 0.9%; (2) A higher stability of improvement from optimization is achieved with the standard deviation of improvement percentage drops from 0.310 to 0.298 for LSP and 0.024 to 0.020 for GSD; (3) With a weaker confidence on the meta-model, a higher degree of sacrifice on GSD is observed. The novelty of the proposed approach lies in its capability to not only predict and optimize the damage from excavation adjacent to an existing tunnel, but also consider various types of uncertainties from geological conditions and meta-models to guarantee reliability.