Forecasting long-term behavior of chronological tunnel longitudinal settlements using machine learning techniques

Abstract The metro tunnels of Shanghai suffered from long-term longitudinal settlements, which are detrimental to their functionality with potential risks of hazards in their daily operations. It is crucial to develop prediction models to be able to assess the extent of such settlements in tunnels so that necessary precautionary measures can be taken ahead of time. The existing research either tries to predict overall tunnel responses including various localized factors or uses only short-term predictions based on immediate earlier measurements. This research focuses on performing long-term predictions of the chronological settlement behavior along the tunnel operational years using limited records of settlements during the initial operational years. Three state-of-the-art machine learning (ML) algorithms, namely the random forest, the multilayer perceptron, and k-nearest neighbors, were adopted to develop predictive models. A stochastic hill climbing technique with adaptive steps was used for the optimization of the hyperparameters. The training and validation of the ML models were performed using the settlement records of the Shanghai Metro Line 1. With a high prediction accuracy (R2 > 0.94), all the ML models demonstrated successful prediction performances for the time-varying settlements of the tunnel as far as 14 years apart.