Self-Adaptive Parameters Optimization of a Physics-Data Hybrid Driven Model For Surface Settlement Prediction

Abstract

Tunnel boring machines play a pivotal role in the construction of urban underground railway networks. However, accurately predicting surface settlement resulting from shield tunneling is challenged by intricate geological factors, posing a significant hurdle in ensuring prediction precision. Moreover, prevailing neural network models for settlement prediction suffer from poor generalization and interpretability. This paper addresses the research question: how to train a neural network for predicting ground subsidence caused by tunnel excavation using a limited dataset under complex geological conditions while improving its interpretability. To achieve this, we propose a physics-data hybrid driven model that leverages prior domain knowledge to improve interpretability while upholding accuracy. Our model comprises: (1) Establishing a deep neural network for predicting ground settlement; (2) Integrating physics knowledge into the DNN-based prediction model; (3) Employing the Tree Structured Parzen Estimator for hyperparameter estimation to optimize network parameters and enhance predictive capability. We evaluate the effectiveness and feasibility of our proposed approach using the San-yang Road tunnel project in Wuhan, China. Experimental results demonstrate that our physics-data hybrid driven model accurately predicts ground settlement resulting from tunnel excavation. Our methodology enhances model interpretability while effectively mitigating high-risk situations.