Data-Driven Method for Predicting Long-Term Underground Pipeline Settlement Induced by Rectangular Pipe Jacking Tunnel Construction

Abstract

The long-term settlement of deep pipelines is a crucial factor in the construction of rectangular shield tunnels. Conventional prediction techniques primarily depend on empirical models and statistical analyses, which often fail to accurately forecast pipeline deformation and settlement in intricate environments. To tackle this challenge, this article introduces a long-term settlement prediction model for pipelines based on long short-term memory (LSTM) networks, incorporating time decay and multiscale improved self-attention mechanisms (ISA), referred to as the LSTM-ISA model. Initially, settlement data that matched theoretical expectations were selected from 69 sets of monitoring data to create a data set for predicting pipeline settlement during shield tunnel construction. The LSTM-ISA model was then developed, using the LSTM network to capture temporal dependencies in time-series data. The time-decay mechanism gives greater importance to more recent data, while the multiscale self-attention mechanism identifies features across various time scales. The model’s effectiveness and reliability were tested using actual measurement data from the Changsha Metro Line 6 project and compared with predictions from traditional LSTM and LSTM-SA networks. The findings indicate that the LSTM-ISA model surpasses both the LSTM and LSTM-SA models, achieving a 12.9% and 30.7% decrease in mean squared error (MSE) and a 6.4% and 21.8% decrease in mean absolute error (MAE), respectively. These results imply that the LSTM-ISA model can serve as an effective tool for providing early warnings regarding long-term pipeline settlement caused by the construction of rectangular shield tunnels. In urban tunnel engineering, accurately predicting settlement is crucial to prevent pipeline damage and ensure the safety of surrounding structures. This study proposes a data-driven predictive approach, introducing a pipeline settlement prediction model based on long short-term memory (LSTM) networks, which is enhanced by a time-decay mechanism and a multiscale self-attention mechanism. The model captures the temporal dependencies in settlement data, assigns higher weights to more recent data using time decay, and extracts features from multiple time scales with self-attention. It predicts the next day’s pipeline settlement using data from the previous days, enabling early detection and mitigation of potential settlement risks and improving decision-making during the construction phase. To validate the robustness of the LSTM-ISA model, we applied it to predict the settlement of five pipelines in the Changsha Metro Line 6 project using time-series data. The results show that the LSTM-ISA model achieved a prediction correlation coefficient of over 90% for all test data sets, demonstrating its accuracy under complex geological conditions. Furthermore, by incorporating additional pipeline settlement data from different construction strata, the model’s generalization and scalability can be further enhanced.