<p>In recent years, the settlement of foundation pits has garnered significant attention, with settlement deformation and prediction being particularly crucial. However, the nonlinear fluctuations caused by surrounding loads and uncertain factors reduce the accuracy of traditional numerical simulation statistical methods. This study focuses on the deep excavation project at a Station of Metro Line 8. It proposes a displacement back-analysis method based on a Backtracking Search Algorithm (BSA) optimized Back Propagation (BP) neural network (BSA-BP) to invert and optimize soil layer parameters. Additionally, a time-series prediction model using Long Short-Term Memory (LSTM) neural networks is introduced to predict surface and underground pipeline settlements. The model utilizes historical settlement data for rolling predictions. Results indicate that the BSA-BP inversion model excels in generalization, robustness, and convergence speed, outperforming traditional BP networks with a relative error of only 1.07% between predicted and measured values. The LSTM model demonstrates high precision and stability, with superior accuracy compared to traditional BP and Support Vector Machine (SVM) methods, effectively leveraging historical data for accurate predictions.</p>