With the increasing development in urban underground spaces towards greater depth, scale, and complexity, the prediction and control of deformations in deep excavation engineering have become critical challenges in geotechnical engineering. This study investigates the ultra-deep excavation of the Tianjin Metro Line 8 Liulitai Station, analyzing the deformation characteristics of the retaining structure during top-down construction in soft soil based on field monitoring data. The results reveal a typical “bulging” pattern in the horizontal displacement of the diaphragm wall, which accumulates progressively with excavation depth. To enhance deformation prediction accuracy, a self-developed beam-plate finite element method (BPFEM) platform, implemented in Python (version 3.11.9), is introduced. The platform integrates code-specified analytical methods and the incremental approach to simulate the internal forces and deformations of the support system with high precision. By incorporating a dual-parameter back-analysis technique—adjusting both the horizontal subgrade reaction modulus and active earth pressure—the numerical model achieves significantly improved agreement with monitoring data. The proposed method demonstrates strong predictive capability, with a maximum error of only 4.4% in subsequent construction stages, confirming its feasibility and reliability for deformation forecasting in top-down deep excavations. The BPFEM framework and parameter inversion strategy presented herein provide an effective technical basis for intelligent prediction and dynamic control in deep excavation projects under complex geological conditions.