【Objective】In order to solve the parameter optimization problem of the three-step excavation method for shallow-buried soft surrounding rock (Class V) tunnels,【Method】factors like step advance to tunnel half-span ratio (R), cohesion, internal friction angle, surrounding rock elastic modulus, and Poisson"s ratio are chosen. Multilevel orthogonal finite element analysis gets vault settlement values for different parameter combinations. The support vector regression model (SVR) is optimized by the adaptive particle swarm algorithm (APSO algorithm) to build a vault settlement prediction model for three-step excavation and an SVR model for comparison. The APSO-SVR model uses vault settlement as input and R, cohesion, etc. as output to invert R. Based on actual engineering, multilevel orthogonal finite element analysis obtains parameter combination values. Midas GTX NX verifies step excavation limit footage parameter accuracy. 【Result】Results: 1) The ASPO-SVR model has lower error and higher accuracy than the SVR model, validating its optimization. 2) Midas inversion shows a 0.72% difference between the simulated and specified arch top settlement values, confirming the optimized parameter scheme"s accuracy. 【Conclusion】The inversion optimization model based on APSO-SVR for three-step excavation has high accuracy and can guide similar project safety construction.