为解决传统初始地应力场反演方法存在边界条件筛选能力弱、易受数据过拟合干扰以及难以解析多重边界相互作用的问题，提出一种基于LASSO-OLS（least absolute shrinkage and selection operator-ordinary least squares)的两阶段初始地应力场反演方法。该方法首先通过对候选边界条件应力矩阵和实测应力矩阵进行Frobenius范数标准化处理，消除不同边界条件数据量级差异的影响；然后，利用LASSO回归的L1正则化约束，从候选边界条件的回归系数路径图中筛选关键影响因素，剔除冗余与弱相关项；最后，针对筛选出的核心变量，采用普通最小二乘回归进行无偏估计，构建兼具稀疏性与准确性的地应力场反演模型。研究结果表明： 1）在工程应用实例中，借助LASSO回归从11个候选边界条件中筛选出5个关键因素，显著降低模型复杂度； 2）模型正则化参数在标准误差内取值，拟合结果能够保持较高的复相关系数（R=0.995 2），表明筛选后的边界条件有效捕捉了初始地应力场特征； 3）初始地应力场反演模型通过LASSO回归筛选，在解析多重边界相互作用时表现出较高的稳定性和物理合理性； 4）与传统方法相比，该方法能有效避免初始地应力场反演出现过拟合问题，提高反演结果的鲁棒性。

Traditional initial geostress inversion methods suffer from limited boundary condition selection capability, susceptibility to overfitting, and challenges in resolving interactions among multiple boundaries. To overcome these drawbacks, a two-stage inversion method based on the least absolute shrinkage and selection operator (LASSO) and ordinary least squares (OLS) is proposed. First, the candidate boundary condition stress matrix and the measured stress matrix are normalized using the Frobenius norm to eliminate the effect of magnitude differences among boundary conditions. LASSO regression, with its L1 regularization constraint, is then applied to identify key influencing factors from the regression coefficient path diagram, thereby eliminating redundant and weakly correlated terms. For the selected core variables, OLS regression is subsequently employed to achieve unbiased estimation, constructing a geostress inversion model that balances sparsity and accuracy. The results show that: (1) LASSO regression selects five key factors from 11 candidate boundary conditions, significantly reducing model complexity. (2) When the regularization parameter is chosen within the standard error range, the fitting results maintain a high multiple correlation coefficient (R=0.995 2), demonstrating that the selected boundary conditions effectively capture the characteristics of the initial geostress field. (3) The inversion model screened via LASSO regression exhibits improved stability and physical reasonableness when analyzing interactions among multiple boundaries. (4) Compared with traditional methods, this approach effectively avoids overfitting in initial geostress inversion and enhances the robustness of the results.