为提升地铁施工过程中周边地表沉降预测的精度与可靠性，基于某市轨道交通1号线工程数据，提出一种基于小波去噪(WTD)、相空间重构(PSR)和网格搜索优化支持向量机回归(GSM-SVR)的组合预测模型，实现地表沉降预测的精细化研究。首先，利用小波阈值去噪法(db4函数)对原始监测数据进行预处理，有效剔除噪声和奇异值，信噪比(R_SNR)提升至11.814 4，均方根误差(E_RMSE）降至0.373 83； 其次，通过相空间重构(嵌入维数m=5，延迟时间τ=1)将静态时间序列转化为多维动态特征数据，增强数据的非线性表征能力； 最后，采用网格搜索法(GSM)优化SVR模型的超参数(惩罚参数c=50，核参数g=0.38)，构建GSM-SVR预测模型，并采用传统SVR、PSO-SVR、GWO-SVR等6种模型进行对比分析，检验GSM-SVR模型的优劣度。结果表明： 1)WTD-PSR有效地剔除了噪声和奇异值，且提升了数据多维动态特征； 2)GSM-SVR模型预测结果与其他6种模型相比，预测误差(E_MAPE、E_MAE、E_RMSE)降低了8.44%~64.35%，拟合度(R²)提高了0.51%~22.08%，平均误差仅3.68%； 3)去噪处理使模型性能显著提升，预测误差降低了12.33%~19.70%，可对地铁施工沉降进行有效的预测。

To predict the settlement of surrounding surfaces during metro construction with high accuracy and reliability, the authors propose a combined model based on wavelet threshold denoising (WTD), phase space reconstruction (PSR), and a grid search method-optimized support vector machine regression model (GSM-SVR). The model is built using engineering data of a city′s Metro Line 1. First, the WTD method (db4 function) is applied to preprocess the raw monitoring data, effectively eliminating noise and singular values. This step raises the signal-to-noise ratio to 11.814 4 and reduces the root mean square error (E_RMSE) to 0.373 83. Second, PSR (embedding dimension=5 and delay time=1) is used to convert the static time series into multidimensional dynamic feature data, thereby enhancing the nonlinear representation capability of the data. Finally, the hyperparameters of the SVR model (c=50 and g=0.38) are optimized using the GSM to construct a GSM-SVR prediction model. Subsequently, six comparative models, including the traditional SVR, particle swarm optimization-SVR, and gray wolf optimizer-SVR models, are used for performance evaluation. Results indicate the following: (1) WTD-PSR effectively removes noise and singular values while enhancing the multidimensional dynamic characteristics of the data. (2) Compared with the other models, the GSM-SVR model reduces prediction errors (E_MAPE, E_{MAE, and} E_RMSE) by 8.44%－64.35% and improves the goodness of fit (R²) by 0.51%－22.08% while achieving an average error of only 3.68%. (3) Denoising remarkably enhances model performance, reducing prediction errors by 12.33%－19.70%. Thus, the proposed model effectively predicts the settlement during metro construction.