目前，我国地铁建设正不断推进，新建地铁车站与既有车站的线路交汇，使得新建地铁车站基坑向更大、更深、更复杂的趋势发展。这提升了地铁车站基坑施工的难度，安全事故也更容易发生，威胁紧邻环境和居民的生命财产安全。同时基坑施工过程中围护结构变形的控制存在不足，如变形预测不准确、变形控制针对性不强等问题。因此，本文基于地铁基坑变形的监测数据，对基坑变形的预测、基坑变形主动控制进行研究，提出了系统的深基坑变形管理方法。首先对基坑施工变形风险进行分析，总结深基坑变形类型及机理，提出基于时间序列模型预测基坑变形，再通过数值模拟方法提出基坑变形主动控制措施，由此来对基坑变形进行预警及控制。具体的实验步骤如下：首先，采集某地铁车站基坑的监测数据，针对原始数据非等时距特点，利用 Akima 插值法进行填补。其次，将预处理后的数据分为单特征和多特征数据集，输入到深度神经网络（DNN）、卷积神经网络（CNN）、极限梯度提升（XGBoost）和门控循环单元（GRU）模型中分别进行训练，对比各模型的性能指标，得出预测深基坑地连墙变形的最优特征输入以及最优预测模型。针对预测到的变形超限情况，建立了采用液压伺服钢支撑和传统钢支撑的深基坑数值模拟模型，模拟液压伺服钢支撑在抑制基坑变形中主动控制的过程，对比不同支护方式下的地下连续墙深层水平位移结果。最后，针对某地铁深基坑变形风险，提出采用轴力伺服系统增设临时钢支撑和受力最大的第五道钢支撑优化，取得了良好变形控制效果。本研究中基于同一环境多特征作为输入的 GRU 模型在预测基坑地连墙变形预测性能最优（R2=0.94）。对比不同支护方式数值模拟结果，结果表明液压伺服钢支撑能利用主动增加轴力实现基坑地连墙变形控制，将地连墙变形控制在限值（32mm）内。并进一步以某地铁车站施工为背景，对提出的方法和模型进行工程验证，结果表明该方法能够准确的预测基坑变形，实现基坑变形的主动控制。

关键词：深基坑； 变形预测； 机器学习； 变形主动控制；数值模拟

At present, China's metro construction is constantly developing, and the intersection of new metro stations with the lines of existing stations makes the new metro station pits develop towards a larger, deeper and more complex trend, which promotes the difficulty of metro station pit construction, and makes safety accidents more likely to occur, thus threatening the immediate environment and the life and property security of residents. The management of metro station pit deformation is inadequate, with issues such as inaccurate deformation prediction and lack of targeted deformation control. Therefore, based on the monitoring data of metro station pit deformation, this study conducts research on the prediction of pit deformation and active control of pit deformation, and proposes a systematic method for deep pit deformation management.
Firstly, we conducted an analysis of the deformation risks during pit construction, which summarized the types and mechanisms of deep pit deformation, and proposed a time series model for predicting pit deformation based on this analysis. Subsequently, active control measures for pit deformation are proposed using numerical simulation methods to provide early warning and control of pit deformation. The specific experimental steps are as follows: Initially, we collected monitoring data from a certain metro station pit. To address the non-equidistant nature of the raw data, the Akima interpolation method is employed for data filling. Next, the preprocessed data is divided into single-feature and multi-feature datasets, which are then input into Deep Neural Network (DNN), Convolutional Neural Network (CNN), Extreme Gradient Boosting (XGBoost), and Gated Recurrent Unit (GRU) models for training. The performance metrics of each model are compared to determine the optimal feature input and prediction model for forecasting deep pit wall deformation. For cases where the predicted deformation exceeds limits, we established a numerical simulation model for deep pits using hydraulic servo steel support and traditional steel support which simulated the active control process of hydraulic servo steel support in restraining pit deformation, and compared the deep-level horizontal displacement results of underground continuous walls under different support methods. Finally, to address the deformation risks of a specific metro station deep pit, the addition of temporary steel supports using an axial force servo system and optimization of the fifth steel support with the maximum force gotpositive results in deformation control. This study demonstrates that the GRU model with multiple features as input in the same environment shows the optimal predictive performance for predicting pit wall deformation (R2=0.94). A comparison of numerical simulation results for different support methods reveals that hydraulic servo steel support can actively increase axial force to control pit wall deformation, and maintain the wall deformation within the specified limit (32mm). Furthermore, using a specific metro station construction scenario as a backdrop, the proposed method and model are validated in engineering applications. The results confirm that this approach can accurately predict pit deformation and achieve active control of pit deformation.