CNN⁃LSTM⁃Attention 的狭窄型基坑开挖变形预测

作者：李忠享1，胡科1，太俊1，徐伟1，孙浩然2，袁家威2，崔德山2

作者单位：

(1.中国市政工程中南设计研究总院有限公司，湖北武汉430010；2.中国地质大学（武汉）工程学院，湖北武汉430078)

基金项目：

*国家自然科学基金(42277171)

摘要：

随着城市化进程的加速，地下管廊工程中狭窄型基坑的应用日益广泛。在狭窄型基坑的数值模拟计算中主要应用HSS模型。然而，目前的HSS小应变土体硬化模型在参数选取上多依赖经验值，导致数值模拟结果存在主观性偏差，缺乏更加准确且客观的模型参数。为了提高预测准确性，提出将CNN-LSTM-Attention模型引入基坑实例计算中，该模型结合了卷积神经网络（CNN）对空间特征的提取能力、长短期记忆网络（LSTM）对时序特征的挖掘能力以及注意力机制（Attention），能够更全面地处理基坑开挖变形问题，可以用于其HSS模型中参数的修正。结果表明，利用CNN-LSTM-Attention模型对HSS模型参数进行了反演修正，并得到了更为精确的修正后参数。反演修正后的模型预测值与实测值对比精度显著提高，并且将其与单一BP模型和单一LSTM模型进行了对比。该模型展现出显著优势，其平均绝对误差（MAE）、平均相对误差（MAPE）和均方根误差（RMSE）均优化于对比模型。研究结果表明，CNN-LSTM-Attention模型在狭窄型基坑开挖变形预测中表现出色，能够更准确地反演HSS模型参数，相比于参数反演前其精度提高了312%。同时相较于其他常用模型，CNN-LSTM-Attention模型在预测精度和优化效果上也均实现了显著提升。

关键词：地下工程；综合管廊；基坑；数值模拟；CNN-LSTM-Attention模型；参数反演；预测

English：

With the acceleration of urbanization, narrowtype foundation excavation are increasingly applied in underground pipe gallery projects. In the numerical simulation of narrowtype foundation excavation, the HSS model is mainly used. However, the current HSS smallstrain soil hardening model often relies on empirical values for parameter selection, leading to subjective deviations in numerical simulation results and a lack of more accurate and objective model parameters.To improve prediction accuracy, the CNN-LSTM-Attention model is proposed to be introduced into the calculation of foundation excavation examples. This model combines the spatial feature extraction ability of the convolutional neural network (CNN), the time series feature mining ability of the long shortterm memory network (LSTM), and the attention mechanism (Attention), which can comprehensively handle the deformation problem of foundation excavation and be used to correct the parameters of the HSS model.The results show that the CNN-LSTM-Attention model was used to inversely correct the parameters of the HSS model, and more accurate corrected parameters were obtained. The comparison between the predicted values of the inversely corrected model and the measured values shows a significant improvement in accuracy. Moreover, it was compared with the single BP model and the single LSTM model. This model demonstrated significant advantages, with its mean absolute error (MAE), mean absolute percentage error (MAPE), and root mean square error (RMSE) all being optimized compared to the comparison models.In conclusion, the research results show that the CNN-LSTM-Attention model performs well in the prediction of deformation in narrowtype foundation excavation, and can more accurately inversely correct the parameters of the HSS model. Compared with before parameter inversion, its accuracy has improved by 312%. At the same time, compared with other commonly used models, the CNN-LSTM-Attention model has also achieved significant improvements in prediction accuracy and optimization effects.

岩土与地下工程 > 2026年 第卷 第13期