基于CNN-LSTM的深基坑挡墙变形时空分布预测方法OA

To achieve accurate prediction and effective control of retaining wall deflection in soft soil excavations and to ensure construction safety,this study developed a spatiotemporal distribution matrix of retaining wall displacement based on its significant spatiotemporal distribution characteristics and created a hybrid CNN-LSTM prediction model that integrates convolutional neural networks(CNN)and long short term memory(LSTM).The research synchronously forecasted and compared the wall deflection from both temporal and spatial dimensions through a deep excavation project in Shanghai.The results show that:1)Compared with four traditional conventional models,the CNN-LSTM hybrid prediction model,based on the spatiotemporal distribution matrix of the retaining wall displacement,demonstrates precise prediction of the spatiotemporal distribution of horizontal displacement through the extraction of spatiotemporal distribution characteristics and deep learning.2)For spatial distribution prediction,the extraction of spatial distribution features combined with deep learning enables not only accurate identification of the deflection mode of the retaining wall but also precise prediction of distribution features such as deformation curvature and positions corresponding to maximum deflection.The predicted MAE in depth and horizontal directions is 0.532 mm and 0.742 mm,respectively.3)For time distribution prediction,the dynamic forecasting of retaining wall displacement at various construction stages is achieved through the extraction of horizontal displacement time series features and deep learning,while considering both short-and long-term data dependencies.The predicted results during the construction period demonstrate good robustness,with an MAE of 0.841 mm.