Prediction Method of Pit Deformation Based on CNN-BiLSTM-Attention

Accurate prediction of pit deformation has always been one of the key problems in pit engineering, and the complexity of underground space and the diversity of pit construction environments make the traditional prediction methods incompetent in dealing with this problem. In order to improve the accuracy of pit deformation prediction, a deep pit deformation prediction model combining convolutional neural network (CNN), bi-directional long and short-term memory neural network (BiLSTM), and attention mechanism (Attention) is proposed. By constructing a spatio-temporal grid, spatial features of foundation pit deformation are extracted using convolutional neural network, temporal features are modeled by combining bidirectional long and short-term memory network, the attention mechanism is introduced to improve the model’s attention to the key spatio-temporal locations, and finally the features are integrated by the fully-connected layer to output the predicted monitoring values. Based on the monitoring data of a deep foundation pit of a talent apartment in Guangzhou City for engineering case validation, the results of ablation and comparison experiments show that the proposed method has higher accuracy in the prediction of deformation in deep foundation pits.