Prediction and analysis of foundation pit support deformation for ultra-large dock chambers based on CNN-LSTM-Attention model

The foundation pit of an ultra-large dock chamber adopts the anchor type composite steel sheet pile retaining system. To study the deformation law of foundation pit, the settlement, horizontal displacement and stress of retaining structures such as dock wall bearing platform, composite steel sheet pile and anchor wall were monitored, and their change trend, causes and temporal and spatial correlation were analyzed. The results show that the top settlement and horizontal displacement of the retaining structure and the deep horizontal displacement of the composite steel sheet pile are large in the excavation process due to the influence of soft soil characteristics and precipitation, and the data change of each monitoring item has a strong spatio-temporal correlation effect. A CNN-LSTM-Attention combined neural network model which can extract the key features of monitoring time series data is established. The monitoring data of different monitoring items are taken as the model input, and the predicted values are output through CNN neural network convolution operation, LSTM neural network regression prediction, Attention weight distribution, full connection layer and other operations, and compared with the measured data. According to the experimental results, the combined model of CNN-LSTM-Attention with attention mechanism has higher prediction accuracy than LSTM model and BiLSTM model. In the prediction of adjacent monitoring points, the model has good applicability, which verifies that the change trend of monitoring points’ data has strong spatial correlation.