Prediction of surface settlement around subway foundation pit based on Self-CGRU model

To improve the prediction accuracy of surface settlement around subway foundation pit, a deep attention hybrid prediction model, termed self-Attention convolutional gated recurrent units (Self-CGRU), is proposed based on the self-attention mechanism and deep learning. The Self-CGRU model can capture the spatio-temporal characteristics of settlement data. The Self-CGRU model is constructed by integrating a spatial module and a temporal module. In the spatial module, the convolutional neural network is selected to capture the spatial correlations of settlement data obtained from the adjacent monitoring points. In the temporal module, the gated recurrent units neural network is used to analyze the temporal rules of settlement data. In addition, the self-attention mechanism is introduced into the Self-CGRU model to capture the autocorrelation in settlement data. Then, the predicted values of settlement can be obtained. Surface settlement data around the subway foundation pit in Shenzhen, China are selected to verify the performance of Self-CGRU model. The results indicate that the Self-CGRU model outperforms existing models, achieving a prediction accuracy improvement ranging from 17.48% to 29.17% compared to these models. The research results can provide an accurate and stable new model for the prediction of surface settlement around subway foundation pit.