Abstract: Deformation control is essential when assessing the safety state of underground structures during construction. Due to the difficulty in quantifying the dynamic influence of underground construction, a deformation prediction model based on traction algorithm in control phases are proposed. The bidirectional long short-term memory (Bi-LSTM) is adopted to predict the time series of the monitoring data. The results of crucial construction stages obtained by numerical simulation are taken as traction points and are updated along with the stages according to the existing monitoring data. By using the attention mechanism combined with the bidirectional long short-term memory (Bi-LSTM-AM), the prediction results of the data-driven models in the control phases can thereafter be modified according to the updated traction points, achieving a more accurate and intelligent prediction of deformation during underground construction. Setting the traction relative weight allows the current reasonable traction degree to be adaptively determined, thereby realizing a valid fusion of Bi-LSTM and numerical simulation. The effectiveness of the traction prediction model is verified through relevant historical cases and data. Combined with the automatic monitoring case of Xi’an Metro Line 1 affected by the upper-span underpass of Fenghao third road in Xixian New District, the advantages and shortcomings of traction prediction are discussed. The results show that the traction effects improve the delay problem of data-driven prediction at the key construction stages, and the error is decreased by 24.34% on average. The continuously optimized traction points gradually approach the true values, which reduces the impact caused by the deviation of numerical simulation. The proposed model provides a fresh perspective on the deformation prediction during underground construction.

Key words: underground engineering, deformation prediction, time series, traction algorithm, Bi-LSTM, attention mechanism