English：

With the increasing demand for the use of underground space and development of undergroundspace structure technology, the safety monitoring and earlywarning of foundation excavation retainingstructure in the construction stage is becoming more and more important. To investigate an effectiveprediction method for the horizontal displacement of foundation excavation retaining piles, based on theautomatic monitoring data for a project, a LSTM network with multi-input and multi-output data isproposed to establish the horizontal displacement prediction model of the foundation excavation retainingpiles, and the prediction results were compared with those generated by the random forest model and themulti-variable grey prediction model. The model not only can consider the horizontal displacement of themeasuring points, but also comprehensively can consider multi-modal data, such as the supporting axialforce of adjacent measuring points and the groundwater level. The results show that, by considering themulti-factor data, the multi-input and multi-output LSTM model exhibits higher prediction accuracy andcan better adapt to the complex time series patterns and nonlinear relationship, the root mean square errorof the prediction results is 52.5%, less than that generated by the single-input and single-output LSTMmodel, considering only the horizontal displacement of measuring points, and it is 62.3% and 59.0%less than that generated by the random forest model and the multi-variable grey prediction model,respectively.