Abstract:Disaster monitoring and prediction are critical tasks in geotechnical engineering. However, the inherent non-stationarity and non-linearity of engineering monitoring data have long posed challeng es for accurate forecasting. In response to this challenge, this study proposes an improved combination prediction model for the deformation of deep foundation pits in subway stations. The model integrates data-driven algorithms, including Extreme Learning Machine (ELM) and Long Short-Term Memory (LSTM) neural networks, along with Complete Ensemble Empirical Mode Decomposition with Adap tive Noise (CEEMDAN) and the Sparrow Search Algorithm (SSA). Initially, CEEMDAN was em ployed to decompose the horizontal displacement sequence of the retaining piles into trend and fluctua tion components, thereby reducing the data's non-stationarity. Furthermore, to fully capture the non linear characteristics of the differences among each decomposed sequence, SSA-optimized ELM and LSTM models were employed to predict the low-frequency trend component and high-frequency fluc tuation component, respectively. The results were then combined to reconstruct the final prediction values. Finally, the accuracy and practicality of the model were systematically evaluated through abla tion, comparative and generalization validation experiments using a deep foundation pit example in Zhengzhou subway station. The results demonstrated that the proposed model exhibited superior per formance in terms of both accuracy and stability when compared to other models. The R2 improve ments ranged from 2.88% to 23.62%, while the reductions in RMSE and MAPE were observed to be between 6.63% and 41.13% and between 8.08% and 64.79%, respectively. The model's efficacy in addressing data's non-stationarity and capturing nonlinear features is evident, offering high reliability and broad application prospects. The model provides novel insights to and technical support for disas ter prevention in geotechnical engineering.
