Shallow-buried subway station construction period: Comparison of intelligent early warning and optimization strategies for surface deformation risk

Abstract

In the context of rapid urbanization, ensuring the safety of subway station construction is vital for the stability of urban infrastructure. Conventional intelligent construction risk prediction methods typically utilize large volumes of monitoring data for training to enhance model accuracy, often neglecting the relationship between the time-series width of the data and the prediction results. To address this issue, and to better serve the construction of shallow-buried subway stations at an earlier stage, this study proposed a bagging algorithm with an improved base learner combination strategy. This algorithm forms the basis for the Bayesian optimization-based random forest model (BA-RF) and the marine predators’ algorithm-optimized random forest model (MPA-RF). By examining trends in real-time data, such as surface and building settlements above the main structure, displacement at key points of the vertical shafts, and crown settlement, the short-term maximum values of key displacements were predicted. This study emphasized the impact of the time width of the input data on the accuracy of the predictive models. Through empirical analysis, the optimal time-series width was determined, allowing for effective short-term structural risk prediction and early warning using a smaller time series. The findings indicate that the BA-RF model, utilizing an improved base learner strategy, achieves higher prediction accuracy than the more complex MPA-RF model, effectively mitigating overfitting. Specifically, when the preceding measured data time widths were 5, 15, and 25 d, the BA-RF model’s mean absolute error was 0.168, 0.160, and 0.349, respectively, whereas the root mean square error was 0.853, 0.463, and 0.509, respectively. Combined with short-term future prediction applications at construction sites, it was demonstrated that appropriately selecting the time-series width can significantly enhance prediction accuracy even with relatively small data volumes. This study provides a method for selecting training data for intelligent risk management during subway station construction and offers practical data selection strategies for risk assessment in other large-scale construction projects. Thus, this method has significant scientific and practical applications.