Research on subway settlement prediction based on the WTD-PSR combination and GSM-SVR model

Due to the speeding up of urban development and the rapid population expansion in China, the subway has become the preferred mode of transportation for people, and urban underground spaces are continuously being improved. However, during the construction of subways, surface settlement around the area is inevitably caused, which can significantly impact the safety of the building procedure and surrounding buildings. Therefore, accurately predicting surface settlement around subway tunnels is of great practical significance. This study proposes a GSM-SVR model for subway settlement prediction, based on WTD-PSR data preprocessing. The data of surface settlement data from three measurement points in a section of the 1st Line of the urban rail transit system was taken as the research subject. By upgrading the one-dimensional settlement data sequence to a multi-dimensional data sequence, and utilizing the Grid Search Method Optimized Support Vector Regression (GSM-SVR) regression model to predict subway settlement with small sample data, the aim is to offer a more precise and reliable data analysis method and theoretical approach for small sample subway settlement prediction. First, wavelet denoising (WTD) is applied to the field measured data, and the denoised time series data is subjected to phase space reconstruction (PSR) to obtain multi-dimensional time series data. The validity of the embedding dimension determined by the phase space reconstruction is verified, providing rich multi-dimensional features for the subsequent prediction models. Based on the reconstructed data, traditional Support Vector Regression (SVR) models and SVR models optimized by the Grid Search Method (GSM) are constructed. Furthermore, Particle Swarm Optimization (PSO), Gray Wolf Optimization (GWO), Marine Predators Algorithm (MPA), and Whale Optimization Algorithm (WOA) are introduced to optimize the SVR model, and the prediction performance is compared with that of the Long Short-Term Memory (LSTM) model. By comparing the prediction accuracy of the seven models, the results show that the Grid Search Method optimized SVR model performs the best in the light of prediction accuracy, with a Mean Absolute Percentage Error (MAPE) of 2.96%, a Mean Absolute Error (MAE) of 0.028 mm, a Root Mean Square Error (RMSE) of 0.032 mm, and a coefficient of determination (R²) of 0.995. Compared with the other six models, the three-error metrics are reduced by 25.25%-64.72%, 32.93%-68.81%, and 34.43%-70.53%, respectively, and the R² value is increased by 0.75%-6.39%, significantly outperforming traditional empirical models. These results indicate that the GSM-SVR model based on WTD-PSR significantly outperforms single-algorithm optimization strategies, making it more suitable for predicting future ground settlement data. This approach provides a reusable hybrid framework for small-sample settlement prediction in subway systems, offering improved guidance for practical engineering applications.