Accurate prediction of tunnel surrounding rock deformation is crucial for preventing casualties and economic losses caused by collapses, ensuring the smooth progress of construction. How to predict enclosure deformation more accurately when there is less monitoring data, this paper proposes a method that integrates ensemble empirical mode decomposition (EEMD) with particle swarm optimization (PSO) to fine-tune the hyperparameters of a long short-term memory (LSTM) network, forming the EEMD-PSO-LSTM prediction model. Using monitoring data from three points in the Aktepu tunnel in Xinjiang—focusing on vault subsidence and peripheral convergence—this study decomposes the data into five intrinsic mode function (IMF) components and a residual displacement trend (Res) through the EEMD method. These feature components are then input into the PSO-LSTM model, where PSO optimizes the LSTM’s learning rate and the number of hidden layers to improve surrounding rock deformation predictions. When compared with non-optimized models, the EEMD-PSO-LSTM model demonstrated superior accuracy and generalizability across all three monitoring points. The results confirm that the EEMD-PSO-LSTM model significantly enhances prediction accuracy and that PSO effectively optimizes LSTM performance, providing a reliable method for determining the optimal timing for secondary lining construction.