[Purposes] Underground pipeline settlement is a key parameter during the construction of rectangular pipe jacking tunnels. Traditional prediction methods mostly rely on empirical models and statistical analysis, which struggle to accurately predict settlement in complex tunnel structures. [Methods] To address this challenge, this paper proposes a Self-Attention mechanism integrated prediction method for underground pipeline settlement based on particle swarm optimization (PSO) and a long short-term memory (LSTM) neural network. First, the Self-Attention mechanism is introduced to adaptively assign different weights to different time steps, which enhances focus of the LSTM neural network on critical time points and is conducive to capturing important information from the time series. Next, the PSO algorithm is used for hyperparameter optimization of the LSTM neural network to ensure that it can process the input data with optimal configuration. The LSTM neural network is responsible for extracting temporal features from the input data and capturing long short-term dependencies. Finally, the model outputs the predicted surface settlement through a fully connected layer of the LSTM neural network. To verify the superiority and robustness of the proposed algorithm, a dataset of underground pipeline settlement was constructed using measured data from the Changsha Metro Line 6 project. The prediction performance of pipe jacking settlement at different tunnel locations was compared using the LSTM，LSTM-Self-Attention，and PSO-LSTM-Self-Attention models. [Findings] The results show that the PSO-LSTM-Self-Attention model outperforms the others in terms of mean squared error, root mean squared error, mean absolute error, and the coefficient of determination. [Conclusions] This study validates that the PSO-LSTM-Self-Attention model exhibits superior performance in predicting underground pipeline settlement.