Abstract: The traditional prediction of surrounding rock deformation in large section loess tunnels relies on fitting monitoring and measurement data, which is time-consuming and laborious. In the case of large data fluctuations, it can lead to a decrease in prediction accuracy. Deep learning algorithms can significantly improve the monitoring efficiency and prediction accuracy of surrounding rock deformation in loess tunnels, while reducing labor costs. This article is based on the Long Short Term Memory (LSTM) model and uses the Artificial Hummingbird Algorithm (AHA) to achieve hyperparameter optimization of the LSTM model, constructing a computational process framework for the AHA-LSTM model. Taking the Shaojiatang Tunnel project as an example, on-site measurements were conducted on the large deformation section of the surrounding rock to obtain monitoring data on arch crown settlement and arch foot convergence. These data were then inputted into the AHA-LSTM model for calculation and compared with traditional machine learning algorithms for analysis. The results showed that the average R² of the AHA-LSTM model was 0.9983, MAE was 0.5802, RMSE was 0.7323, MAPE was 1.8539%, and the relative error was controlled within -4.0% to 4.0%. It performed the best in the comparative model; The actual engineering verification shows that the average R² value is 0.9927 and the average MAPE is 4.2770%, achieving accurate prediction of the deformation of the surrounding rock of the loess tunnel. The prediction method proposed in this article provides a reliable reference for introducing the LSTM prediction method in deep learning into tunnel construction and promoting tunnel construction safety.