The deformation monitoring of underground structures is no longer sufficient to prevent disasters. Therefore, the prediction of underground structure deformation is necessary. In the study, two dynamic prediction models of GM(2,1) and LSTM are combined to predict the tunnel settlement under artificial mountain, which used the NSGA-Ⅱ algorithm considering two evaluation indicators to achieve better prediction results, based on the long-term monitoring data of a tunnel project. The results show that LSTM can ensure higher prediction accuracy and can better simulate the development trend of settlement, meanwhile the gray model is difficult to simulate more local settlement changes. The combined models can quickly obtain the optimal weighting scheme under multiple evaluation indexes with the NSGA-Ⅱ algorithm. It can combine the advantages of each model, and get a more accurate prediction effect.