Abstract: The monitoring value of ground settlement is characterized by complexity and nonlinear dynamic change. Aiming at the problems that the previous static models are easily disturbed by historical monitoring data and the model input weights and thresholds are more difficult to choose, a dynamic prediction method for ground settlement of the portal section of tunnels is proposed. The ground settlement is equidistant by the cubic-spline function interpolation method and decomposed into the trend and random term displacement by the time series analysis theory and the variational mode decomposition (VMD). By using the grey wolf optimizer (GWO) to optimize the weights and thresholds of the online sequential extreme learning machine (OSELM), the GWO-OSELM dynamic prediction model is established to predict the displacement components separately. Taking the portal section of Xinglong tunnel in Chongqing as an example, the proposed model is compared with the traditional model. Finally, the influences of the choice of activation function on the prediction performance of the model and some factors influencing the random term displacement are analyzed. The results show that the model can effectively predict the displacement components after the preprocessing of non-equidistant time series data, and it has high prediction accuracy and small prediction error. Moreover, the Sigmoid activation function is more suitable for the model, and the rates of the ground settlement and the vault subsidence have important influences on the random term displacement. The model provides a new way of thinking and a method for the long-term prediction of ground settlement in the portal section of mountain tunnels.