The existing slope settlement prediction models have low accuracy and cannot effectively reflect the time series information contained in the settlement. A hybrid prediction model for slope settlement was proposed based on the gated deep recurrent belief network（GDRBN）. The adaptive learning rate is introduced to improve the efficiency of model training. Taking the project of the second phase of Guangfo-Zhaoqing Expressway as an example, different settlement prediction models was established, and the outcomes were contrasted. The results show that the proposed models will increase the accuracy of the prediction by 69%, 54%, 38%, and 26%, respectively, which is compared with traditional prediction models (such as GM, BP, RNN, DBN).It can provide a accurate calculation method for slope prediction.