Abstract: A two-layer fusion prediction model based on landslide displacement time series decomposition is proposed to address the problems of existing landslide displacement prediction models which rely too much on the environmental influence factors and do not mine the fluctuating term displacements sufficiently. The landslide displacement time series is decomposed using the improved complete ensemble empirical mode decomposition with adaptive noise (ICEEMDAN) method, and a number of intrinsic mode function (IMF) components with different frequencies are obtained. The model training is accelerated by Gaussian filtering and manual extraction of nonlinear features in the time and frequency domains to reduce the model dependence on the amount of data and arithmetic power. Four machine learning algorithms, namely, error back propagation neural network (BP neural network), extreme learning machine (ELM), long-short-term memory network (LSTM), and convolutional neural network-bidirectional long-short-term memory network (CNN-BiLSTM), are used as the first-layer learners, to make the preliminary prediction of each IMF component separately, and the hyper-parameter optimization is performed using the crested porcupine optimization (CPO) algorithm. For the outputs of the first layer learners, the Ridge regression model and XGBoost regression model were used as the second layer learners for feature fusion, respectively, to construct the final landslide displacement fusion prediction model. The results show that the CNN-BiLSTM-XGBoost model performs the best in prediction performance among the above eight combined models, with the mean absolute error (MAE) of 0.072 mm, the root mean square error (RMSE) of 0.097 mm, and the coefficient of determination (R²) of 0.998 1, which shows its high accuracy and stability in landslide displacement prediction. On this basis, ablation experiments are conducted for the CNN-BiLSTM-XGBoost model to evaluate the independent contributions of the three algorithms, CNN, BiLSTM and XGBoost, to the model performance, and to demonstrate the superiority of the CNN-BiLSTM combination algorithm compared to the single CNN or BiLSTM algorithm and the CNN-BiLSTM-XGBoost dual-layer learner model over the CNN-BiLSTM single layer learner model is necessary.