Abstract: Landslide displacement prediction is a critical component of mine disaster early warning. However， landslide displacement time series exhibit significant nonlinear behavior， making it difficult for a single prediction model to capture these characteristics simultaneously， which limits prediction accuracy. To fully extract characteristic information from the displacement time series and improve prediction stability and accuracy， time series decomposition was combined with deep learning methods in this study and a short-term displacement prediction model for mining landslides based on the time series of ground-based micro-deformation monitoring radar was proposed. First， landslide displacement time series were extracted from radar images， and filtering algorithms were applied to denoise the raw displacement data. Next， variational mode decomposition （VMD） was used to decompose the displacement time series into a baseline term and a fluctuation term. Considering the distinct temporal characteristics of each component， an autoregressive integrated moving average （ARIMA） model was used to predict the baseline displacement， while the bidirectional temporal convolutional network-long short-term memory （BiTCN-LSTM） hybrid model was adopted to predict the fluctuation displacement. The total displacement was obtained by superimposing the predicted terms. The proposed model was validated using two monitoring points with different deformation characteristics in an open-pit mine and was compared with existing prediction models. The results showed that VMD effectively separated displacement series into terms with different frequency characteristics， reducing prediction complexity. By integrating the local feature extraction capability of the bidirectional temporal convolutional network （BiTCN） with the long-term dependency learning ability of the long short-term memory （LSTM） network， the hybrid model significantly improved the prediction accuracy of the fluctuation term displacement. Compared with the single BiTCN and LSTM models， the proposed model reduced the root mean square error and average absolute error by 20%~60%， and the fitting coefficient reached or exceeded 0.98. The prediction error was mainly concentrated in the range of 0~0.5%， demonstrating good stability and generalization ability. This paper provides an effective method for predicting landslide displacement and a new approach for landslide prediction and early warning in mines.