This paper is devoted to treat an issue with poor performance of CEEMDAN-TCN model in high-frequency signal prediction by employing a CEEMDAN-VMD double-layer time-frequency feature extraction method. Firstly, the short-term water level features are extracted through CEEMDAN (complete ensemble empirical mode decomposition with adaptive noise) and TCN (Temporal Convolutional Network) parallel prediction is performed on the extracted series of sub-signals. After that, the VMD (Variational mode decomposition) secondary feature extraction method is employed on the high frequency complex sub-signals and the traversal method is applied to determine the optimal decomposition method for achieving high tracking accuracy of TCN on the sub-signals. Finally, prediction results for each sub-signal are superimposed linearly. The validity of the proposed method for improving the prediction accuracy on short-term water level is verified by predicting the short-term water level data of Hung-tse Lake. To be specific, prediction accuracy achieves 99.7%, MAE and RMSE increase 29.6% and 30%, respectively.