Abstract: In order to effectively improve the prediction accuracy of non-stationary and non-linear runoff series,the TVF-EMD-LSTM prediction model is proposed by coupling the time varying filter empirical mode decomposition（TVF-EMD）with adaptive sequence characteristics and Long Short-Term Memory neural network（LSTM）. Firstly,the TVF-EMD method is used to adaptively decompose the runoff series into high-frequency series and low-frequency series;secondly,the LSTM neural network is used to predict the decomposed series. Finally,the predicted results are added and reconstructed into the final runoff prediction results. The proposed model is applied to the monthly runoff forecast of the Changshui Hydrological Station in the Luohe River Basin,and compared with LSTM model,EMD-LSTM model and CEEMDAN-LSTM model. The results show that the TVF-EMD-LSTM neural network coupling model has the highest prediction accuracy and the smallest error. Therefore,TVF-EMD can better alleviate the mode mixing problem and provide a better way for the data preprocessing of runoff series. The TVF-EMD-LSTM hybrid model provides an effective new method for monthly runoff forecasting.