Aiming at solving the problems of more noise, poor stationarity and low prediction accuracy of traditional time series models, a water level prediction model is proposed in combination with improved complete ensemble empirical mode decomposition with adaptive noise (ICEEMDAN) and Long Short-Term Memory (LSTM). The modeling process is as follows: The data mirror is extended to improve the endpoint effect caused by ICEEMDAN decomposition; The extended data is decomposed into several intrinsic mode components (IMF) by ICEEMDAN; The high-frequency noise component (denoted by IMF 1) will be eliminated, and the LSTM parallel prediction model of middle and low frequency components will be established; The final prediction result is obtained by reconstructing the prediction result of medium and low frequency components. Experiments show that this model has higher prediction accuracy than LSTM, LightGBM, EMD-LSTM, ICEEMDAN-LightGBM, and ICEEMDAN-LSTM model without mirror continuation and high-frequency noise elimination. In the prediction of the upper water level of Luoma Lake Reservoir, MAE is 0.008m and RMSE is 0.022m, and R2 reaches 99.8%.