Abstract: Accurate water level prediction has important application value in fields of natural disaster early warning, water resource management and ecological environmental protection.Therefore, a hybrid water level prediction model based on robust local mean decomposition (RLMD), sample entropy (SampEn), convolutional neural network (CNN) and regularized extreme learning machine (RELM) is proposed.Firstly, RLMD is used to decompose the historical water level data, and the SampEn method is introduced to reorganize features of the component data in order to reduce data volume.Then, CNN is used to extract features of the reorganized data to improve the training speed.Finally, RELM is used to predict each sub-sequence, and the prediction results are superimposed to get the final prediction value of the water level sequence.Taking the daily water level data of Gaochang hydrological station in the lower reaches of Minjiang River Basin from 1997 to 2020 as the research object, the predictive performance of the model is verified.The results show that, in terms of predicting the water level 1-day ahead, the proposed hybrid model achieves accuracy improvement of 5.93%, 5.91%, and 0.52% compared to the RELM, CNN-RELM, and RLMD-CNN-RELM models, respectively.For three different forecast period (1, 2, and 3 days), the NSE values of the hybrid model′s prediction results are 0.934 657, 0.932 588, and 0.922 955, respectively, and the prediction accuracies all reach Class-A level.The established RLMD-SE-CNN-RELM model demonstrates high prediction accuracy and strong stability, providing a reference for water level prediction and precise water resource scheduling.