Abstract: To reduce the impact of the nonlinear and non-stationary characteristics of vibration signals on the accuracy of vibration prediction, a vibration prediction method of pumped storage units based on adaptive feature and optimized kernel extreme learning machine (KELM) is proposed in this paper. Firstly, the improved complete ensemble empirical mode decomposition with adaptive noise (ICEEMDAN) is used to decompose the vibration signal, and the intrinsic mode components of different frequency components are obtained. Secondly, autoencoder is used to adaptively extract the features of the obtained components, and the key features are dynamically captured. Then, the KELM prediction model is established to predict each component respectively, and DEIHHO is proposed to optimize the regularization parameter and kernel parameter of KELM, and then the final prediction result of unit vibration is obtained by superadding the prediction results of each component. Finally, the experimental results show that the proposed method has better prediction performance and can effectively improve the accuracy of vibration prediction of pumped storage units.