This paper presents a vibration prediction method of pumped storage units based on adaptive features and an optimized kernel extreme learning machine (KELM) to reduce the impact of the nonlinear

non-stationary characteristics of vibration signals on the accuracy of vibration predictions. First

we use improved complete ensemble empirical mode decomposition with adaptive noise (ICEEMDAN) to decompose a vibration signal and generate the intrinsic mode components of different frequencies. And

an autoencoder is used to extract the features of these components adaptively and capture their key features dynamically. Then

a KELM prediction model is developed to predict each component separately

using a modified DEIHHO algorithm to optimize its regularization parameter and kernel parameter. Finally

the final prediction result of unit vibration is obtained by superadding the predictions of all the components. Comparison with previous experimental data shows our new method is better in vibration prediction of pumped storage units and improves the accuracy effectively.