Degradation trend prediction of pumped storage unit based on a novel performance degradation index and GRU-attention model

Abstract

Degradation trend prediction (DTP) is an essential approach to ensure the secure operation of pumped storage unit (PSU). Its accuracy is mainly reflected through the reliability of performance degradation index (PDI) and prediction model. However, the inevitable fitting error of machine learning method bring uncertainty to PDI. Moreover, the precise DTP is challenging due to the complex nonlinear PDI sequence. In this paper, a combined degradation trend prediction (DTP) model which construct reliable feature extraction-based PDI and achieve accurate DTP is proposed. Firstly, a feature fusing the kernel density estimation (KDE) and Wasserstein distance called KDE-W, is presented to extract the degradation information from monitoring data directly. Then, the feature extraction-based PDI is constructed based on KDE-W, state update strategy (SUS) and the warning state are devised to improve the reliability of PDI. Finally, the GRU-A prediction model, which combines GRU with the attention mechanism, is designed to predict the degradation trend of PSU. Model verification and comparative analysis are carried out on a PSU in China. The results reveal that the proposed combined DTP model can not only effectively extract the degradation trend, but also achieve outstanding prediction results compared with other popular prediction models.