Health assessment and deterioration prediction of pumped storage units based on improved KAN and optimal asynchronous characteristic spatial-temporal method

It is imperative to conduct a comprehensive health assessment and deterioration prediction of pumped storage units (PSUs) to guarantee the stable operation under prolonged and intricate conditions. To attain the optimal asynchronous health state of each component, a systematic and effective health assessment and prediction method is essential. An optimal asynchronous feature spatial-temporal data acquisition method and an improved Kolmogorov-Arnold network (KAN) network are proposed. Firstly, based on the asynchrony of the monitoring data distribution for each component of the unit, the data with the optimal distribution characteristics of each monitoring component of the unit is integrated to build optimal state health data (OSTHD). Compared with the traditional health data after overhaul, it can reflect the health status of the unit more comprehensively. Secondly, an improved model combining KAN and autoencoder is constructed based on OSTHD training, which has the capability of extracting features from complex, multi-dimensional data. Subsequently, in view of the coordinate vibration data of the monitoring parts of three components, the triangular area deterioration index (ADI) with the ability to reflect the change of the overall performance status of the unit is proposed and verified to be effective. Finally, based on the ADIs and performance degradation indexes, the health degradation trend is predicted using a prediction model combining Gaussian filter and bidirectional gated recurrent unit. A pumped storage power plant in China is used as the research case. The efficiency and feasibility of the presented methodology are demonstrated by comparing the results of several experiments. • Integrate the segmented data with the best distribution of each component to form the optimal health data for the PSU. • A triangle area deterioration index is constructed using measured and predicted vibration coordinates. • An AE-KAN network is established as the health model.