Abstract: Hydropower units deteriorating poses a significant effect on the safe and stable operation of hydropower stations and even on power grid systems; accurate analysis of their operation status needs an accurate prediction of the deterioration trend. This paper presents a hybrid model for predicting this trend based on the extreme gradient boosting (XGBoost) algorithm, the variational mode decomposition (VMD) algorithm, the bidirectional gated recurrent (BiGRU) neural network, and the attention mechanism (AM). First, we use the XGBoost algorithm to construct a health state model of hydropower units considering the influences of working head, active power, and guide vane opening. And this model is applied to predict the deteriorating trend in a period of several years. Then, we decompose the deteriorating trend using VMD and obtain several intrinsic model functions (IMF) that are relatively stable; for each IMF component, we construct a BiGRU-AM model. Finally, all the components are superimposed to give the final trend prediction. Application in a case study shows our method can accurately describe the deterioration trend of hydropower units and improve the accuracy of unit deterioration predictions significantly.

Key words: health state model, deterioration trend prediction, XGBoost, VMD, BiGRU, attention mechanism