Health Prediction of Hydropower Units Based on VMD and Spatio-Temporal Transformer

To address the issues of low accuracy and significant fluctuations in the prediction results of hydropower unit operating condition trend forecasting techniques, this paper proposes a health condition trend prediction model for hydropower units based on variational mode decomposition (VMD) and a multi-scale spatiotemporal Transformer. The method first uses VMD to adaptively decompose the original health index sequence into a set of intrinsic mode functions (IMFs), thereby suppressing noise and reducing non-stationarity. Each IMF component is then fed into a multi-scale spatiotemporal Transformer model for prediction. This model employs a parallel multi-scale dilated convolution module to extract multi-dimensional features and uses a Transformer encoder to capture long-term temporal dependencies globally. Finally, the predictions for all components are integrated to reconstruct the complete health status trend. The model’s effectiveness and superiority are validated using actual operational data from Unit 3 of a hydropower station in Fujian Province. Results demonstrate that the proposed model outperforms baseline models such as GRU and standard Transformer, with key performance metrics improving by up to 70%. Specifically, the RMSE is as low as 0.00425, the MAPE is 0.51%, and the R² reaches 0.9874.