Hybrid SE-attention-CNN-LSTM network for hydraulic turbine water guide bearing temperature prediction in hydropower generation station

Abstract

Accurate prediction of the water turbine bearing temperature is of great significance to ensure safe operation and improve economic benefits of the hydropower station. The water turbine bearing temperature forecasting is a challenging problem due to the running data being nonlinear and non-stationary. This work proposes a hybrid SE-CNN-LSTM forecasting model by integrating the advantages of convolutional neural Networks (CNN), long short-term memory (LSTM) networks and squeeze-and-excitation (SE) attention mechanism. CNN is designed to extract local features from raw temperature time series data to capture short-term dependencies in temperature data. LSTM is designed by incorporating a state-space representation to better model the dynamic characteristics of long-term correlations in temperature data. SE attention mechanism is integrated into the CNN-LSTM to form the SE-CNN-LSTM network for accurate prediction of the water turbine’s guide bearing temperature. Experiments are conducted by using real operational data from a pumped-storage power station. Experimental results demonstrated that the proposed SE-CNN-LSTM model outperforms traditional prediction methods in terms of multiple evaluation metrics, exhibiting superior prediction accuracy and robustness.