Deep learning-based prediction of efficiency and pressure in a reversible pump-turbine

The prediction of the efficiency and pressure in pump-turbine is crucial for the efficient and stable operation of the power plant. Deep learning methods can effectively overcome the limitations of traditional mathematical approaches in handling high-dimensional and large-scale data problems. In this study, a Fully Connected Neural Network (FCNN) is constructed to directly predict multi-condition efficiency using the runner design parameters of the pump-turbine. Three deep learning models Long Short-Term Memory (LSTM), Convolutional Neural Network combined with LSTM (CNN-LSTM), and Recurrent Neural Network (RNN) are developed to predict pressure in both temporal and spatial dimensions. The results indicate that the FCNN can employ the blade loading distribution and blade lean at the high-pressure side of the runner to realize the efficiency prediction, with an error within 1 %. In time dimension of pressure prediction, when predicting the pressure at the same monitoring point in the volute and the draft tube, all three models exhibit high reliability. In spatial dimension, employing the pressure data from the volute as the training set to predict the pressure of other components shows higher prediction accuracy than using the data from the draft tube. Among the three models, RNN achieves the highest prediction accuracy.