Pressure pulsation prediction of pump-turbine during transient processes based on physical model and data-driven approach: A case study of load rejection transient process

As intermittent energy sources like wind, solar, and hydropower are integrated on a wide scale, pumped storage power stations (PSPs) are compelled to frequently switch between multiple operating conditions, leading to a significant increase of transient processes. Predicting the pressure pulsations (PP) during these transient processes in advance is of great significance for ensuring the safety of the PSP. An intelligent prediction model of PP combining physical model and data-driven approaches is proposed in this study. First, a physical model based on the method of characteristics is constructed to calculate the water hammer pressure (WHP) at key measurement points of the PSP. Second, a data enhancement model based on successive variational mode decomposition is proposed, which can expand a single PP dataset into an arbitrarily large number of samples. Next, a PP prediction model is developed, which is a deep learning model combining one-dimensional multi-scale convolutional neural networks, bidirectional long short-term memory networks, and attention mechanism. The intricate nonlinear interaction between WHP and PP can be accurately represented by this model. Finally, the computed WHP are fed into the trained model to forecast the PP. Historical transient data from load rejection testing at three domestic PSPs are utilized to validate the effectiveness of the suggested strategy. The findings demonstrate that the prediction error for the maximum amplitude of PP at the volute inlet during 100% load rejection is only 0.76%. This paper presents a thorough foundational model for the intelligent prediction of transient PP in PSP.