Optimizing cavitation fault forecasting in pump-turbine systems through gated recurrent units

The pressure fluctuation data in the pump-turbine runner region exhibit significant nonlinearity. The method of utilizing neural networks is employed to analyze pressure fluctuations in order to determine the occurrence of cavitation phenomena. This paper presents a model that utilizes a VMD (variational mode decomposition)-optimized algorithm combined with GRU (gated recurrent unit)–attention for the prediction of pressure fluctuations, aiming to facilitate the forecasting of cavitation-induced failures. Using data collected from a real machine over the course of one day, predictions were made using three different models: a standalone GRU, a combination of GRU and Attention mechanisms, and a combination of VMD and three different optimization algorithms. The evaluation of prediction performance indicates that the VMD–dung beetle optimization–GRU–attention model not only captures the nonlinear characteristics of the actual values but also aligns more closely with the trend of the real data. The error assessment results demonstrate that this model exhibits superior predictive performance. Analyze the prediction of pressure pulsation at three different locations between the runner and the guide vane, top and bottom cover. This method enables effective predictions of cavitation conditions up to 50 minutes in advance, showcasing its potential for practical engineering applications.