Pressure pulsation forecasting in pump-turbines via variational mode decomposition and optimized gated recurrent unit-attention network

In order to solve the challenge posed by the strong nonlinearity of pressure pulsation data in the runner region of pump-turbines for the prediction of cavitation faults, this study utilizes neural networks to explore the potential connection between cavitation coefficients and pressure pulsations. The Variational Mode Decomposition-Dung Beetle Optimizer-Gated Recurrent Unit-Attention (VMD-DBO-GRU-Attention) composite model is applied for pressure pulsation prediction to identify cavitation faults. Based on the measured data of a unit, a single gated recurrent unit-attention (GRU) model, a GRU combined with Attention mechanism model, and the newly constructed VMD-DBO-GRU-Attention model are compared for prediction. After evaluation of the prediction results, the VMD-DBO-GRU-Attention model accurately portrays the nonlinear signals of pressure pulsation, and the prediction curves are highly consistent with the trend of the actual values, with a number of error indicators better than those of the comparison model. Meanwhile, by analyzing the pressure pulsations at the three positions, including between the runner and the top cover, and comparing them with the actual engineering data, the method can predict the operating status of the unit more than 15 min in advance, which shows good engineering practical value.